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.1000413 | A Novel Role for MAPKAPK2 in Morphogenesis during Zebrafish Development | One of the earliest morphogenetic processes in the development of many animals is epiboly. In the zebrafish, epiboly ensues when the animally localized blastoderm cells spread, thin over, and enclose the vegetally localized yolk. Only a few factors are known to function in this fundamental process. We identified a maternal-effect mutant, betty boop (bbp), which displays a novel defect in epiboly, wherein the blastoderm margin constricts dramatically, precisely when half of the yolk cell is covered by the blastoderm, causing the yolk cell to burst. Whole-blastoderm transplants and mRNA microinjection rescue demonstrate that Bbp functions in the yolk cell to regulate epiboly. We positionally cloned the maternal-effect bbp mutant gene and identified it as the zebrafish homolog of the serine-threonine kinase Mitogen Activated Protein Kinase Activated Protein Kinase 2, or MAPKAPK2, which was not previously known to function in embryonic development. We show that the regulation of MAPKAPK2 is conserved and p38 MAP kinase functions upstream of MAPKAPK2 in regulating epiboly in the zebrafish embryo. Dramatic alterations in calcium dynamics, together with the massive marginal constrictive force observed in bbp mutants, indicate precocious constriction of an F-actin network within the yolk cell, which first forms at 50% epiboly and regulates epiboly progression. We show that MAPKAPK2 activity and its regulator p38 MAPK function in the yolk cell to regulate the process of epiboly, identifying a new pathway regulating this cell movement process. We postulate that a p38 MAPKAPK2 kinase cascade modulates the activity of F-actin at the yolk cell margin circumference allowing the gradual closure of the blastopore as epiboly progresses.
| One of the earliest cell movement processes in the development of many animals is epiboly. In the zebrafish, epiboly ensues when the blastoderm cells spread over and enclose the yolk cell. Only a few factors are known to function in this fundamental process. We identified a maternal-effect mutant, betty boop (bbp), which displays a novel defect in epiboly, wherein the blastoderm margin constricts dramatically, precisely when half of the yolk cell is covered by the blastoderm, causing the yolk cell to burst. We demonstrate that Bbp functions in the yolk cell to regulate epiboly. We positionally cloned the bbp mutant gene and identified it as the serine-threonine kinase Mitogen Activated Protein Kinase Activated Protein Kinase 2, or MAPKAPK2, which was not previously known to function in embryonic development. We show that the regulation of MAPKAPK2 is conserved within a p38 MAP kinase pathway, thus identifying a new pathway in the regulation of this fundamental cell movement process. We postulate that a p38 MAPKAPK2 kinase cascade modulates F-actin contraction at the yolk cell margin circumference, allowing the gradual closure of the cells over the yolk cell as epiboly progresses.
| Early embryonic development is marked by cellular movements that ultimately generate the shape of the embryo in a process known as morphogenesis. One of the earliest morphogenetic events in many animals is the process of epiboly, whereby embryonic tissues spread and thin [1]–[5]. In the zebrafish embryo, three distinct cell layers lying at the animal pole of the embryo undergo epiboly: the enveloping layer (EVL) and yolk syncytial layer (YSL), both of which are extraembryonic, and an intermediate deep cell layer that forms the embryo proper (Fig. 1). About 1 hour after the mid-blastula transition, the morphogenetic process of epiboly begins. The deep blastomeres radially intercalate, while the underlying yolk moves animalward in a process called doming. At completion of this initial phase of epiboly, an inverted bowl-shaped blastoderm covers ∼50% of the yolk surface, referred to as the 50% epiboly stage (Fig. 1). During the second phase of epiboly, the deep cells begin gastrulation cell movements converging dorsally and undergoing involution/ingression movements to form the germ layers [5]. At the same time, epiboly continues with all three cell layers spreading over the yolk to the vegetal pole of the embryo, ultimately resulting in the complete internalization of the yolk [6]. The morphogenetic process of epiboly also occurs in numerous other vertebrates and invertebrates, including amphibia, sea urchins, and C.elegans [1]–[4].
The YSL actively participates in epiboly. Within the YSL, microtubule organizing centers associated with the yolk syncytial nuclei (YSN) extend microtubule arrays vegetally into the cortical yolk cytoplasmic layer (YCL) (Fig. 1). Ablation of microtubules with UV treatment or nocodazole slows or arrests epiboly progression [7],[8]. Studies of YSN movements suggest that movement of blastomeres and YSN are coordinated [9]. Although the mechanism remains unknown, E-cadherin is required for the coupling of the deep cells to the YSL and EVL in coordinating this movement between tissue layers in zebrafish [10]–[12]. In Xenopus fibronectin-integrin cell adhesion interactions act in radial intercalation during epiboly [13]. As the EVL and blastoderm cells move over the yolk, the yolk cell membrane is actively removed via endocytosis [14]. Also within the YSL is an actin band, first identified in Fundulus heteroclitus, which is required for epiboly movements and is postulated to close the blastopore (the uncovered vegetal yolk surface) as epiboly progresses (Fig. 1) [15]–[17].
Although epiboly is a fundamental morphogenetic process, only a handful of factors have been identified regulating this process. To identify additional molecular regulators, we investigated the zebrafish maternal-effect mutant, betty boop (bbp), which displays a novel defect in epiboly [18],[19]. In bbp mutants the leading edge of the blastoderm constricts dramatically at 50% epiboly causing the yolk cell to burst. This defect in epiboly is not seen in other mutants or by pharmacological treatments, suggesting that a novel aspect of epiboly is affected. Through whole blastoderm transplants and mRNA microinjection rescue, we determined that Bbp functions in the yolk cell. Consistent with it having a novel function in epiboly, we found that microtubules, actin band formation and endocytosis in the YSL appear normal prior to the onset of the phenotype. We positionally cloned the bbp mutant gene and identified it as the zebrafish homolog of Mitogen Activated Protein Kinase Activated Protein Kinase 2 (MAPKAPK2). Mutation of the p38 MAP kinase phosphorylation sites in MAPKAPK2 implicates p38 in regulating MAPKAPK2 function. Expression of a dominant-negative p38 MAP kinase demonstrates that it functions in a similar manner to MAPKAPK2 in epiboly. Furthermore, we found a dramatic increase in calcium release in bbp mutants, possibly reflecting altered actin contraction. Thus, we identified the p38 MAPAPK2 pathway as a new regulator of the fundamental morphogenetic process of epiboly. We propose that the p38 MAPKAPK2 kinase cascade modulates actin contraction at the blastoderm margin to close the blastopore during normal epiboly progression.
In a maternal-effect screen, we identified a mutant, bbp, which displays a striking morphogenesis defect in epiboly [18]. bbp mutant embryos appear to develop normally until 50% epiboly (Fig. 2 A,B); however, just as the blastoderm cells reach 50% epiboly, the equator constricts around its circumference, causing the yolk cell to burst (Fig. 2 B″). This defect is a 100% penetrant, maternal-effect phenotype, in which all embryos from homozygous females are mutant (hereafter called bbp mutant embryos), irrespective of their paternal genotype. Time-lapse microscopy of mutant (n = 9) and wild-type (WT) (n = 9) embryos (Fig. 2 C–D) shows that bbp embryos undergo abnormal shimmying movements periodically during blastula stages. That is, rapid, abnormal movements of large regions of the blastoderm, which increase in amplitude until the constriction and bursting of the yolk (Video S1).
We investigated if defects in patterning could account for the bbp phenotype. We found that the expression of goosecoid (Fig. 3 A, D), a dorsal organizer marker, and no tail (Fig. 3 B, E), a mesodermal marker, was normal in bbp embryos (n = 10, 13 respectively). Analysis of bmp4 and eve1 expression in ventral and ventrolateral regions, as well as foxb1.2 [20] in dorsal regions confirmed that patterning is normal in bbp embryos (data not shown). Thus, the defect in bbp appears to be specific to the morphogenetic process of epiboly.
Proper epiboly progression depends on microtubules that are present in the yolk cell [7],[8]. We analyzed microtubule array formation in the YCL (yolk cytoplasmic layer) at 50% epiboly just prior to bursting. Anti-tubulin stainings showed that microtubules are properly formed and robust in mutant embryos (Fig. 3 F, n = 23), similar to WT embryos (Fig. 3 C, n = 17). Thus, microtubule array formation does not appear defective in bbp embryos.
As epiboly progresses, the yolk cell membrane adjacent to the advancing blastoderm cells is removed via endocytosis. This process decreases the yolk cell membrane during epiboly, as the deep cells, EVL and YSN move over the yolk cell, and may drive their vegetal movement [14],[16]. We found that bbp mutant (n = 8) and WT (n = 10) embryos both internalized Rhodamine-dextran dye (Fig. 3 G,H), indicating that endocytosis was normal and is unlikely the cause of the epiboly defect.
It was unclear if the primary cause of the bursting of bbp embryos is the marginal constriction or a loss of yolk membrane integrity, which secondarily leads to the buckling of the blastoderm margin. To distinguish between these possibilities, we incubated WT and bbp embryos in hypertonic media to stabilize the yolk membrane in an attempt to prevent it from failing during the manifestation of the phenotype. The hypertonic medium caused no defects in WT embryos (Fig. 3 N) and stabilized the yolk membrane of bbp mutants, as evident by the lack of yolk globules protruding through the membrane (Fig. 3 K, L arrowheads). Importantly, despite the stabilized membrane, bbp embryos continued to constrict around their circumference at 50% epiboly in the hypertonic medium without causing the yolk to burst (Fig. 3 L, n = 56). Eventually the blastoderm pinched off from the vegetal yolk and healed, and the embryo continued to develop to 1 day post fertilization (dpf) (Fig. 3 M). This result indicates that the bursting phenotype results from a mechanical constriction, rather than from a yolk membrane integrity defect.
It is thought that a slow and controlled constriction of an actin ring present in the YSL closes the blastopore as epiboly progresses in the Fundulus embryo [15]. Similarly, in zebrafish a punctate actin band forms just vegetal to the blastoderm/EVL margin in the yolk cell at the 50% epiboly stage [16]. This actin band coincides with the region of yolk cell membrane endocytosis [16] and the marginal constriction of EVL cells during epiboly [21]. Phalloidin staining shows that an F-actin band forms in the YSL and at the EVL margin of bbp embryos (Fig. 3 J, n = 15), as in WT embryos (Fig. 3 I, n = 20). However, it remains unclear if the F-actin functions normally during epiboly.
We investigated the timing of the marginal constriction in bbp mutant embryos to distinguish between two alternative defects. One possibility is that the vegetal-ward movement of the blastoderm is arrested in bbp embryos, while blastopore closure is unaffected, thus causing the observed marginal constriction as the blastopore gradually tries to close in the absence of vegetal-ward cell movement. Such a constriction defect is seen in a subset of embryos in which microtubule function is blocked [7],[8] and in embryos depleted of Mtx2, a presumptive transcription factor acting in the yolk cell to regulate epiboly [22],[23]. Alternatively, the marginal constriction could arise due to too rapid closing of the blastopore, i.e. precocious blastopore closure, caused by unregulated actin constriction. We found that the margin does not constrict gradually over a ∼3 hour period as in WT, but instead occurs rapidly within a 20 to 30 minute period in bbp mutants (Fig. 2 and Video S1). This result is consistent with a model in which the blastopore closes precociously in bbp mutants through an unregulated marginal F-actin constriction.
We examined the behavior of the yolk syncytial nuclei (YSN) in bbp mutant embryos to determine if YSL morphogenesis is affected. Below the blastoderm at 50% epiboly, the internal YSN (I-YSN) are widely distributed, while at the margin the external YSN (E-YSN) are more densely packed [7]–[9]. In time-lapse analysis of fluorescently labeled nuclei, we observed a normal distribution of the I-YSN during epiboly in bbp mutant and WT embryos. However, following a strong shimmying movement and coincident with the onset of the marginal constriction shortly before bbp mutants burst, the I-YSN withdrew from the location of a strong shimmying movement and collapsed into a small area of the I-YSL (Fig. 3 O–V, Video S2, S3). The E-YSN remained in place until the yolk cell burst minutes later. Thus, YSN epiboly movements appeared normal in bbp mutants until 50% epiboly, when strong shimmying movements and the marginal constriction likely cause them to behave abnormally secondarily.
During early epiboly stages, calcium levels are low throughout the embryo; however, beginning at 50% epiboly calcium levels become dynamic and are required for formation of the yolk cell actin band and epiboly progression [16], [24]–[26]. We investigated if calcium dynamics are altered in bbp mutants during epiboly, which could reflect a change in the yolk cell F-actin function. We examined calcium dynamics by ratiometric imaging with the fluorescent calcium indicator, Fura-2. Time-lapse microscopy and transient composite analysis of calcium activity showed a dramatic increase in calcium release in bbp mutants (n = 7) compared to WT embryos (Video S4 and S5). During early epiboly when doming of the yolk occurs, calcium activity is maintained at a sustained level in WT embryos (Fig. 4 A, arrowhead), whereas in age-matched bbp embryos, ectopic calcium release activity was observed (Fig. 4 B, arrowheads). This ectopic release increased in frequency and intensity as epiboly progressed (Fig. 4 C, arrowheads) until eventual bursting. Analysis of calcium flux throughout early epiboly in a composite pseudo-colored image shows low calcium activity at the margin in WT (Fig. 4 D), which is clearly increased in bbp embryos (Fig. 4 E, F). Thus the increased calcium release may lead to increased contraction of F-actin at the yolk margin and the dramatic constriction observed in bbp mutant embryos.
The Bbp protein may function in the yolk cell, deep cells or EVL. To determine the embryonic domain in which Bbp functions, we performed whole blastoderm transplants to separate yolk, YSL, and YCL structures from the deep cells of the blastoderm and EVL [27]. WT blastoderms were then placed on bbp yolks, and vice versa (Fig. 5 A). Chimeric embryos containing WT yolk and mutant blastoderm completed epiboly (Fig. 5 B–B′″, n = 5) and were viable through at least 6 dpf (data not shown). In contrast, chimeric embryos containing mutant yolk and WT blastoderm constricted at the equator and burst at 50% epiboly (Fig. 5 C–C″, n = 7), similar to bbp embryos. These data indicate that Bbp functions in the yolk cell to regulate epiboly, consistent with it modulating the closing of the blastopore via actin constriction within the yolk cell.
To identify the molecular nature of bbp, we mapped the bbp mutation to the centromere of chromosome 11 using SSLP (simple sequence length polymorphic) markers [28] in bulk segregational analysis of homozygous mutant versus WT sibling adult female fish. The interval was narrowed by fine recombination mapping using more than 1100 meiotic events to a 900 kb interval based on the Sanger Centre zebrafish genome sequence (www.ensembl.org/Danio_rerio). Ten novel ESTs and one known cDNA were present in the critical interval. RT-PCR of ovary cDNA identified 7 ESTs and the one known cDNA as maternally-supplied RNAs (data not shown). Systematic sequencing of these genes from WT and bbp mutant ovary cDNA identified a non-sense mutation (Fig. 6A) in the known serine-threonine kinase gene, Mitogen Activated Protein Kinase Activated Protein Kinase 2 (MAPKAPK2). Sequence analysis of the genomic region of the parental mutagenized fish demonstrated that this mutation did not exist in the background and thus was induced by the ENU mutagenesis. The mutation introduces a premature stop codon, predicting a carboxy-terminal 33 amino acid truncation, which removes an identified nuclear localization signal and a p38 docking domain (Fig. 6A) [29],[30]. The MAPKAPK2 transcript is present maternally in the egg and is found ubiquitously in the blastoderm through the 50% epiboly stage (data not shown, Fig. 6C,D).
The bbp phenotype could be rescued by injection of 20 pg of WT MAPKAPK2 mRNA at the 1-cell stage (n>500, 100% rescue), demonstrating that MAPKAPK2 is the gene defective in bbp mutants. This is among the first maternal-effect mutant genes to be cloned positionally in a vertebrate. Importantly, injection of WT MAPKAPK2 mRNA into the YSL at the 512- to 1000-cell stage rescued the constriction defect and the calcium defect of bbp mutant embryos, confirming that MAPKAPK2 is required in the yolk cell (n = 20, 100% rescued). Injection of 1.5 ng of mutant MAPKAPK2 mRNA (75-fold overexpression compared to WT mRNA) had no rescuing activity (n = 75). Furthermore, injection of mutant MAPKAPK2 mRNA (250 pg to 1.25 ng, n>50) into WT embryos caused no phenotypic consequences, indicating that the mutant protein is inactive and has no dominant-inhibitory activity when overexpressed. In conjunction with rescuing amounts of MAPKAPK2 mRNA, we injected a translation blocking morpholino to MAPKAPK2 into bbp mutant embryos. Rescue was inhibited by the morpholino (n = 17, 100% no rescue; 100% rescued by mRNA injection alone), suggesting that it can block translation. However, morpholino injection alone into WT embryos did not phenocopy bbp (n = 14), possibly due to maternally-supplied MAPKAPK2 protein or high maternal levels of transcript.
MAPKAPK2 is a well-characterized direct target of p38 MAP kinase (MAPK). p38 MAPK activates MAPKAPK2 by phosphorylating key residues on MAPKAPK2. MAPKAPK2 has been extensively studied in cell culture, including structure function analysis, and the identification of nuclear and cytoplasmic targets [31]. However, very few of the results from these cell culture studies have been tested in an animal model.
Studies of mammalian MAPKAPK2 in cell culture show that the protein is localized to the nucleus under basal conditions, but upon phosphorylation by p38 MAPK, an overriding nuclear export signal (NES) stimulates its export to the cytoplasm [29],[30],[32],[33]. The carboxy-terminal truncation of MAPKAPK2 in bbp mutants results in loss of the NLS (Fig. 6A), suggesting that the mutant protein may be constitutively localized to the cytoplasm and fail to phosphorylate nuclear targets, thus causing its loss of function. To address this question, we examined the subcellular localization of myc-tagged WT and Bbp MAPKAPK2 proteins in embryos. The myc-fusion did not compromise MAPKAPK2 activity, as it fully rescued the bbp mutant phenotype (Fig. 6E, n = 108, 100% rescued). We found that WT myc-MAPKAPK2 predominantly localized to the nucleus, with additional weak localization in the cytoplasm, consistent with previous studies in mammalian cells (Fig. 6F, n = 12). Bbp myc-MAPKAPK2 showed increased localization to the cytoplasm and the cell cortex. Surprisingly, a significant amount of the mutant fusion protein remained in the nucleus, despite loss of the NLS (Fig. 6G, n = 10). GFP-fusions to WT and the Bbp mutant proteins behaved similarly, except that the WT GFP-fusion was found exclusively in the nucleus (data not shown). The inability of the mutant fusion proteins to rescue the mutant phenotype (myc-Bbp, n = 14; GFP-Bbp, n = 14) despite their nuclear localization, indicates that other properties of MAPKAPK2 are deficient in the Bbp protein. These results also suggest that interacting factors or additional features of MAPKAPK2 can localize it to the nucleus in the absence of the NLS.
We next investigated the kinase activity of the mutant protein to determine if it was constitutively active in the cytoplasm or otherwise misregulated. We analyzed the phosphorylation status of a well-established cytoplasmic substrate of MAPKAPK2, heat shock protein 27 (HSP27) [31]. Antibodies specific to the phosphorylated form of HSP27 did not detect the endogenous zebrafish protein in embryos. Therefore, we analyzed the activity of WT and Bbp MAPKAPK2 in transfected HeLa cells. We first confirmed that the subcellular localization of the proteins in HeLa cells recapitulated that seen in intact embryos (data not shown). We next examined their ability to induce phosphorylation of HSP27. As shown in Fig. 6J (top panel), the Bbp protein was expressed at levels comparable to the WT protein, suggesting that the stability of the mutant protein is not grossly affected. Expression of WT MAPKAPK2 induced robust phosphorylation of endogenous HSP27 (Fig. 6J, bottom panel). Unexpectedly, Bbp was significantly impaired in its ability to induce phosphorylation of HSP27, despite the fact that its kinase domain is intact, as well as the three predicted MAPK phosphorylation sites (Fig. 6A, J). These results indicate that the Bbp mutant protein is defective in its kinase activity, causing its loss-of-function phenotype.
To test directly if MAPKAPK2 kinase activity is required for it to regulate epiboly, we generated a myc-tagged full-length kinase dead protein by changing the critical lysine in the catalytic domain to a methionine (K73M) [34], and assayed activity in both cell culture and zebrafish embryos. As expected, kinase-dead MAPKAPK2 failed to induce phosphorylation of HSP27 in a HeLa cell culture assay (Fig. 6I, bottom panel). In bbp mutant embryos, injection of 200 pg of the kinase dead MAPKAPK2, although stably expressed in the embryo (Fig. 6K), failed to rescue the mutant phenotype (n = 99, Fig. 6B), in contrast to the WT protein (Fig. 6E). These results demonstrate that the kinase activity of MAPKAPK2 is required for its function in epiboly, indicating that the Bbp mutant protein fails to phosphorylate a critical target in the yolk cell that regulates epiboly.
The carboxy-terminal region truncated in Bbp also contains a p38 docking site, which is important for p38 to phosphorylate MAPKAPK2 efficiently [35],[36]. The lack of kinase activity of the Bbp mutant protein may be due to failure of p38 to efficiently phosphorylate and thus activate the Bbp protein. To investigate if p38 regulates MAPKAPK2 activity, we mutated the three p38 phosphorylation sites of MAPKAPK2 [29], [30], [37]–[39]. Based on the mammalian MAPKAPK2 structure, zebrafish MAPKAPK2 is expected to be phosphorylated on Threonine 202, Serine 252, and Threonine 315 by p38 MAPK (Fig. 6A). We mutated these three residues of zebrafish MAPKAPK2 to Alanines, any two of which when mutated in cell culture cause a failure in MAPKAPK2 activation [37]. In contrast to these cell culture studies, we found that the T202A/T315A double mutant MAPKAPK2 rescued bbp mutant embryos (40 pg, 15/16 rescued; 90 pg, 47/47 rescued). However, injection of 150 pg of the triple phospho-mutant RNA, although stably expressed (Fig. 6H, K), failed to rescue the mutant phenotype (n = 70), indicating the importance of these three sites in MAPKAPK2 function in epiboly. We also investigated the activity of the MAPKAPK2 phosphorylation site mutants in our HeLa cell culture assay and found similar results to the bbp mutant rescue data (Fig. 6I).
Thus, our results show that Ser252 is sufficient in the zebrafish embryo and HeLa cells for MAPKAPK2 protein function, contrasting previous studies in other cell culture systems. Considering that the three p38 phospho-residues are conserved in all MAPKAPK2 genes, we postulate that one or more of these residues is required to activate MAPKAPK2 depending on the cellular context. Taken together, we postulate that the lack of kinase activity of the truncated Bbp MAPKAPK2 is due to a loss of the p38 docking site in the Bbp mutant protein, resulting in an inability of p38 to bind, phosphorylate, and thus activate the mutant protein.
To test directly if p38 MAPK could be the upstream activator of MAPKAPK2, we investigated p38 function in epiboly progression by expressing a dominant negative p38a (DNp38) in WT embryos. p38a is expressed maternally and throughout blastoderm stages in the zebrafish [40],[41]. Microinjection of 250 pg of DNp38a mRNA caused 60% of the embryos to display a phenotype similar or identical to that of bbp mutant embryos (Fig. 7A,B, n = 155). In time-lapse microscopy analysis shimmying movements were observed in DNp38-injected embryos, similar to those seen in bbp mutants. Likewise, at or slightly after 50% epiboly, the blastoderm margin constricted strongly, followed by the yolk cell bursting (Video S6). These results support a role for p38 MAPK in regulating the activity of MAPKAPK2 in epiboly in the zebrafish.
The coordination of cell movements during developmental processes such as epiboly is not well understood. Although a fundamental cell movement process, few molecular components regulating epiboly have been described. Those that have been described are isolated components, not yet integrated into gene pathways. Here we identified a novel function for MAPKAPK2 and p38 MAPK in modulating this morphogenetic process in the early zebrafish embryo. MAPKAPK2 is required to prevent the premature constriction of the blastopore observed in bbp mutants. p38 MAPK is also required in this process, since mutating the three p38 MAPK phosphorylation sites in MAPKAPK2 abrogates its function and a dominant negative p38 MAPK phenocopies the bbp mutant phenotype. These results indicate that p38 MAPK regulates MAPKAPK2 activity during epiboly in zebrafish through the well-known MAP kinase cascade pathway.
Loss of MAPKAPK2 or p38 MAPK function causes premature constriction of the blastoderm margin, which ruptures the yolk cell and causes lethality. We can block yolk cell rupture by incubation in hypertonic media (Fig. 3 L); however, the marginal constriction persists in these conditions, indicating that the rupture and constriction are not simply due to a defect in yolk cell membrane integrity. We show that MAPKAPK2 functions within the yolk cell and postulate that it modulates actin-based contractility to close the blastopore during epiboly. The blastopore is at its greatest circumference at 50% epiboly (Fig. 1). Following that stage, the blastoderm cells migrate uniformly vegetally causing the blastopore circumference to continuously decrease in a purse-string like fashion until the blastopore is completely closed at 100% epiboly. Electron microscopy studies in Fundulus reveal an actin ring in the YSL adjacent to the blastoderm margin that is postulated to act as the strings during blastopore closure [15]. In zebrafish an F-actin band is first evident at 50% epiboly in a similar location in the YSL [16]. This actin band is associated with an active form of myosin, phosphorylated-myosin 2 [17], indicating the presence of an actin-myosin contractile activity in the YSL margin. This actin band is also associated with endocytosis [16], which removes the yolk membrane as the advancing blastoderm/EVL cells move vegetally over the yolk during epiboly and may be a driving force in their vegetal movement. The most intense region of this F-actin band associates with the leading edge of the EVL cells and is implicated, together with the Misshapen kinase in regulating the constriction of the marginal edge of EVL cells as they advance vegetally [17].
Pharmacological inhibitors of actin or myosin can slow the later stages of epiboly in zebrafish, implicating actin function in epiboly progression [16]. Higher doses of these inhibitors can arrest epiboly, but also cause either a dissociation of the EVL and blastoderm cells or yolk herniation due to loss of the vegetal actin mat that maintains the yolk integrity [16],[42]. We tested several actin or myosin inhibitors (cytochalasin B and D, Latrunculin A and B, blebbistatin) for their ability to suppress the marginal constriction in bbp mutants at 50% epiboly, however, none of them can suppress the phenotype at doses that slow epiboly in WT embryos (data not shown). Higher doses of some inhibitors can arrest epiboly in WT embryos and block the very strong constriction in bbp mutants; however, due to the many other roles that actin plays in development, including cytokinesis, cell adhesion, and general cell integrity, higher doses arrest development in general and cause lethality (data not shown), precluding our ability to block specifically the yolk cell constriction.
Loss of Mtx2, a predicted transcription factor, by morpholino knockdown in zebrafish results in reduction of the YSL punctate actin band and arrest in the vegetal movement of cells at the 50% epiboly point [22],[23]. Interestingly, mtx2 morphants also constrict around the margin with a very similar phenotype to bbp mutants [23]. However, the temporal progression of the constriction in mtx2 morphants coincides with the normal timing of blastopore closure in a WT embryo. That is, the margin constricts as if epiboly were progressing normally, although the cells fail to move vegetally. This contrasts the marginal constriction in bbp mutant embryos, which occurs rapidly during a 20 to 30-minute window, rather than over a 3-hour time period. The reduced punctate actin band in mtx2 morphants may result in epiboly arrest. While the remaining strong F-actin at the EVL margin may mediate the marginal constriction observed in mtx2 morphants and may be precociously activated in bbp mutant embryos.
In zebrafish, endogenous calcium release activity, as well as a requirement for calcium during epiboly, supports the importance of calcium signaling in epiboly progression [16]. Calcium levels are low during early epiboly [24]–[26],[43], but increase and become dynamic from 50 to 100% epiboly. Spikes of calcium are evident in the yolk cell beginning at 50% epiboly, followed by waves of calcium that traverse the blastoderm margin from 65% epiboly to blastopore closure [16],[24]. Loss of calcium causes a loss of the yolk cell actin band and a blockage in epiboly progression [16]. Considering that calcium positively regulates actin contraction [44], the dramatic increase in calcium release observed in bbp mutant embryos (Fig. 4) is consistent with increased actin contraction causing the abnormal morphogenesis movements. During early epiboly when calcium release is normally infrequent, we observed striking calcium dynamics in bbp mutants, coincident with the abnormal shimmying movements observed prior to 50% epiboly, suggesting abnormal F-actin contractile movements prior to 50% epiboly. Furthermore, we observed increased and sustained levels of calcium at the margin, when morphological constriction is apparent. The constriction phenotype is remarkable in its precise timing at specifically the 50% epiboly point in all mutant embryos, coincident with the timing of robust F-actin band formation at the YSL margin. We postulate that only when the F-actin band fully forms at 50% epiboly in conjunction with EVL marginal F-actin does the abnormal calcium regulation cause lethality through an unregulated F-actin constrictive force.
One well characterized target of MAPKAPK2, HSP27, plays a positive role in actin polymerization when phosphorylated by MAPKAPK2 [45]–[49]. Our studies suggest that actin polymerizes normally in bbp mutants, since the yolk cell actin band forms at 50% epiboly in bbp mutants. To our knowledge, there are no known MAPKAPK2 targets that inhibit actin-myosin contraction or calcium release. Thus, our results suggest a novel target of MAPKAPK2 that normally restricts sustained actin constriction to regulate tissue morphogenesis.
Although well studied in cell culture, MAPKAPK2 has been little studied in model organisms. While a MAPKAPK2 gene exists in both Drosophila and C. elegans, no mutant alleles have been reported. RNAi screens in Drosophila cell culture suggest a role for the fly homolog in cell cycle progression and cell shape regulation [50],[51].
In the mouse a targeted mutation of MAPKAPK2 is viable and fertile, but exhibits defects in mediating inflammatory responses [52]–[54]. Double mutants of MAPKAPK2 and the closely related MAPKAPK3 in the mouse display more severe defects in the inflammatory response, but do not exhibit developmental abnormalities, although they are widely expressed in development [55]. The third less related subfamily member, MAPKAPK5, exhibits an incompletely penetrant embryonic lethal phenotype in the mouse [56]. Generation of double and triple mutants of these MAPKAPKs will be required to reveal potential overlapping functions. Zebrafish have a duplicate MAPAPK2 gene that is not expressed until 3 dpf (E. Brito and DSW, unpublished observation). Zygotic roles for MAPKAPK2 in zebrafish development may be masked by the duplicate homolog or by the activity of homologs of the other family members MAPKAPK3 and/or MAPKAPK5, since MAPKAPK2/bbp homozygous zygotic mutants in zebrafish are viable to adulthood, with no obvious developmental defects.
The dramatic cell movements driving epiboly are crucial to the development of the body plan of anamniote vertebrates. We show that the p38 MAPK pathway is a critical component in regulating this process within the teleost yolk cell. Future studies will be required to reveal the mechanism by which this pathway regulates blastopore closure in this fundamental cell movement process. The early requirement for MAPKAPK2 and the accessibility of the zebrafish embryo will provide an excellent in vivo model for investigating the function and regulation of MAPKAPK2, which has primarily been studied in cell culture. In particular understanding the role of MAPKAPK2 during zebrafish epiboly will be valuable in the identification of MAPKAPK2 inhibitors to modulate the inflammatory response that MAPKAPK2 mediates in chronic inflammatory conditions in humans.
The animal work in this study was approved by the Institutional Review Board of the University of Pennsylvania School of Medicine.
All analyses were performed with the bbpp58cd allele [18]. Optimized fixation protocols for the cytoskeletal proteins were followed. Tubulin was visualized using KMX-1 (Boerhinger Mannheim) and F-actin with rhodamine-phalloidin (Molecular Probes) as described [57]. Myc-tagged protein was visualized by fixing embryos in 4% PFA in PBS, blocking with 5% BSA and 0.5% Tween-20 in PBS and staining with anti-myc monoclonal antibody 9E10 (University of Pennsylvania, Cell Center Service Facility), followed by confocal microscopy. In situ hybridization was performed using gsc and no tail [58], eve1 [59], bmp4 [60], and foxb1.2 (formerly fkd3) [20].
Endocytosis was analyzed by incubating manually dechorionated embryos in 1% rhodamine-dextran (MW 10,000, Molecular Probes) in E3 medium from sphere stage to 50% epiboly as described [8],[61]. Fixed embryos were analyzed by confocal microscopy.
For yolk cell membrane integrity studies, WT and bbp embryos were incubated in hypertonic Ringers solution (116 mM NaCl, 2.9 mM KCl, 0.8 mM MgSO4, 1.8 mM CaCl2, 5 mM HEPES) from the 4-cell stage through 24 hpf.
Whole blastoderm transplants were performed as previously described [27]. Briefly, mutant and WT embryos were injected at the one-cell stage with 1 nl 2.5% rhodamine-dextran (MW 10,000, Molecular probes) and 1 nl 0.05 mM Sytox Green (Molecular Probes), respectively, in 0.1 M KCl. Embryos were dechorionated and kept in E3 medium until the 1000-cell stage. Transplants were performed in 1× Ringers solution (116 mM NaCl, 2.8 mM Kcl, 5 mM HEPES, 1 mM CaCl2, pH 7.2) containing 1.6% whipped and cleared chicken egg whites. Blastoderms were separated from yolks using a glass knife and hybrid embryos formed by placing blastoderms onto naked yolks using slight pressure until adhered, then transferred to 1/3× Ringers solution after 15 minutes and analyzed.
Genomic DNA was pooled from mutant and WT sibling females, and the bbp mutation, p58cd [18], mapped to a chromosomal position using SSLP markers spaced throughout the genome, as described [19],[28]. SSLP markers flanking the mutation (z22766 and z22355) were used to genotype individual fish. Fish were generated for fine mapping by crossing bbp−/+ females to bbp−/− males. Both bbp−/+ and bbp−/− females were scored for recombination between markers z22766 and z22355.
cDNA was made from ovary RNA purified from bbp−/− and bbp+/+ fish using Superscript II (Invitrogen). MAPKAPK2 (MK2) cDNA was amplified using the primers: 5′-CCATCGATGGGTGTTGCCAAAGAAAGAC-3′ and 5′-GCTCTAGATCCACCGAGTTATTGCTTCC-3′. The product was sequenced and cloned into Cla1 and Xba1 sites in pCS2+. Full-length WT and bbp MK2 were cloned into a pCS2+ vector containing six tandem copies of a myc epitope (gift of Dr. Peter Klein), yielding an N-terminal myc-fusion protein. Kinase-dead myc-MK2 (K73M) was made by site-directed mutagenesis using the primers: 5′-GTGGGGAGAAGCTCGCTTTAATGATGCTTCATGACTGCCCAAA-3′ and 5′-TTTGGGCAGTCATGAAGCATCATTAAAGCGAGCTTCTCCCCAC-3′. Non-phosphorylatable myc-MK2 (T202A, S252A, T315A) was made by site-directed mutagenesis using the following primers in three subsequent reactions: 5′-ACACACAACTCTCTGGCCGCCCCCTGCTATACTCCTTATTAT-3′ with 5′-ATAATAAGGAGTATAGCAGGGGGCGGCCAGAGAGTTGTGTGT-3′ (T202A), 5′-GAATCATGGATTGGCAATTGCTCCTGGTATGAAGAAACGAAT-3′ with 5′-ATTCGTTTCTTCATACCAGGAGCAATTGCCAATCCATGATTC-3′ (S252A), and 5′-CAATCAATGGAGGTTCCACAGGCACCCCTACACACCAGCCGT-3′ with 5′-ACGGCTGGTGTGTAGGGGTGCCTGTGGAACCTCCATTGATTG-3′ (T315A). DNp38a (T181A, Y183F) was generated from a WT p38a cDNA in pCS2+ (gift of T. Hirano) by site directed mutagenesis with the primers 5′-GACACACAGATGATGAGATGGCCGGCTTTGTGGCCACAAGGTGGTATC-3′ 5′-GATACCACCTTGTGGCCACAAAGCCGGCCATCTCATCATCTGTGTGTC-3′. Capped mRNA was produced from pCS2+ constructs with mMessage mMachine (Ambion) and injected into embryos, as previously described [62]. A MK2 morpholino, GTTGGCGTTAGTCAACATCTCCCAC (Gene Tools, Philomath, Oregon) was injected at 5 ng/nl in 0.1 M KCl at the one-cell stage.
For in situ hybridization of MK2, a probe was generated from the pCS2+MK2 vector digested with HinDIII and transcribed with T7 polymerase. SacII digestion, followed by SP6 polymerase synthesis generated the sense control probe.
HeLa cells were grown in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% fetal bovine serum (FBS), penicillin and streptomycin, and GlutaMax (Gibco BRL) to a density of 2×105 per 35 mm plate at 37°C in 5% CO2. Cells were transfected with tagged and untagged MK2 WT and mutant plasmids, as described [63]. Cell lysates were analyzed by a standard Western blot protocol using anti-MK2, anti-HSP27, anti-Phospho-HSP27 (Assay Designs), and anti-Phospho-MK2 (Cell Signaling). Zebrafish embryonic proteins were resolved as described [64] with each lane containing 2 embryo equivalents. Western blotting was performed with anti-myc monoclonal antibodies.
Still images of live embryos and in situ hybridization were captured with iVision (DVL Software). For time lapse-microscopy, embryos were immobilized in individual chambers in 1% agarose in E3 and covered by 3% methylcellulose in E3. Time-lapse movies were created using OpenLab (Improvision, Beverly, MA). Confocal analysis of tubulin and phalloidin staining, endocytosis, and MAPKAPK2 localization was performed using a Zeiss confocal and LSM510 software. For in vivo calcium imaging, 1-cell stage embryos were microinjected with Fura-2 Dextran or Bis-Fura2 (Molecular Probes) and imaged as described [65]. Image pairs were collected at 15-second intervals through epiboly stages. For rescue of the calcium defect, MAPKAPK2 RNA (90 ng/ul) co-mixed with Texas Red (TxR) or TxR alone as a control was injected in the yolk cell below the blastoderm at the 512-cell stage. Periodically frames were collected at 560 nm to locate the TxR lineage marker and in vivo calcium imaging performed as above. For YSN labeling, embryos were injected in the YSL at 1000-cell stage with 1 nL of 0.25 mM Sytox Green in 0.1 M KCl. Embryos were mounted in 0.12% low melt agarose. Images were acquired on a Zeiss Axiovert 200 and processed in Axiovision software. For time-lapse movies, images were processed in Photoshop, ImageJ and Quicktime Pro.
|
10.1371/journal.pntd.0004084 | Discrimination between E. granulosus sensu stricto, E. multilocularis and E. shiquicus Using a Multiplex PCR Assay | Infections of Echinococcus granulosus sensu stricto (s.s), E. multilocularis and E. shiquicus are commonly found co-endemic on the Qinghai-Tibet plateau, China, and an efficient tool is needed to facilitate the detection of infected hosts and for species identification.
A single-tube multiplex PCR assay was established to differentiate the Echinococcus species responsible for infections in intermediate and definitive hosts. Primers specific for E. granulosus, E. multilocularis and E. shiquicus were designed based on sequences of the mitochondrial NADH dehydrogenase subunit 1 (nad1), NADH dehydrogenase subunit 5 (nad5) and cytochrome c oxidase subunit 1 (cox1) genes, respectively. This multiplex PCR accurately detected Echinococcus DNA without generating nonspecific reaction products. PCR products were of the expected sizes of 219 (nad1), 584 (nad5) and 471 (cox1) bp. Furthermore, the multiplex PCR enabled diagnosis of multiple infections using DNA of protoscoleces and copro-DNA extracted from fecal samples of canine hosts. Specificity of the multiplex PCR was 100% when evaluated using DNA isolated from other cestodes. Sensitivity thresholds were determined for DNA from protoscoleces and from worm eggs, and were calculated as 20 pg of DNA for E. granulosus and E. shiquicus, 10 pg of DNA for E. multilocularis, 2 eggs for E. granulosus, and 1 egg for E. multilocularis. Positive results with copro-DNA could be obtained at day 17 and day 26 after experimental infection of dogs with larval E. multilocularis and E. granulosus, respectively.
The multiplex PCR developed in this study is an efficient tool for discriminating E. granulosus, E. multilocularis and E. shiquicus from each other and from other taeniid cestodes. It can be used for the detection of canids infected with E. granulosus s.s. and E. multilocularis using feces collected from these definitive hosts. It can also be used for the identification of the Echinococcus metacestode larva in intermediate hosts, a stage that often cannot be identified to species on visual inspection.
| The canid adapted intestinal tapeworms, Echinococcus granulosus, E. multilocularis and E. shuiqucus are well known to be endemic in Northwestern China. The first two species can cause fatal disease in humans. Although E. shiquicus has not been reported to infect humans, all three species can be transmitted by dogs. The very close relationship between dogs and humans can readily lead to human infection. To aid the surveillance and management of echinococcosis, effective diagnostic approaches are urgently needed. We developed a single tube multiplex PCR assay for the accurate identification and discrimination of the three Echinococcus species for use in both clinical diagnosis and epidemiological studies.
| In the most recent taxonomic revision, nine species were recognized in the genus Echinococcus [1]. Of these, the most important and widespread are E. granulosus sensu stricto (genotypes G1-G3) and E. multilocularis, which cause cystic echinococcosis (CE) and alveolar echinococcosis (AE), respectively. The former is commonly associated with livestock and human infections worldwide whereas the latter is primarily found in voles and humans and is geographically limited to the northern hemisphere [2]. To date, E. granulosus s.s., E. canadensis (G6), E. multilocularis and E. shiquicus have been identified in China [3–5]. Both E. multilocularis and E. granulosus s.s. are particularly widespread in western China, including Qinghai, Ningxia, Gansu, Xinjiang and Sichuan provinces/autonomous regions, and are well known as major public health and medical threats. Unlike the other species, E. shiquicus has a very restricted distribution, being reported only from Qinghai Province, China. This species is not known to cause human echinococcosis. The intermediate hosts are plateau pikas (Ochotona curzoniae), in which unilocular cysts occur.
For Echinococcus species in general, dogs, wolves, other canids and cats are definitive hosts in which adult worms cause sub-clinical infections [6–9]. However, larval Echinococcus spp. can cause morbidity and mortality in their intermediate hosts which include cattle, sheep, small mammals (including rodents, plateau pikas, etc.) and humans [10, 11]. It can be difficult to discriminate morphologically adults of some Echinococcus species, such as E. multilocularis and E. shiquicus [12].
To replace traditional morphological methods, a number of molecular approaches targeting parasite DNA have been developed for identification/discrimination of different life stages of Echinococcus species in definitive and intermediate hosts [13–15]. Multiplex PCR approaches, simultaneously using multiple specific primers in a single tube and detecting more than one target species, are material- and time-saving, precise, efficient and cost-effective when DNA from a mixture of pathogens may be present in a sample. This approach is also suitable for mass-screening of samples that may be generated from epidemiological investigations in endemic areas. Several multiplex PCR methods have been developed for identifying certain Echinococcus species, but none for the identification of E. shiquicus [16–17].
Based on interspecific variation in mitochondrial genes of the genus Echinococcus, we designed a multiplex PCR assay with three pairs of specific primers in a single reaction tube for rapid identification of E. granulosus s.s., E. multilocularis and E. shiquicus originating from either intermediate or definitive hosts. Further assessment of the sensitivity and specificity of the multiplex PCR assay was performed using metacestode DNA and copro-DNA to determine the reliability and accuracy of the new diagnostic tool developed in this study.
Dogs and mice used in this study were handled in strict accordance with good animal practice according to the Animal Ethics Procedures and Guidelines of the People's Republic of China (Regulations for Administration of Affairs Concerning Experimental Animals, China, 1988). No endangered/protected species were involved in this study. The dogs and mice used were also treated in accordance with the institutional procedures and guidelines for animal husbandry issued by the Ethics Committee of Lanzhou Veterinary Research Institute, Chinese Academy of Agricultural Sciences (Approval No. LVRIEC2010-005).
Adult worms were collected from stray dogs during routine work of the endemic echinococcosis prevention and control program in Dari County, Qinghai Province, P.R. China. A total of 86 Echinococcus spp. metacestode samples from yaks, sheep, Qinghai voles (Microtus/Neodon fuscus) and plateau pikas were collected on the Qinghai-Tibet plateau, P.R. China. Ten yak lungs and 16 sheep livers harboring hydatid cysts were collected from abattoirs in Maqu County, Gansu Province and Xining City, Qinghai Province, respectively. Thirty Qinghai vole livers and 30 plateau pika lungs harboring hydatid cysts were provided by the epidemic prevention station of Dari County, Qinghai Province.
Parasite materials were dissected from the host tissue and stored either in 70% ethanol before molecular analyses, or temporarily stored at 4°C prior to experimental infections of dogs.
Fifteen dogs (mixed breeds) aged 6–8 months were purchased in Lanzhou City, Gansu Province, China. These were de-wormed using praziquantel and confirmed to be free of intestinal parasites by examination of their feces two weeks later. Samples of these feces were retained as negative controls for the multiplex PCR assay. Live protoscoleces (100,000) of each Echinococcus spp. were fed independently to five dogs after their viability for dog challenge was confirmed by microscopy.
Dogs were euthanized three months after challenge with protoscoleces. Fecal samples were collected from the dogs each day prior to sacrifice. After removal of the coarse gut contents, the small intestine was cut into 15–20 cm lengths and opened to expose the mucosa. Samples, taken by scraping the mucosa with glass strips, were placed in petri dishes in bio-safety containers [18]. After addition of a small volume of sterile phosphate-buffered saline (PBS, pH 7.2), the contents were checked for the presence of worms (intact or fragmented) and/or eggs. Adult worms were removed using a glass needle and washed in PBS three times. All procedures were performed following appropriate bio-safety conditions [19].
Ten stray dogs, provided by the epidemic prevention station in Dari County, Qinghai Province, were processed as above to obtain mucosal samples, worms and eggs. Additionally, five fecal samples from captive foxes were collected from a fur farm in Lanzhou City, Gansu Province. All the collected fecal samples were frozen at -80°C for at least seven days for bio-safety reasons. Worm samples were preserved either in 70% ethanol or frozen at (-80°C) in PBS for further analyses.
DNA samples, extracted from a variety of cestodes (identities confirmed by sequencing and morphology), were used to determine the specificity of the newly developed multiplex PCR assay (Table 1). They were kindly provided by the Key Laboratory of Veterinary Parasitology of Gansu Province, Lanzhou Veterinary Research Institute, CAAS.
DNA extracted from host tissues was used to check for nonspecific reactions or assay interference that might be caused by contamination of parasite samples with host DNA. Host tissues included dog intestines, and liver and lung samples from cattle, sheep, Qinghai voles and plateau pikas.
Two hundred mg of each metacestode sample was frozen in liquid nitrogen and ground to powder after removal of ethanol or PBS by rinsing with ddH2O. Total genomic DNA was extracted using a QIAGEN DNeasy Blood & Tissue Kit (QIAGEN, Hilden, Germany) according to the manufacturer’s instructions and stored at -20°C until use.
To minimize the impact of inhibitors on PCR using copro-samples as template, an additional step of stool flotation in saturated zinc chloride solution was used before copro-DNA extraction [20]. Briefly, about 20 g (20 ml) fecal material was placed in a 50 ml centrifuge tube, which was then filled with zinc chloride solution. The tube was vortexed until the fecal material was completely broken up. The tube was then centrifuged at 1000 ×g for 5 min. Five hundred μl of the supernatant (usually containing helminth eggs, proglottids or cells of parasites) was transferred to a 2 ml centrifuge tube, 1.5 ml ddH2O was added to dilute the solution, and the tube was centrifuged at 12,000 ×g for 10 min. The supernatant was carefully discarded and 200 μl ddH2O added to suspend the sediment for DNA extraction. Total genomic DNA was extracted using a QIAGEN QIAamp DNA Stool Mini Kit (QIAGEN, Hilden, Germany) following the manufacturer’s instructions, and the DNA concentration was determined using a spectrophotometer (Thermo, NanoDrop 2000, USA) after elution in 50 μl ddH2O for use in the PCR assay.
Genomic DNA was extracted from host tissues using a QIAGEN DNeasy Blood & Tissue Kit (QIAGEN, Hilden, Germany), according to the manufacturer’s instructions, and stored at -20°C until use.
The complete mt genomes (mtDNA) of various cestodes (Table 1) available in GenBank (http://www.ncbi.nlm.nih.gov/) were retrieved to facilitate design of primers specific for E. granulosus s.s., E. multilocularis and E. shiquicus (Table 1). The sequences were aligned automatically using Clustal in MEGA5.0 [21]. Primer pairs, expected to be specific for E. granulosus s.s. (S1 Fig), E. multilocularis (S2 Fig) and E. shiquicus (S3 Fig), were thus obtained. After some preliminary experimentation, one pair of primers specific for each Echinococcus spp. was selected for inclusion in the multiplex PCR assay. Sequences of these primers, target genes and other related information are presented in Table 2.
PCR amplification was carried out in a 25 μl mixture containing 2 μl dNTPs (2.5 mM of each), 2.5 μl 10× ExTaq Buffer (Mg2+ free), 2 μl MgSO4 (25 mM), 0.25 μl ExTaq DNA polymerase (5U/μl) (TaKaRa, Dalian, Liaoning), 100 pg DNA template of each Echinococcus sample, and all three primer pairs were added according to the final concentrations given in Table 2. Fragments were amplified using the following optimized thermocycling conditions: 95°C/5 min for denaturation; 30 cycles of 94°C/30 sec, 55°C/30 sec, 72°C/40 sec; and 72°C/10 min extension. For all the multiplex PCR assays, positive DNA (DNA templates of the three Echinococcus spp.) and negative (no-DNA) controls were included.
Amplicons were visualized by electrophoresis in 2.0% (w/v) agarose gels in 1×TAE (40 mM Tris-acetate, 2 mM EDTA, pH 8.5), stained with ethidium bromide (EB), and viewed under UV light. The fragments were purified using an agarose Gel DNA Purification Kit (TaKaRa, Dalian, Liaoning), and then cloned into pMD18-T Simple vectors using a TA cloning strategy. The recombinant vectors were identified by enzyme digestion and at least two clones for each DNA region were sequenced by the Shanghai Invitrogen Biotechnology Co. Ltd.
Infections of E. granulosus s.s. and E. multilocularis were successfully achieved in all the experimentally infected dogs with 5539, 8562, 12535, 18932 and 20775 E. granulosus s.s. and 2893, 3153, 3762, 3864 and 5322 E. multilocularis adult worms being recovered from each group of 5 dogs that were fed with protoscoleces of each species. No adult worms were found in any of the 5 dogs fed larval E. shiquicus. None of the stray dogs was found harboring E. shiquicus or E. multilocularis; only E. granulosus s.s. adult worms were found in their intestinal contents (identity confirmed by both morphology and cox1 sequencing). Worm burdens were relatively low (circa 100–200 worms) in the ten stray dogs examined.
Expected PCR products of 219, 584 and 471 bp were obtained for E. granulosus s.s. (nad1), E. multilocularis (nad5) and E. shiquicus (cox1), respectively (Fig 1), and products of mixed templates of the three Echinococcus species are shown in Fig 2. The multiplex PCR products contained 3 DNA bands (219, 471 and 584 bp) with mixed DNA templates of E. granulosus s.s., E. multilocularis and E. shiquicus; 2 DNA bands (219 and 584 bp) with E. granulosus s.s. and E. multilocularis DNA templates; 2 DNA bands (219 and 471 bp) with E. granulosus s.s. and E. shiquicus DNA templates; and 2 DNA bands (471 and 584 bp) with E. multilocularis and E. shiquicus DNA templates. DNA sequences of these products corresponded in each case with the relevant reference sequences in GenBank: E. granulosus (G1) (NC_008075) [22], E. multilocularis (NC_000928) [23] and E. shiquicus (NC_009460) [24].
China is the most severe pandemic country for cystic echinococcosis (CE), in humans and livestock, due mainly to E. granulosus s.s., and for alveolar echinococcosis (AE) due to E. multilocularis in humans and small wild mammals. E. shiquicus is also endemic although it has not been reported to infect humans. Dual infections of animal hosts with different Echinococcus spp have been reported in the eastern Qinghai-Tibet plateau region of China [4, 25]. The very close relationship between dogs and humans can lead readily to human infection. The increasing number of human AE and CE cases in northwestern China, where considerable numbers of dogs are present, places a heavy burden on public health and veterinary services. To aid surveillance, management and diagnosis, effective methods are needed for rapid and accurate detection and identification of different life cycle stages of the three Echinococcus spp. simultaneously. The multiplex PCR assay developed in this study provides such a method.
Traditional epidemiological surveys for tapeworms often involve recovery of eggs from feces of potential definitive hosts. However, morphological identification of Echinoccocus eggs to species level is practically impossible, prompting the development of several molecular approaches [26, 27]. Inhibitors present in fecal material that co-purify with parasite DNA extracted from feces often present a problem for PCR-based methods [28]. In this study, the QIAGEN QIAamp DNA Stool Mini Kit, containing InhibitEX tablets for removing inhibitors in fecal samples, was used to purify copro-DNA. The sieving-flotation method was helpful in overcoming this problem due to its enrichment of worm eggs [29]. The positive control (protoscolex DNA in fecal samples) used in this study demonstrated the lack of inhibitor effects in our copro-multiplex PCR assay.
E. granulosus s.s. has been reported as having a pre-patent period of 6 weeks (42 days) [30, 31], while E. multilocularis eggs have been observed in feces at 42–46 days post infection [32]. However, in the current study we first identified eggs of E. granulosus s.s. at 47–56 days post-challenge and those of E. multilocularis at 36–44 days post-challenge by microscopy similar to reports by others [30, 33]. The discrepancies between these studies may be due to the use of different dog-breeds, ages, nutrient status or the conditions under which the dogs were maintained. We were unable to experimentally infect dogs with E. shiquicus although the viability of the challenge sample of protoscoleces was confirmed by microscopy.
PCR-positive signals in this study were obtained from dog fecal samples much earlier (17 days for E. multilocularis and 26 days for E. granulosus) than any other previous studies using microscopy as a method of detecting infected canid hosts. The much earlier detection of an Echinococcus infection by the multiplex PCR method compared with egg recovery from feces and microscopic examination is a marked improvement that can aid surveillance programs aimed at preventing echinococcosis transmission.
The method developed in this study has achieved high species specificity because it produced no amplicon from any other helminth (including several that might dual infect with Echinococcus species in dogs) or from the negative copro-samples (no-DNA). The primer set (three pairs of primers) multiplex reaction in a single tube worked well with all templates tested and yielded specific amplicons of the expected length for each of the three Echinococcus spp. examined.
E. granulosus s.s and E. multilocularis are of major public health concern in many endemic countries globally [34]. A cost effective diagnostic tool is required for echinococcosis surveillance of definitive and intermediate hosts, and for monitoring the effectiveness of control programs. The multiplex PCR assay developed in this study provides an effective method that can be applied in both clinical and epidemiological settings for the identification of Echinococcus spp in diverse hosts, and would be particularly useful for identifying infected hosts in areas co-endemic for AE and CE.
In this study, we focused on Echinococcus samples collected from the Qinghai-Tibet plateau region of China, where three species (E. granulosus, E. multilocularis and E. shiquicus) are known to be endemic. In total, nine species are now recognized in the genus Echinococcus, including E. granulosus sensu stricto (genotypes G1-G3), E. equinus, E. canadensis (genotypes G6, G7, G8 and G10), E. ortleppi, E. multilocularis, E. shiquicus, E. vogeli, E. oligarthrus and E. felidis [1]. None of the three specific pairs of primers developed in this study produced a PCR-amplified product using DNA isolated from E. canadensis (G6 genotype) showing in Fig 3 (the lane 2 with non-band as a negative result). This is supported by inspection and comparison of the primer target sequence for the G6 genotype with those of the three Echinococcus spp., which showed six base pair differences between them (S1 Fig, S2 Fig and S3 Fig in the Supporting Supplementary Information). Furthermore, six or more base pair differences are apparent between the target sequences for E. equinus, E. canadensis (genotypes G7, G8 and G10), E. ortleppi, E. vogeli, E. oligarthrus and E. felidis. Therefore, it is highly unlikely that any amplicon would be produced from these species during the multiplex PCR due to its high species specificity.
|
10.1371/journal.pcbi.1000717 | Predicted Auxiliary Navigation Mechanism of Peritrichously Flagellated Chemotactic Bacteria | Chemotactic movement of Escherichia coli is one of the most thoroughly studied paradigms of simple behavior. Due to significant competitive advantage conferred by chemotaxis and to high evolution rates in bacteria, the chemotaxis system is expected to be strongly optimized. Bacteria follow gradients by performing temporal comparisons of chemoeffector concentrations along their runs, a strategy which is most efficient given their size and swimming speed. Concentration differences are detected by a sensory system and transmitted to modulate rotation of flagellar motors, decreasing the probability of a tumble and reorientation if the perceived concentration change during a run is positive. Such regulation of tumble probability is of itself sufficient to explain chemotactic drift of a population up the gradient, and is commonly assumed to be the only navigation mechanism of chemotactic E. coli. Here we use computer simulations to predict existence of an additional mechanism of gradient navigation in E. coli. Based on the experimentally observed dependence of cell tumbling angle on the number of switching motors, we suggest that not only the tumbling probability but also the degree of reorientation during a tumble depend on the swimming direction along the gradient. Although the difference in mean tumbling angles up and down the gradient predicted by our model is small, it results in a dramatic enhancement of the cellular drift velocity along the gradient. We thus demonstrate a new level of optimization in E. coli chemotaxis, which arises from the switching of several flagellar motors and a resulting fine tuning of tumbling angle. Similar strategy is likely to be used by other peritrichously flagellated bacteria, and indicates yet another level of evolutionary development of bacterial chemotaxis.
| Chemotaxis of bacteria plays an important role in their life, providing them with the ability to actively search for an optimal growth environment. The chemotaxis system is supposed to be highly optimized, because on the evolutionary time scale even a modest enhancement of its efficiency can give cells a large competitive advantage. For a long time it was believed that the only navigation mechanism of bacteria is increasing the run length toward the preferred direction. The tumble was assumed to be a purely random change of direction between runs. We analysed recently published experimental data that demonstrate a dependence of tumbling angle on the number of CW-switched motors. We introduced such a dependence into our model of chemotactic E. coli, and simulated it under different conditions. Our simulations show that this dependence is an important additional mechanism of bacterial navigation, which was previously unrecognized because it lays below the experimental errors of conventional single-cell tracking. We show that such a fine tuning of tumbling significantly improves efficiency of chemotaxis, and represents a new level of evolutionary optimization of bacteria.
| Many motile unicellular organisms are known to direct their movement in gradients of specific chemical substances – the process called chemotaxis. Chemotaxis plays an important role in the microbial population dynamics with chemotactic bacteria in a nonmixed environment – that is in presence of nutrient gradients – having significant growth advantage [1]–[4]. Modeling of microbial population dynamics indicates that motility and chemotactic ability can be as important for evolutionary competition as cell growth rate [5],[6]. The chemotaxis system is thus expected to be highly optimized, as has been indeed suggested by several studies [7]–[10].
The best example of such optimization is bacterial chemotaxis strategy itself. While eukaryotic cells are able to sense the gradients by direct comparison of concentrations at the opposite sides of the cell [11], bacteria like E. coli employ temporal comparisons along their runs [12]. Theoretical analysis suggested that such strategy is superior to direct spatial comparisons for objects of bacterial size and swimming speed [7]. Adapted E. coli has two swimming modes: runs, which are periods of long straight swimming, and tumbles, when bacterium stops and changes its orientation. The runs of a swimming bacterium are interrupted by tumbles which abruptly change the swimming direction. For cells swimming up an attractant gradient, the runs become longer due to suppression of tumbles, and the cell population migrates up the gradient. The frequency of tumbles is controlled by the chemotaxis network through switching of individual motors. During a run, flagellar motors rotate counter-clockwise (CCW) causing flagella to form a bundle, whereas switching of one or several flagellar motors to clockwise (CW) rotation breaks up the bundle and initiates a tumble. The direction of motor rotation depends on the concentration of phosphorylated CheY molecules, which bind to the motor and switch its direction in a highly cooperative mode. The CheY phosphorylation is controlled by the histidine kinase CheA, which forms sensory clusters together with transmembrane receptors and the adaptor CheW. Each receptor can be either active or inactive, depending on ligand binding and on the methylation level. The active receptor activates CheA, eliciting downstream phosphorylation of the response regulator CheY. Phosphorylated CheY (CheYp) is dephosphorylated by CheZ. Receptors can be methylated by the methyltransferase CheR and demethylated by the methylesterase CheB. Methylation regulates the receptor activity. Because the reaction of receptor methylation is much slower than the initial response, methylation provides chemical ‘memory’, which allows the cell to compare the current ligand concentration with the recent past.
Early single-cell tracking experiments reported no dependence of the tumbling angle, i.e. turning angle between consequent runs, on the direction of the gradient and the inclination of a run [12], and it was thus presumed to be random in subsequent modeling of bacterial chemotaxis. However, in recent study that used high-resolution fluorescence video microscopy [13], it was shown that the cell turning angle depends on the number of CW-rotating filaments involved in the tumble, and thereby the turning angle rises proportionally to the number of motors that switched to CW rotation. Because the CW switch probability is set by the chemotaxis system dependent on the cellular swimming direction along the gradient, the tumbling angle can be expected to depend on the swimming direction, too. If the cell swims up a gradient of attractant, the probability of CW rotation is smaller, and fewer motors are likely to change directions. Therefore, even if the cell makes a tumble, the tumbling angle should be small. When the cell swims down the gradient of attractant, the probability of CW rotation is higher and more motors are likely to change directions during a tumble, with the consequence that the tumbling angles will be larger.
The goal of this study was thus to investigate the magnitude of the tumbling angle dependence on the swimming direction and the effect of such dependence on the chemotactic efficiency. We introduced dependence of the turning angle on the number of CW-rotating motors in a recently constructed hybrid model of chemotactic E. coli, RapidCell simulator [14]. Our simulations demonstrate that although the estimated difference of tumbling angles up and down the gradient is only few degrees, even such a small difference significantly improves the chemotactic efficiency of E. coli. We thus suggest that tuning of tumbling angle depending on swimming direction serves as an additional navigation mechanism for E. coli and other peritrichously flagellated bacteria with similar chemotaxis behavior.
The tumbling angle dependence on the number of switching motors was investigated by extending the recently published hybrid model of chemotactic E. coli [14]. First, a more detailed model of tumbling was developed to bring the model in a closer agreement with the tracking experiments of [12]. While previous version of the model relied on a simple voting model of tumbling, which started the tumble as soon as the majority of motors rotate CW, our new model takes into account the duration of CW-rotation of every motor (Fig. 1A). The complex hydrodynamics of multiple flagella is described in simplified form, through a distortion factor which is a function of of each motor (see Methods). Despite this simplification, the simulated swimming of E. coli is in a very good agreement with the original tracking experiments [12]. The model realistically reproduces nearly all data provided by tracking experiments: mean cellular speed, run times, tumbling angles (Tab. 1), as well as individual motor switching and graduate recovery of cellular speed after a tumble.
Second, we introduced a dependence of tumbling angle on the number of CW-rotating motors that cause the tumble (Fig. 1B). This was done by fitting the experimental data of [13] with a realistic choice of discrete tumbling angles at each number of CW-switched motors (Fig. 1C). To ensure consistency with experimental data, we further assumed dependence of tumbling angle on the total number of motors. This model was called anisotropic, and it was compared to a conventional model of isotropic tumble, which chooses the tumbling angle stochastically. In simulations without a gradient, both models produce equal cellular drift velocities, with the accuracy of estimation error. To keep the mean angles of both models consistent, we defined the frequencies of the discrete angles in the anisotropic model as shown in Fig. 1D.
The model of swimming proposed here allows tumbling with variable number of motors, as soon as the sum of their CW-rotation times exceeds 0.15 s threshold needed for tumbling (). A cell swimming down the gradient will sooner reach the threshold, because each motor has higher probability of switching to CW. As a first consequence, the average run down the gradient will be shorter. As a second consequence of higher switching probability, the average number of motors that switch CW during that tumbling period will be higher than in case of up-gradient swimming. For example, cells with 3 motors when swimming down the gradient N1 tumble with motors while up the gradient with motors (means.e.m.).
Therefore, the tumbling angles for anisotropic model depend on the swimming direction prior to tumbles (Fig. 2A). This dependence naturally arises from the dependence of tumbling angle on the number of CW-rotating motors. The simulated cells which turned with the smallest were swimming in slightly skewed directions up the gradient before the tumble, whereas the cells which turned with the highest were swimming with even smaller skew down the gradient before the tumble. A more detailed analysis shows that the total angular difference between tumbling angles that correspond to the movement up and down a gradient is only about 3 (Fig. 2B). Such a small difference is within the error of the early tracking experiments, about [15], which explains why it remained undetected.
Despite such a small difference of mean angles, it can significantly increase the chemotactic performance, with the mean drift velocity being up to two times higher for anisotropically tumbling cells (Fig. 2C). The positive effect of anisotropic tumble becomes more visible in steeper gradients and for higher number of motors, which suggests that highly flagellated cells can adjust their tumbling angle more precisely.
In the case of motors and moderate gradient (N1), the mean tumbling angle is . This value is only smaller than the angle in ligand-free simulations, so the increase of the drift velocity in the anisotropic model cannot be attributed to the change of the total mean tumbling angle. The mean tumbling angle up the gradient , while down the gradient it is . Therefore, the difference in mean tumbling angles causes a 52% increase in the population drift velocity, from 0.92 to 1.4 (Fig. 2C).
As a control, we simulated chemotactic cells that tumble with a constant angle (67.5 deg.), and compared them to cells that tumble with slightly smaller angle (67.5−), when they swim up the gradient, and with slightly higher angle (67.5+), when they swim down the gradient. Here, the was a constant parameter changed from 1 to 5 deg. A difference of degrees increased the drift velocity by about 100% in the gradient N1, and by 50% in the gradient N2 (Fig. 3A). This confirms that the observed increase in drift velocity shown in Fig. 2C is due to small changes in tumbling angles of up- and down-swimming cells, and does not arise from model-specific parameters.
Bacterial movement in gradients is further affected by the Brownian motion for both isotropic and anisotropic tumbling models (Fig. 3B). In our simulations we used (Tab. 1). At lower coefficients of rotational diffusion, both models demonstrate better chemotaxis, and the advantage of the anisotropic tumbling is most pronounced, which is due to lower noise factor arising from rotational diffusion [16]. Since rotational diffusion depends on the cells size, flagellar length, media viscosity and temperature [17],[18], predicted effects of anisotropic tumbling can be even more pronounced for other bacteria or under different environmental conditions.
Taken together, our results suggest that in addition to extending the run length while swimming up the gradient, E. coli uses an auxiliary mechanism of tumbling angle tuning according to the swimming direction. This fine tuning of tumble is mediated by the same adjustment of tumbling frequency that underlies the conventional chemotaxis strategy of E. coli (Fig. 4). Since both navigation mechanisms arise from the same basic mechanism of altered motor switching, evolutionary optimization of the basic mechanism depends on both the effect from the tumble frequency and the number of flagella that reverse per tumble. The previously unrecognized mechanism shown here is expected to be shared by other peritrichously flagellated bacteria with similar chemotactic behavior, and it seems to represent yet another level of evolutionary optimization of the chemotaxis system.
We applied the recently proposed Monod-Wyman-Changeux (MWC) model for mixed receptor clusters [19],[20], which accounts for the observed experimental dose-response curves of adapted cells measured by in vivo FRET experiments [19],[21], as shown in [20],[22],[23]. According to the MWC model, an individual receptor homodimer is described as a two-state receptor, being either ‘on’ or ‘off’, with the free energy being a function of methylation level and ligand concentration (1)where is the ‘offset energy’, and , are the dissociation constants for the ligand in the ‘on’ and ‘off’ state, respectively. Groups of receptors form larger sensory complexes, or signaling teams, with all receptors in a team being either ‘on’ or ‘off’ together. The teams are composed of mixtures of Tar and Tsr receptors, and the total free energy of the team is given by(2)The probability (A) that a team will be active is a function of its free energy(3)
The adaptation is modeled according to the mean-field theory [24],[25], assuming that the CheB demethylates only active receptors, CheR methylates only inactive receptors, and both enzymes work at saturation(4)This equation implies that both enzymes work in the zero-order regime. The linear products and () mean that a bound CheR (CheB) can only act if the receptor team is inactive (active), with probability and , respectively.
The average methylation level is assumed to be a continuously changing variable within the interval , with linear interpolation between the key offset energies, , as suggested in [25],[26]. The ODE for methylation (Eqn. 4) is integrated using the explicit Euler method to ensure high computational speed of the program, while the time step is chosen as 0.01 s to keep the simulation error low.
The details of network model were previously described in [14]. CheA kinase activity is assumed to be equal to the activity of the receptor complex . The rate of phosphotransfer from active CheA to CheY is much faster than the rate of CheA autophosphorylation [9],[27]. Therefore, the relative concentration of CheYp is obtained as a function of active CheA from the steady-state equation(5)where is a scaling coefficient, , , are the rate constants according to [9],[28],[29].
The relative concentration of CheYp is converted into the CCW-motor bias using a Hill function [30]:(6)where [30], [30],[31].
To simulate the experimentally observed hydrodynamics of bacterial swimming and tumbling [13],[32] in simple terms, we introduce a distortion factor which reflects how one CW-rotating flagellum influences the cellular speed and angular deviation(7)This functional form implies that the distortion rises proportionally to the CW rotation time as long as it is below the threshold (the first period). After this threshold is reached, the distortion exponentially decays (the second period). The first period corresponds to unwinding of a flagellum from the bundle and its rotation in the right-handed semicoiled form, which initiates a tumble. In the second period, when the flagellum rotates CW longer than the threshold time, a rapid transformation from semicoiled to curly 1 form occurs, and the flagellum twists around the bundle during the new run, due to high flexibility of the latter form [32].
The influence of several simultaneously CW-rotating motors is assumed to be proportional to the sum of their distortion factors(8)This implies that the tumble can occur if a single motor rotates CW for at least period, or if two or more motors rotate CW together for a shorter time. Formally, a tumble occurs when , where is a threshold value. In principle, the threshold depends on the total number of motors: the larger , the higher is required to generate a tumble. This is consistent with experimental data of [13], Fig. 12 therein, where switching of 1 motor is sufficient for a tumble at , but for at least 2 motors are necessary for a tumble. However, we keep the same for for simplicity, to avoid additional arbitrarily chosen thresholds. The simulated run lengths in a ligand-free medium have distribution close to exponential.
The cellular swimming speed depends on the distortion in a piece-wise linear form(9)In our model, we considered only ‘complete’ tumbles, which occur when reaches and the swimming speed falls to zero: at this time point the cell instantly changes its orientation by the tumbling angle , which is determined by two alternative models, isotropic and anisotropic. For simplicity, we assumed that if the distortion does not reach , it causes only a drop of speed, without a change of the swimming direction.
During a run, the direction of cellular swimming is affected by the rotational diffusion [12],[17]. After each time step, the swimming direction is changed by adding a stochastic component with normal distribution , where the diffusion coefficient equals [17].
In order to measure the chemotactic efficiency accurately and to avoid the effects of receptors saturation, we simulated the cells in an artificial constant-activity gradient, which ensures a constant chemotactic response CheYp and a constant cell drift velocity over a wide range of ligand concentrations, in contrast to commonly used Gaussian and linear gradients [14]. Drift velocity in constant-activity gradient was measured by a linear fit of in the time interval from 200 to 500 s. The constant-activity gradient has the following form:(11)where is the ligand concentration in position , and is the geometric mean of Tar methyl-aspartate dissociation constants. Here is a free parameter which determines the steepness of the gradient, and thereby the drift velocity of cells up the gradient. We compare the drift velocities in three constant-activity gradients, with relative steepness changing two-fold from one to another, and designate them as N0, N1 and N2. The corresponding gradient functions are(12)with mm for N0, N1 and N2, respectively. Here is the size of square 2D domain, where cells were simulated starting from the center of left wall .
|
10.1371/journal.pgen.1004866 | The nphp-2 and arl-13 Genetic Modules Interact to Regulate Ciliogenesis and Ciliary Microtubule Patterning in C. elegans | Cilia are microtubule-based cellular organelles that mediate signal transduction. Cilia are organized into several structurally and functionally distinct compartments: the basal body, the transition zone (TZ), and the cilia shaft. In vertebrates, the cystoprotein Inversin localizes to a portion of the cilia shaft adjacent to the TZ, a region termed the “Inversin compartment” (InvC). The mechanisms that establish and maintain the InvC are unknown. In the roundworm C. elegans, the cilia shafts of amphid channel and phasmid sensory cilia are subdivided into two regions defined by different microtubule ultrastructure: a proximal doublet-based region adjacent to the TZ, and a distal singlet-based region. It has been suggested that C. elegans cilia also possess an InvC, similarly to mammalian primary cilia. Here we explored the biogenesis, structure, and composition of the C. elegans ciliary doublet region and InvC. We show that the InvC is conserved and distinct from the doublet region. nphp-2 (the C. elegans Inversin homolog) and the doublet region genes arl-13, klp-11, and unc-119 are redundantly required for ciliogenesis. InvC and doublet region genes can be sorted into two modules—nphp-2+klp-11 and arl-13+unc-119—which are both antagonized by the hdac-6 deacetylase. The genes of this network modulate the sizes of the NPHP-2 InvC and ARL-13 doublet region. Glutamylation, a tubulin post-translational modification, is not required for ciliary targeting of InvC and doublet region components; rather, glutamylation is modulated by nphp-2, arl-13, and unc-119. The ciliary targeting and restricted localization of NPHP-2, ARL-13, and UNC-119 does not require TZ-, doublet region, and InvC-associated genes. NPHP-2 does require its calcium binding EF hand domain for targeting to the InvC. We conclude that the C. elegans InvC is distinct from the doublet region, and that components in these two regions interact to regulate ciliogenesis via cilia placement, ciliary microtubule ultrastructure, and protein localization.
| Cilia are sensory organelles that are found on most types of human cells and play essential roles in diverse processes ranging from vision and olfaction to embryonic symmetry breaking and kidney development. Individual cilia are divided into multiple functionally and compositionally distinct compartments, including a proximal “Inversin” compartment, which is located near the base of cilia. We used the nematode C. elegans, a well-defined animal model of cilia biology, to characterize the genetics, components, and defining properties of the proximal cilium. The Inversin compartment is conserved in C. elegans, and is established independent of another proximal ciliary region, the microtubule doublet-based region. We showed how components of both the doublet region and the Inversin compartment genetically interact to regulate many pathways linked to core aspects of cilia biology, including ciliogenesis, cilia placement, cilia ultrastructure, microtubule stability, and the protein composition of ciliary compartments. In addition to expanding and clarifying our knowledge of basic cilia biology, these results also have direct implications for human health research because several of the genes and pathways explored in our work are linked to ciliopathies, a group of diseases caused by dysfunctional cilia.
| Cilia are cellular “antennae” that mediate the transduction of environmental signals into intracellular pathways. Cilia play an integral role in many cellular functions, including developmental signaling, symmetry breaking, cell-cell adhesion, cell-cycle control, stress response, and DNA damage response (e.g., [1]–[6]). The vast majority of cilia share a set of evolutionarily conserved features: cilia are supported by a microtubule-based backbone, the axoneme; are built by intraflagellar transport (IFT), a microtubule motor driven cargo transport system [7]; and can be divided into structurally and functionally distinct compartments. These compartments include the microtubule triplet basal body which roots the cilium to the cell, the microtubule doublet transition zone (TZ) which anchors the cilium to the membrane, and the microtubule doublet cilia shaft where IFT occurs. The basal body and TZ also act as selective filters for inbound and outbound ciliary cargo, functioning through physical occlusion and cargo-specific recognition mechanisms [7]–[9]. The cilia shaft has traditionally been treated as an undifferentiated whole [10], though recent evidence has shed light on subdivisions of the cilia shaft [11].
Inversin/Nephrocystin-2 specifically localizes to the Inversin compartment (InvC), a proximal portion of the cilia shaft adjacent to the TZ [12]. This region has been suggested to play a role in signal transduction and amplification [13]–[15], cilia placement, and ciliogenesis [16], [17]. Products of several genes—including INVS/NPHP2, NPHP3, NEK8/NPHP9, and ANKS6/NPHP16—localize to the InvC. Interactions between InvC genes and other cilia genes have only recently begun to be explored and have not been well generalized across animal and cell culture models [18]–[22].
The mechanisms that initially establish the InvC are currently unknown, though recruitment pathways for several InvC components are known. Work in vertebrate models has shed light on an InvC-specific physical interaction complex composed of Inversin, Nek8, Nphp3, and Anks6 [23]–[25]. Nek8, Nphp3, and Anks6 localize to the InvC in an Inversin-dependent manner, but Inversin itself localizes independently of the other proteins [23], [24]. Unc119b—a possible, though not proven, InvC component—may mediate an InvC-targeting pathway. In mammalian cells, Unc119b binds myristoylated cargo, including Nphp3, and shuttles it into the cilium. Once the Unc119b-Nphp3 complex has translocated into the cilium, the small GTPase Arl3 triggers Unc119b to release bound cargo [26]. InvC components other than Nphp3 are not known to be myristoylated and shuttled via Unc119b, suggesting additional InvC targeting pathways must exist. Whether Unc119b is required for the localization of Inversin and how Unc119b itself is targeted to the proximal cilium is unknown.
The nematode Caenorhabditis elegans is a well-studied model of cilia biology [27]. C. elegans possesses a ciliated nervous system [28]–[31] which is primarily used by the roundworm to detect internal and external cues and signals. Amphid channel cilia in the head and phasmid cilia in the tail are exposed to and sense the external environment through cuticular pores [32], [33]. Unlike most mammalian primary cilia, the cilia shafts of C. elegans amphid channel and phasmid cilia are divided into two regions: a proximal microtubule doublet-based region attached to the TZ, and a distal microtubule singlet-based region that extends from the doublet region [32], [33]. As both the doublet region of C. elegans cilia and the InvC of mammalian primary cilia lie at the proximal end of the cilium, directly adjacent to the TZ at the cilia base, and constitute only a portion of the length of the cilia shaft, previous work has viewed them as compositionally and functionally similar [13], [17], [34]. The relationship between the mammalian InvC and the C. elegans doublet region of cilia has not been well characterized; Here we present evidence that the InvC and the doublet region are distinct, but overlapping, ciliary regions.
The C. elegans genome encodes orthologs for several of the mammalian InvC-associated proteins, including Inversin itself (NPHP-2), Unc119b (UNC-119), Arl3 (ARL-3), and possibly Nek8 (the uncharacterized paralogous pair NEKL-1 and NEKL-2), but likely not Nphp3 or Anks6. Of these, NPHP-2 and ARL-3 have previously been shown to be doublet region-localizing in C. elegans [17], [35]. C. elegans also possess several doublet region-enriched proteins, which are not InvC restricted in mammalian primary cilia; these include the kinesin-II IFT motor KLP-11 and the membrane-associated small GTPase ARL-13 [35]. The IFT motors Kinesin-II and OSM-3 work cooperatively to carry the IFT assemblies IFT-A and IFT-B and to build the doublet region—OSM-3 alone is sufficient to build the singlet region [36]. ARL-13 likely stabilizes the interaction between IFT-A and IFT-B particles and is required for ultrastructural integrity of the doublet region [35], [37]. arl-13 mutants exhibit multiple ciliary defects, some of which can be suppressed by deletion of histone deacetylase hdac-6, through an unknown mechanism [35]. Mammalian HDAC6 also antagonizes ciliogenesis in mammalian primary cilia: HDAC6 knockouts can suppress ciliogenesis defects arising from INVS/NPHP2 RNAi in MDCK cells [16].
In this work, we aimed to molecularly dissect the proximal cilium of C. elegans and to gain insight into the nature of the InvC and doublet region. We examined interactions between genes associated with the doublet region, determined the territories and localization dependencies of the protein products of these genes, and performed ultrastructural analysis of deletion mutants of these genes. We find that the InvC is conserved in C. elegans, is established early in development, and is distinct from the doublet region. nphp-2 interacts with doublet region genes to regulate cilia placement, microtubule ultrastructural patterning, tubulin glutamylation, and territory sizes of NPHP-2 and ARL-13. Finally, we show that nphp-2, arl-13, klp-11, and unc-119 fall into two parallel redundant genetic modules, and that interactions between the two modules are modulated by hdac-6 and arl-3. Together, the InvC and the doublet region function in concert to regulate many critical aspects of ciliogenesis and cilia biology.
nphp-2 and arl-13 single mutants have statistically similar, moderate ciliogenic defects (Fig. 1). As hdac-6 and arl-3 modulate several arl-13 phenotypes [35], we sought to determine if nphp-2 and arl-13 genetically interact and if hdac-6 and arl-3 modulate nphp-2 phenotypes. We examined double, triple, and quadruple mutant combinations using “dye filling” of ciliated neurons as a gross indicator of ciliogenesis and cilia integrity [32]. Properly formed and placed cilia are environmentally exposed and take up fluorescent DiI dye, whereas stunted and misplaced cilia are not exposed and cannot take up DiI. Unlike the mild dye-filling defects (Dyf) of nphp-2 and arl-13 single mutants, arl-13; nphp-2 double mutants were severely synthetic dye-filling defective (SynDyf) in both the amphids and phasmids (Fig. 1). hdac-6 deletion did not suppress nphp-2 or arl-13 Dyf; this is contrary to previously published data indicating that hdac-6 can partially suppress the weak arl-13 single mutant Dyf, and may be due to a difference in dye filling or scoring method [35] (Materials and Methods). However, hdac-6 suppressed the arl-13; nphp-2 severe SynDyf phenotype to the mild Dyf severity of the single mutants in both amphids and phasmids (Fig. 1). hdac-6 may function by suppressing defects arising from one of the pathways or at a point where the two pathways converge. Combined, this data indicates that arl-13 and nphp-2 act in partially redundant parallel pathways antagonized by hdac-6.
arl-3 has also been implicated as a modulator of the arl-13 pathway [35]. We found that, unlike the interactions with hdac-6, interactions with arl-3 are cell-type specific (S1A Figure). Both arl-3 single mutants and arl-3; hdac-6 double mutants were nonDyf in both amphids and phasmids. In amphids, arl-13; arl-3 was mildly SynDyf, whereas arl-3; nphp-2 is nonDyf (cf. Fig. 1 and S1A Figure). hdac-6 deletion suppressed the arl-13; arl-3 phenotype to a severity similar to that of the arl-13 single mutants. In phasmids, both arl-13; arl-3 and arl-3; nphp-2 double mutants exhibited a moderate SynDyf phenotype. hdac-6 suppressed arl-13; arl-3 defects, but not arl-3; nphp-2 defects. Strikingly, in both amphids and phasmids, arl-3 deletion prevented the hdac-6 Dyf suppression in the arl-13; hdac-6; nphp-2 triple mutant described above. The SynDyf phenotype of arl-13; arl-3 is qualitatively different from the suppression of arl-13 Dyf defects by arl-3 found previously [35]. Similarly to the difference in hdac-6 suppression of arl-13 defects discussed above, this may be due to different scoring methods or assay conditions. We conclude that arl-3 functions parallel to both nphp-2 and arl-13 pathways, and likely lies in the same regulatory pathway as, but acts antagonistically to, hdac-6.
To gain a better understanding of the defects present in nphp-2 single and arl-13; nphp-2 double mutants, and to examine the effects of hdac-6 mediated suppression of the double mutant defects, we used serial-section transmission electron microscopy (TEM) to examine the ultrastructure of amphid channel cilia. Wild-type amphid cilia are divided into three segments based on microtubule ultrastructure: the TZ, the doublet region, and the singlet region (Fig. 2A). One of the two microtubules within a doublet, the A-tubule, extends to form the microtubule singlet seen in the distal cilium (Fig. 2A1); the second tubule of the doublet, the B-tubule, terminates at the distal end of the doublet region. The doublet microtubules of the TZ and doublet region are arranged in a circular pattern in close proximity to the ciliary membrane (Fig. 2A2, 2A3). Within a particular cilium, microtubule ultrastructural characteristics—the location in the lumen, membrane association, and singlet/doublet architecture—are similar across all nine outer microtubule doublets.
In nphp-2 single mutants, in a given amphid cross-section at a single level, we observe singlet regions of some cilia, doublet regions of other cilia, and TZs of the remaining cilia. This is consistent with amphid cilia that were shifted lengthwise with respect to each other, indicating a potential anchoring defect (Fig. 2B3). Additionally, within a given cilium, doublet and singlet microtubule spans were no longer aligned. In sections across the proximal axoneme, this appeared as singlets amongst the expected doublets, and in sections across the distal axoneme, this appeared as doublets interspersed between the expected singlets (Fig. 2B1–3). These defects indicate that nphp-2 is required both for microtubule patterning and for cilia anchoring. While the chemical fixation method utilized here does not preserve Y-link ultrastructure well, nphp-2 mutants exhibited significantly greater Y-link disorder than observed in wild-type animals. The TZ defects seen in nphp-2 animals may be related to both the lengthwise cilia shift and the TZ-placement defect previously reported in nphp-2 mutants [17]. This set of defects has not been reported in any other C. elegans cilia mutant, suggesting that nphp-2 functions in a novel capacity.
arl-13 single mutants also exhibit a range of ultrastructural defects [35], [37]. Doublets are observed in the central lumen of the cilium, which may originate from either displaced outer doublets or mispatterned inner singlets. In both C. elegans arl-13 and hennin/ARL13B mouse mutants, there is also an increased frequency of early B-tubule detachment from the A-tubule, indicative of microtubule stability or patterning defects [38]. Similar to nphp-2 mutants, ectopic microtubule singlets were sometimes visible in the doublet region of arl-13 worms [35], [37]. arl-13; nphp-2 double mutants exhibited extreme ultrastructural defects, likely causative of the severe Dyf phenotype (Fig. 2C1–3). Cilia were almost completely absent from the amphid channel pore. No other cilia were visible even at the distal dendritic level. In one instance, a single cilium was visible in TEM sections. At the TZ level, an incomplete set of doublets was visible within the single cilium (Fig. 2C3). Closer to the socket cell/sheath cell transition, a set of singlets was visible (Fig. 2C2). At the distal pore, we were unable to resolve any internal structure due to electron dense material filling the cilium (Fig. 2C1).
Remarkably, in arl-13; hdac-6; nphp-2 triple mutants almost all defects observed in the double mutant were suppressed (Fig. 2D1–3). The only defects remaining were misplaced TZs (Fig. 2D3), ectopic singlets in the doublet region (Fig. 2D2), and ectopic doublets in the singlet region (Fig. 2D1), similar to those present in the nphp-2 single mutant. We did not observe inner doublets as reported in arl-13 single mutants [35], [37]. These results are consistent with the observed hdac-6 suppression of arl-13; nphp-2 Dyf defects.
Because nphp-2 and arl-13 both exhibit ectopic microtubule singlets, nphp-2 mutants exhibit ectopic microtubule doublets, and nphp-2; arl-13 double mutants exhibit severe defects in ciliogenesis, we conclude that nphp-2 and arl-13 function together redundantly in regulation of microtubule patterning and ciliogenesis. Because ciliogenic defects are suppressed by hdac-6, but ectopic doublets and singlets are not, ciliogenesis and microtubule patterning may be independently regulated.
Localization of InvC and doublet region components can be broken down into three steps: first the protein is targeted to the cilia base, second, the protein is imported into the cilium, and third, the protein is restricted to a subdomain of the cilium [23]. The factors required for the initial establishment of the InvC and doublet region cilia targeting and localization restriction are unknown. The TZ functions as a regulator of ciliary protein import (Reviewed in [39]), and has been implicated in the import of InvC and doublet region components in mammalian cilia [40], [41]. As both nphp-2 and arl-13 genetically interact with TZ-associated genes [17], [37], we wanted to determine if NPHP-2 and ARL-13 ciliary targeting and import requires TZ components. In C. elegans, TZ genes are organized into two genetic and physical modules—the mks module and the nphp-1+nphp-4 module [8], [17], [25], [42]. We examined the localization of NPHP-2 and ARL-13 in mutants missing a component of each module (Fig. 3A,C).
NPHP-2::GFP signal length is similar in wild-type and nphp-2 backgrounds (1.77±0.23 um vs 1.74±0.22 um, st. dev.) and NPHP-2::GFP rescues the SynDyf phenotype of the nphp-2 nphp-4 double mutant (Fig. 4B) [17]. These results indicate that this reporter reflects NPHP-2 functional and endogenous localization. In both nphp-4 and mks-3 single mutants, NPHP-2::GFP was properly targeted to and imported into the cilium. Mislocalized NPHP-2::GFP puncta in the periciliary region were sometimes visible (Fig. 3A). In mks-3; nphp-4 double mutants, there were severe ciliogenic and dendritic extension errors, as previously reported [17], [42]; in phasmid cilia that were visible and placed properly, NPHP-2::GFP localization appeared as in mks-3 and nphp-4 single mutants (S2A Figure). In a wild-type background, ARL-13::GFP localized exclusively to the doublet region in amphid channel and phasmid neurons (Fig. 3C). Like NPHP-2::GFP, ARL-13::GFP was targeted to and imported into the cilium properly in mks-3 and nphp-4 mutants. Similar to published reports, we also observed ARL-13::GFP mislocalization to the periciliary membrane, as judged by a fluorescent “fringe” surrounding the periciliary region where the membrane lies (Fig. 3C, enhanced contrast in S2B Figure), which has been suggested to be due to a failure of the TZ diffusion barrier [34]. In amphid cilia, like in phasmid cilia, both NPHP-2 and ARL-13 were targeted to the cilium and imported properly, and both infrequently exhibited mild mislocalization (S3 Figure).
We investigated whether doublet region-associated genes were required for doublet region restriction of NPHP-2::GFP and ARL-13::GFP. hdac-6 had no obvious effect on the localization of NPHP-2::GFP (S2A Figure). NPHP-2::GFP was targeted and restricted to the proximal cilium in both klp-11 and arl-13 mutants. In both mutants, periciliary puncta similar to those seen in TZ-associated mutants were visible (Fig. 3B). unc-119 mutants also exhibited proper NPHP-2::GFP ciliary targeting. In unc-119 mutants, NPHP-2::GFP exhibited a unique distal dendrite mislocalization pattern, distinct from the periciliary puncta seen in other TZ and doublet region mutants (Fig. 3B).
In klp-11 and nphp-2 mutants, ARL-13::GFP was restricted to a proximal portion of the cilium as in wild type, but with a periciliary membrane mislocalization pattern similar to TZ mutants. ARL-13::GFP also did not require unc-119 for ciliary localization. In unc-119 mutants, ARL-13::GFP and NPHP-2::GFP localized along the distal dendrite in a similar manner (Fig. 3D).
In all TZ, doublet region, and InvC mutants examined, NPHP-2::GFP and ARL-13::GFP still localized to the cilium, suggesting that either unknown factors or redundant pathways are required for establishing ciliary territories. Periciliary mislocalization was observed for both reporters across all mutant backgrounds. This suggests that either a delicate, easily perturbed interaction network is required for NPHP-2 and ARL-13 ciliary import/export, or that the overexpressed reporter constructs are “leaking” out of the cilium in sensitized mutant backgrounds, or a combination of both possibilities.
We next looked to determine which domains of NPHP-2 were required for InvC localization. In mammalian models, several domains in Inversin are required for ciliary targeting and InvC restriction: the ankyrin repeat region, IQ2 domain, and ninein-homologous region (Fig. 4A) [12], [43]. The IQ and ninein homology domains are not conserved in C. elegans NPHP-2; only the ankyrin repeat region, hydroxylation motif (S8 Figure), and two nuclear localization signals (NLSs) are conserved (Fig. 4A) [17]. NLS motifs are hypothesized to play a role in ciliary protein import, and are required for ciliary import of the IFT motor KIF17 [9]. NPHP-2 also contains a predicted calcium binding EF-hand which may function in the same calcium detection capacity as the IQ domain of Inversin [17].
We built NPHP-2::GFP constructs missing the EF-hand (NPHP-2-EFΔ::GFP, residues 520–533), NLS1 (NPHP-2-NLS1Δ::GFP, residues 441–446), or NLS2 (NPHP-2-NLS2Δ::GFP, residues 598–603) (Fig. 4A). Full-length NPHP-2::GFP localized to a short proximal region of amphid channel and phasmid cilia as well as IL, CEP, OLQ, amphid channel and phasmid cilia, and the AWC wing cilia (Fig. 4B). Native promoter driven NPHP-2::GFP was not visible in AWB cilia, as reported previously for AWB-specific promoter driven NPHP-2 [13]. NPHP-2-EFΔ::GFP was generally faint or absent in amphid channel, AWC, and phasmid cilia, indicating that the EF-hand is strictly required for normal localization of NPHP-2 in these cell types. NPHP-2-EFΔ::GFP properly localized in IL cilia, indicating that the fluorescent reporter was being synthesized and folded correctly. NPHP-2-EFΔ::GFP was additionally present in either CEP or OLQ cilia, though specific identification was difficult due to their close proximity. Deletion of either NLS1 or NLS2 did not perturb NPHP-2 localization (Fig. 4B).
We next tested whether these domain deletion constructs were functional by attempting to rescue the severe Dyf phenotype of nphp-2 nphp-4 mutants; nphp-4 single mutants are nonDyf, allowing for easy determination of rescue [17]. Rescue of the nphp-2 nphp-4 SynDyf phenotype by NPHP-2-NLS1Δ::GFP and NPHP-2-NLS2Δ::GFP was comparable to rescue by full length NPHP-2::GFP. NPHP-2-EFΔ::GFP only weakly rescued the SynDyf phenotype, indicating that the EF-hand is critical for both normal function and localization of NPHP-2 (Fig. 4C). These results indicate that the calcium-binding EF-hand plays a significant role in both localization and function of NPHP-2 in amphid channel and phasmid cilia, but is dispensable for localization in cilia of CEP, OLQ, and inner labial neurons.
Mammalian Unc119b localizes to the proximal cilium and is required for the targeting of several proteins to the InvC [26], and C. elegans unc-119 is required for singlet region biogenesis in amphid cilia [44], suggesting a compartment-specific role. We therefore examined the localization of posm-6::GFP::UNC-119. In phasmid cilia, GFP::UNC-119 localization was similar to NPHP-2::GFP and ARL-13::GFP: its localization was restricted only to a small proximal portion of the cilium and was excluded from the TZ. Similar to localization in phasmid cilia, GFP::UNC-119 likely localized to the doublet region and not the entire cilium of amphid cilia (Fig. 5A). GFP::UNC-119 was not motile in cilia.
We also examined the dependence of UNC-119 localization on doublet region-associated genes. Specific amphid mislocalization was difficult to determine, and we therefore focused on phasmid cilia (S4 Figure). In phasmid cilia, GFP::UNC-119 did not require nphp-2, arl-13, hdac-6, or klp-11 for ciliary targeting or restriction to the proximal cilium (Fig. 5B).
We examined the genetic interactions between the doublet region-associated genes nphp-2, arl-13, and their modulators hdac-6 and arl-3 with the two IFT motors, osm-3 and klp-11. In both amphids and phasmids, arl-13; klp-11 was SynDyf [37]. arl-13; klp-11 dye-filling defects were slightly suppressed by deletion of hdac-6 (Fig. 6A, S5A Figure). klp-11; arl-3 double mutants yielded a very mild SynDyf phenotype (S1B Figure). osm-3 single mutants are missing a singlet region and are completely Dyf, which precludes searching for synthetic interactors of osm-3. Instead, we assayed osm-3 double mutants for suppression of the severe dye-filling defect; in no mutant background examined was the defect suppressed (Fig. 6B, S5B Figure).
We next examined genetic interactions between unc-119 and other doublet region-associated genes. unc-119 single mutants were severely Dyf, with the phasmid phenotype being less severe than the amphids (Fig. 6B, S5A Figure) [44]. In phasmids, both nphp-2 and klp-11 were SynDyf with unc-119. Surprisingly, we found that hdac-6 suppressed Dyf defects in the unc-119; nphp-2 double mutant to the level of the nphp-2 single mutant (Fig. 6B). arl-3 deletion was also able to significantly suppress the unc-119 Dyf phenotype in both amphids and phasmids (S1B Figure).
Genetic interactions between nphp-2, arl-13, klp-11, unc-119, hdac-6, and arl-3 are summarized in Fig. 6C. Doublet region-associated genes fall into two redundant pathways or modules: nphp-2+klp-11 and arl-13+unc-119. Deletion of two genes within a module does not result in an increase in Dyf severity over that present in single mutants, but deletion of any two genes from different modules yields a SynDyf phenotype. Interactions between these modules are regulated by hdac-6 and arl-3.
Double mutants with a deletion in a single TZ gene and a single doublet region gene are SynDyf. To determine whether hdac-6 mediated suppression of SynDyf defects extended to these cross-compartmental genetic interactions, we assayed for suppression of SynDyf defects in nphp-2 nphp-4 mutants by hdac-6 and arl-3. We found that neither hdac-6 nor arl-3 suppressed the severe nphp-2 nphp-4 SynDyf defect (S5D Figure). This suggests that hdac-6 functions specifically in doublet region pathways.
Post-translational glutamylation predominantly occurs on the C-terminal tails of α- and β-tubulin of axonemal B-tubules [45]–[49], and regulates microtubule stability and IFT motor function [50]. Glutamylation is specifically associated with the doublet region, as B-tubules define and are only present in the doublet region. Additionally, in Chlamydomonas and Paramecium, TZ microtubules are not glutamylated [45], [51]. In C. elegans and vertebrates, mutants with defects in tubulin glutamylation or arl-13 exhibit B-tubule degeneration [38], [50], [52]. We therefore determined whether doublet region associated genes regulated tubulin glutamylation, or whether tubulin glutamylation specified the localization of doublet region proteins.
In wild-type animals, the anti-glutamylated tubulin antibody GT335 labeled the doublet region of amphid channel and phasmid cilia (Fig. 7A) [52]. nphp-2 mutants exhibited characteristic cilia displacement in the amphids, but no qualitative changes in head cilia glutamylation. The glutamylation signal in nphp-2 phasmid cilia ranged from wild-type-like to extremely elongated, which is consistent with the TEM observation of B-tubules extending into the distal axoneme (Fig. 2B). arl-13 mutants exhibited elongated staining in amphid channel cilia. Amphid staining in unc-119 mutants was extremely shortened and cilia were angled inwards. hdac-6 mutants appeared to have shortened GT335 staining of amphid channel cilia. We also observed significant differences in the length of the phasmid GT335 ciliary signal. Both arl-13 (3.95±0.25 µm) and nphp-2 (4.30±0.32 µm) mutants had phasmid staining significantly longer than in wild type (2.79±0.07 µm), while unc-119 (2.40±0.05 µm) mutants had staining significantly shorter (S7E Figure). The length of GT335 staining in amphid cilia was not quantified due to the difficulty of unbiased measurement of a single cilium within the amphid bundle.
To determine if doublet region-associated protein localization was dependent on tubulin glutamylation status, we examined the localization of NPHP-2::GFP and GFP::UNC-119 in ccpp-1 and ttll-4 mutants. Mutants of ccpp-1, which encodes a tubulin deglutamylase, display degenerating amphid channel and phasmid cilia with a concomitant dye-filling defect; this is suppressed by deletion of the opposing glutamylase, encoded by the tubulin tyrosine ligase-like gene ttll-4 [50]. In each mutant, NPHP-2::GFP and GFP::UNC-119 reporters were targeted to cilia and were restricted to the proximal cilium similarly to wild-type (Fig. 7B,C). This was surprising in the case of ccpp-1 mutants, as cilia degenerate as the worm ages (Fig. 7B,C). These reporters were still doublet region-associated in earlier larval stages of ccpp-1 mutants when ciliary degeneration was not as severe (S9C-D Figure). Combined, these results indicate that nphp-2, arl-13, unc-119, and hdac-6 lie upstream in regulation of tubulin glutamylation pathways, and the localization patterns of their protein products are not defined by tubulin glutamylation.
To allow for a more direct comparison of subciliary localization, we stained transgenic NPHP-2::GFP, ARL-13::GFP, and GFP::UNC-119 strains with GT335 (Fig. 8). NPHP-2::GFP did not fully overlap with GT335, only colabelling the proximal portion of the doublet region of amphid and phasmid cilia. However, ARL-13::GFP colabelled with a greater portion of the GT335 doublet region signal in amphid cilia than did NPHP-2::GFP, and completely colabelled with GT335 in phasmid cilia (Fig. 8, S6A Figure); this suggests that ARL-13 is associated with the microtubule doublets that define the doublet region. GFP::UNC-119 colabelled with either a significant fraction of the length of the GT335 signal. Additionally, GFP::UNC-119 did not extend beyond the GT335 labelled doublet region, indicating that GFP::UNC-119 is excluded from the TZ, which is not labelled by GT335.
We also examined the territory length of each of the fluorescent reporters as a fraction of the total cilia length. Transgenic animals were incubated with DiI to label the length of the cilium. NPHP-2::GFP marked a significantly shorter fraction of the length of the cilium than did ARL-13::GFP or GFP::UNC-119 (S6C Figure).
To understand how the localization of doublet region-associated proteins is genetically regulated, we measured the length of the cilium marked by NPHP-2::GFP, ARL-13::GFP, GFP::UNC-119, and KAP-1::GFP—a component of Kinesin-II—in different mutant backgrounds (S7 Figure). ARL-13::GFP and NPHP-2::GFP have interdependent localizations: in arl-13 mutants, the NPHP-2::GFP territory was extended along the cilium, and in nphp-2 mutants, the ARL-13::GFP territory was extended. unc-119 mutants exhibited shortened ARL-13::GFP and NPHP-2::GFP territories (S7A-B Figure). Additionally, klp-11 mutants had a shorter NPHP-2::GFP, but not ARL-13::GFP, localization signal (S7A Figure). No significant differences were found in the territory lengths of GFP::UNC-119 and KAP-1::GFP in any of the mutant backgrounds (S7C-D Figure). The length of the tubulin glutamylation signal is also genetically controlled; the GT335 signal is shortened in unc-119 mutants, and elongated in nphp-2 and arl-13 mutants (S7E Figure).
In sum, the NPHP-2::GFP territory size is shorter than the territories of doublet region components ARL-13 and Kinesin-II, marks a shorter length of the cilium than ARL-13 and UNC-119, is shorter than the doublet region-linked glutamylated tubulin signal, and colabels only a portion of both amphid channel and phasmid GT335 staining. We conclude that NPHP-2 marks a region of the cilium distinct from the doublet region, and propose that this region is analogous to the InvC of mammalian cilia.
In this study, we present evidence that (1) C. elegans possesses a conserved InvC compartment, (2) interactions between nphp-2 and arl-13 regulate microtubule ultrastructural patterning, (3) InvC and doublet region sizes are distinct and genetically regulated, (4) hdac-6 and arl-3 modulate interactions between nphp-2, arl-13, klp-11, and unc-119, (5) and that microtubule glutamylation is downstream of the action of InvC and doublet region genes. Additionally, we found that genes associated with the proximal cilium (TZ, InvC, and doublet region) can be grouped into parallel genetic modules, which interact to drive ciliary anchoring and proper ciliogenesis. Finally, we addressed several possible mechanisms for the ciliary targeting and InvC restriction of NPHP-2.
In the last twenty years, fluorescent tagging of proteins has allowed for unparalleled insight into in vivo localization and transport mechanisms. However, the primary method of introducing transgenes into C. elegans, microinjection, yields extrachromosomal arrays containing many copies of the reporter construct [53]. Many aspects of ciliogenesis are tightly regulated, and may require stoichiometric quantities of components for proper function. We previously showed that overexpression of proteins may lead to defects in ciliogenesis and IFT, resulting in the SynDyf phenotype [17], [54].
We sought to minimize these effects on our conclusions by avoiding direct comparisons between multiple reporters, and using antibody colabeling when comparisons were required. In wild-type and mutant backgrounds, we tested all constructs for dominant defects in ciliogenesis (S1 Table). In the respective mutant background, we routinely test for rescue of mutant phenotype, which indicates that the reporter is functional (Fig. 4A) [17], [29], [50], [54]–[57]. For example, NPHP-2::GFP localization is similar in wild-type and nphp-2 mutant backgrounds, and rescues the nphp-2 nphp-4 SynDyf phenotype (Fig. 4B) [17]. Therefore we can conclude that NPHP-2::GFP provides and accurate and functional reflection of the endogenous localization pattern.
The use of ARL-13::GFP is well established in the literature, with five papers examining ARL-13::GFP localization in amphid and phasmid cilia [34], [35], [37], [58], [59], and a sixth paper examining ARL-13::GFP localization in AFD cilia [60]. In two papers, the ARL-13::GFP construct is the same as was used in our work [35], [59]. Moreover, this ARL-13::GFP reporter is functional [35] and displays a similar localization in wild-type and arl-13 backgrounds (Jinghua Hu, personal communication). In three papers, ARL-13::GFP reporters are used to determine subciliary localization of ARL-13::GFP [34], [59], [60]. Additionally, the ARL-13::GFP transgenic lines used in the papers by the Hu and Blacque labs were built independently and exhibit similar localizations.
In the future, genomic engineering (e.g., CRISPR) will allow for easy single-copy fluorescent tagging of proteins and in vivo analysis, addressing many of these concerns [61]. These techniques are not a panacea though, as several genes, including nphp-2, may be expressed at too low a level for single-copy tagged constructs to be visible without advanced microscopic techniques (personal communication, Knudra Transgenics).
The C. elegans doublet region and the mammalian proximal InvC have been considered analogous [11], [13], [17], [34], [37]. In C. elegans, a number of factors have been associated with the doublet region, including NPHP-2, ARL-13, UNC-119, ARL-3, HDAC-6, the Kinesin-II components KAP-1/KLP-11/KLP-20 [35]–[37], and glutamylated tubulin [50]. However, mammalian orthologs of many C. elegans doublet region proteins localize along the entire cilium [26], [62]–[71]; this casts doubt on the equivalence between the C. elegans doublet region and the mammalian InvC. It is likely that the C. elegans ciliary doublet region is analogous to the entire mammalian ciliary doublet-based cilia shaft, and that the mammalian InvC is analogous to a C. elegans InvC (modelled in Fig. 9).
Calcium signaling plays a crucial role in signal transduction and ciliary function [72], [73]; cilia have high intraciliary calcium concentrations, and many TZ proteins possess calcium binding domains [74]. The NPHP-2 vertebrate homolog, Inversin, has two identified calmodulin-binding IQ domains [75], [76], one of which is required for proper localization [12]. Calmodulin detects intracellular calcium concentrations through a calcium-binding EF-hand. Though C. elegans NPHP-2 does not encode a predicted IQ domain, it does possess an EF-hand. This EF-hand is required for the localization and function of NPHP-2 in amphid and phasmid cilia, similarly to the IQ2 domain of Inversin. A significant difference exists between the EF-hand of NPHP-2 and the IQ2 domain of Inversin: deletion of the EF-hand of NPHP-2 results in a complete lack of ciliary localization, whereas deletion of the IQ2 domain of Inversin results in a mislocalization of Inversin along the entire cilium. In both systems, Inversin/NPHP-2 no longer localizes to the InvC. This suggests that Ca2+ detection/binding by and the subsequent hypothetical modulation of the activity of Inversin/NPHP-2 is a critical, conserved feature of the protein. Two possibilities arise for the function of these domains: Ca2+ specifies the localization of NPHP-2, modulates the activity of the protein, or both.
We found that UNC-119 localizes to the proximal cilium and is excluded from the distal region. GFP::UNC-119 and GT335 colabel, indicating that GFP::UNC-119 is excluded from the TZ, as TZ microtubules are not glutamylated (See Fig. 8, in which GT335 colabels with TZ-excluded ARL-13::GFP and NPHP-2::GFP). In C. elegans, UNC-119 labels a shorter portion of the cilium than ARL-13 or GT335, markers associated with the doublet region. Additionally, mammalian Unc119b physically interacts with the InvC component Nphp3 and is proximally restricted in cilia of RPE cells, suggesting that Unc119b is associated with the InvC [26]. However, GFP::UNC-119 marked a larger portion of the cilium than did NPHP-2::GFP.
The C. elegans genome encodes homologs of many of many Unc119b shuttle proteins, including Unc119b, Arl3, and RP2, and two myristoylated ciliary proteins which require unc-119 for ciliary localization [77]. arl-3 genetically interacts with unc-119 and nphp-2, suggesting that in C. elegans the components of the shuttle are in place. These shuttle components do not appear to be required for the localization of NPHP-2, as NPHP-2 is imported into the cilium in unc-119 and arl-3 mutants. In unc-119 mutants, NPHP-2 exhibited a unique distal dendritic localization pattern that cannot be attributable to TZ leakage; it is unknown whether this population represents NPHP-2 that has not been properly imported into the cilium or is mistargeted NPHP-2.
In both mammalian systems and C. elegans, Hdac6/hdac-6 functions as an antagonist of ciliogenesis and cilia stability [16], [62]. In mammalian primary cilia, Hdac6 functions as an α-tubulin K40 deacetylase regulates microtubule function and stability [62], [78]–[81]. The C. elegans genome encodes a single α-tubulin with the acetylatable residue K40, MEC-12. However, there is no direct evidence for mec-12 expression in amphid and phasmid neurons, and the anti-α-tubulin-K40 antibody 6-11b-1 does not label amphid and phasmid cilia in wild-type animals or hdac-6 mutants (S10 Figure) [82]. Alternative tubulin acetylation sites may exist, including on β-tubulin [83]; hdac-6 could deacetylate these secondary sites. NPHP-2 contains a predicted N-terminal acetylation site (at 2S) which may be deacetylated by HDAC-6; this may modulate binding between NPHP-2 and its targets [84]. Additionally, HDAC-6 may have other unidentified ciliary targets [35]. Determining the mechanism by which HDAC-6 acts as a genetic modifier of InvC and doublet region gene defects is an important future direction.
Tubulin glutamylation is associated with the proximal portions of microtubule B-tubules in C. elegans cilia, mouse spermatozoa flagella, and Chlamydomonas flagella [47], [50], [85]. In C. elegans, microtubule glutamylation is linked to microtubule ultrastructure, stability, and maintenance [50], [86]. Multiple doublet region genes regulate microtubule glutamylation, and we observed a correlation between ectopic glutamylation and ectopic microtubule doublets. Additionally, ciliary targeting of doublet region proteins is not dependent on glutamylation status. Therefore, nphp-2, arl-13, hdac-6, and unc-119 function upstream of microtubule glutamylation, which may enable them to exert an influence on microtubule patterning, IFT, ciliogenesis, and, in the case of unc-119, singlet region biogenesis. These pathways may be conserved: in Arl13b/hennin mutant mice, ciliary microtubule glutamylation intensity is reduced, and microtubule B-tubules have ultrastructural defects [38], [64].
In C. elegans, TZ-associated genes can be grouped into two distinct, partially redundant genetic and physical modules [8], [17], [42]. Doublet region- and InvC-associated genes may be grouped in a similar manner into a nphp-2+klp-11 module and an arl-13+unc-119 module. hdac-6 appears to function outside the two modules, negatively regulating both: deletion of hdac-6 in SynDyf double mutants suppresses the SynDyf phenotype. In mammalian primary cilia, Hdac6 also plays an antagonistic role, destabilizing cilia through deacetylation of tubulin, a pathway suppressed by Inversin [16]. arl-3 may function outside of the two modules in a cell-type specific manner; in phasmids it genetically interacts with components from both modules, but in amphids arl-3 only genetically interacts with only the arl-13+unc-119 module and not the nphp-2+klp-11 module. In amphid cilia, nphp-2 also does not interact with the TZ SynDyf network.
Curiously, a further two genetic module organization exists between TZ genes and the InvC/doublet region genes. nphp-2 nphp-4, mks-3; nphp-2, and arl-13; nphp-4 double mutants exhibit severe ciliogenic defects [17]. Deletion of one TZ gene from either TZ module and one doublet region gene from either doublet region module yields a SynDyf phenotype, though not all combinations have been tested. The TZ SynDyf network and the doublet region SynDyf network can be thought of as two genetically interacting “super-modules”, each consisting of two to three sub-modules described in this and previous work [8], [17], [25], [42].
The localization requirements for multiple InvC localizing components have been previously determined, but how the InvC is initially established is not known.
In C. elegans, the InvC is likely established early in cilia development, as NPHP-2 is proximally restricted in phasmid cilia as early as the first larval L1 stage immediately following hatching (S9A Figure), and is not motile (S9B Figure), unlike the larval stage-dependent dynamic localization of ARL-13 [34]. We have eliminated several mechanisms for the establishment of the InvC. Ciliary ultrastructure does not seem to play a role, as in both mammals and C. elegans, the localization of Inversin/NPHP-2 is associated with only a sub-portion of the doublet region where there are no identifiable ultrastructural features [12]. Tubulin post-translational modifications also do not appear to specify the InvC, as nphp-2 (and genetically interacting doublet region components) lies upstream of glutamylation pathways. The TZ does not appear to play a major role in specifying the InvC. NPHP-2 still localizes and is restricted to the proximal cilium in TZ single and double mutants. IFT is another candidate mechanism, but we found that although IFT components genetically interact with InvC and doublet region associated genes, Kinesin-II is not required for NPHP-2 localization. In mammalian cilia, the Unc119b shuttle is required for the ciliary import of Nphp3; this activity is upstream of the action of Inversin in Nphp3 localization. In C. elegans phasmid cilia, NPHP-2 does not require either unc-119 or its effector arl-3 for ciliary import or InvC restriction (S2A Figure).
Several mechanisms for establishing the InvC remain. First, calcium may play a role. Both Inversin and NPHP-2 require a calcium binding domain for InvC localization [12]; the origin and nature of the calcium signal these domains detect is unknown. Second, the InvC may also be initially established by a diffusion of factors from the cilia base. A third possibility is that cilia membrane composition may help define the InvC. The cilium has a distinct membrane composition from the plasma membrane, and the different ciliary subregions may also have differential membrane composition.
We propose that the logic underlying the establishment of the NPHP-2/Inversin compartment may be similar in C. elegans and mammals, in a manner independent of microtubule ultrastructure. We have shown that doublet region- and InvC-associated genes interact to guide ciliogenesis, cilia placement, cilia ultrastructure, protein composition, and tubulin post-translational modification. The next challenge is to determine what initially patterns the doublet region and InvC, and to understand the function of these cilia regions.
Standard protocols were followed for all molecular biological procedures. PCR amplification using Taq polymerase (New England BioLabs, Ipswich, MA, USA) was used for genotyping deletion alleles, and was followed by restriction digest for SNP diagnosis. PCR amplification for construction of transgenic constructs was performed with Phusion High fidelity DNA polymerase (New England BioLabs), templated off C. elegans genomic DNA. Sequencing was performed offsite (GeneWiz, South Plainfield, NJ, USA). PCR primer and construct sequences are available upon request.
Protein BLAST was used to find sequence orthologs [87]. All protein sequence information other than that of C. elegans was provided by NCBI, and all C. elegans nucleotide and protein sequences were provided by WormBase (Releases WS229 and WS234). Structural motif and domain predictions were generated by MotifScan [88]. Acetylation motifs were identified using NetAcet 1.0 [84]. Coiled-coil regions were identified with COILS [89]. ApE 2.0.36 was used for sequence manipulation, annotation, and restriction site identification.
All strains were cultured at room temperature, unless otherwise noted, under standard conditions [90]. Transgenic strains using pha-1 selection were grown at 25°C, and pha-1(e2123) mutants were grown at 15°C. Deletion alleles were outcrossed to him-5 at least four times. Strains used in this study are listed in S2A Table. Alleles used were as follows: nphp-2(gk653), nphp-4(tm925), mks-3(tm2547), arl-13(gk513), unc-119(ed3), hdac-6(ok3203), arl-3(tm1703), him-5(e1490), osm-3(p802), klp-11(tm324), ccpp-1(ok1821), and ttll-4(tm3310). Primers used for diagnosis are listed in S2B Table. All transgenic strains used in Fig. 1–9 were tested using dye-filling for dominant negative defects in ciliogenesis. We observed no adverse effects in NPHP-2::GFP and GFP::UNC-119 transgenic strains but did find dominant negative defects in ARL-13::GFP transgenic strains (S1 Table).
The full length isoform of NPHP-2 was used in all NPHP-2 reporter constructs. The full length isoform, NPHP-2L, differs from the shorter isoform, NPHP-2S, in that the shorter isoform is missing 22 non-conserved amino acids of unknown function. Both isoforms have similar subciliary localization [17].
nphp-2 and arl-13; hdac-6; nphp-2 young adult animals were fixed using 3.5% glutaraldehyde +1% PFA in 0.1M HEPES and then in 1% OsO4+1.25% K4Fe(CN)4 in 0.1M HEPES. Samples were infiltrated and embedded in Embed-812 plastic resin. arl-13; nphp-2 young adult animals were fixed using high-pressure freeze fixation and freeze substitution in 2% OsO4+2% water in acetone as the primary fixative [91]. Samples were slowly freeze substituted in an RMC freeze substitution device, before infiltration with Embed-812 plastic resin. Images for wild-type animals fixed by a comparable immersion fixation method (cf. [32]) are now curated by the Hall lab at Einstein courtesy of E. Hedgecock. These wild type images are also available online at www.wormimage.org.
For TEM, serial sections (70 nm thickness) of fixed animals were collected on copper slot grids coated with formvar and evaporated carbon and stained with 4% uranyl acetate in 70% methanol, followed by washing and incubating with aqueous lead citrate. Images were captured on a Philips CM10 transmission electron microscope at 80 kV with a Morada 11 megapixel TEM CCD camera driven by iTEM software (Olympus Soft Imaging Solutions).
For each strain, we imaged three individuals that were fixed chemically. Additionally, we were concerned that the severe defects seen in the arl-13; nphp-2 double mutant were partially due to the harsh chemical fixation method. We fixed a fourth double mutant using high pressure freeze (HPF) fixation, which introduced fewer artifacts to confirm the validity of the chemical fixation data. We used the same strain in the construction of the double arl-13; nphp-2 and triple arl-13; hdac-6; nphp-2 mutants as was used in the previously published EM of the arl-13 single mutant.
Animals were mounted on 5% Noble agar pads on standard microscope slides, and immobilized with a five minute incubation in 10 mM sodium azide. Worms were imaged using a Zeiss Plan-AXIOCHROMA 100X 1.4NA oil objective on a Zeiss Axio Imager.D1M (Zeiss, Oberkochen, Germany) with a Retiga-SRV Fast 1394 digital camera (Q-Imaging, Surrey, BC, Canada). Exposure time for antibodies was 100 ms, and exposure time for GFP fluorophores was 250 ms. Images were captured and manipulated using Metamorph software (Version 7.6.1.0, MDS Analytical technologies, Sunnyvale, CA, USA). Image stacks were 3D deconvolved using Auto Deblur software (Version 1.4.1, Media Cybernetics, Bethesda, MD, USA). Figures and diagrams were created with Adobe Photoshop CS3 (Version 10.0, Adobe Systems, San Jose, CA, USA) and Adobe Illustrator (Version 13.0.0, Adobe Systems). Image brightness and contrast were modified uniformly across an image, but gamma was not adjusted from 1.00. Brightness manipulations are similar, but not identical, across panels and figures. Significant variations in absolute intensity are noted where appropriate. For all strains, unless noted, worms were picked at L4 stage 24 hours before imaging.
All statistical analysis was performed with a combination of GraphPad Prism (Version 5.01, GraphPad Software, La Jolla, CA, USA) and Microsoft Excel (Version 14.0.7106, 32-Bit, Microsoft Corporation, Seattle, WA, USA). Sample size (n) for all figures is listed in S3 Table. Minimum p value for significance was set at 0.01 for all analyses unless otherwise specified. All parametric and continuous data types were analyzed using unpaired t-tests with Welch's correction to avoid assumption of equal variance. When multiple t-tests were performed on related data sets presented together, the Holm-Bonferroni multiple comparison adjustment was used to ensure the total alpha for the analysis did not exceed 0.01. All nonparametric and discontinuous data types were analyzed using Mann-Whitney U-test. Similar to the analysis of continuous data types, the Holm-Bonferroni multiple comparison adjustment was employed to ensure total alpha for all related comparisons did not exceed 0.01. Specific pairwise comparisons made are described in figure legends. Letters on graphs indicate statistically distinct groups, e.g., all groups marked ‘A’ are significantly different from all groups marked ‘B’.
Staged young adult hermaphrodites were washed of plates with M9, and then rinsed three further times in M9, using gentle centrifugation to pellet the worms between rinses. Worms were then incubated in 40 µg/mL DiI (2.5 mg/mL dimethyl formamide stock, diluted 1∶1000 in M9) (Invitrogen) for 30 minutes in the dark. Worms were then rinsed three times in M9 as before, and were then placed on a seeded plate for a further 30-60 minutes to recover and flush dye from the digestive tract. Animals were anesthetized with 10 mM sodium azide and then immediately scored on a compound microscope (see Imaging section) using for dye-filling by manual counting of filled cell bodies. Cell body counts within the amphid or phasmid organs were averaged together to yield the average number of cells filling per organ per worm, and subjected to statistical analysis. Ectopically dye-filled neurons (e.g., IL2s) were not included in the total count.
Antibody staining was performed under the standard Ruvkun-Finney protocol (Anatomical Methods, [92]) using GT335, an antibody against branch point single and polyglutamylated tubulin. Both primary (GT335, mouse, 1∶100, Enzo Life Sciences) and secondary antibody (Alexa Flour 568 goat anti-mouse, 1∶2000 dilution, Life Technologies) washes were performed at 4°C overnight. Young adult hermaphrodites were selected for imaging using the number of eggs in the animal—between one and ten—as a proxy for age.
|
10.1371/journal.pgen.1005675 | EEPD1 Rescues Stressed Replication Forks and Maintains Genome Stability by Promoting End Resection and Homologous Recombination Repair | Replication fork stalling and collapse is a major source of genome instability leading to neoplastic transformation or cell death. Such stressed replication forks can be conservatively repaired and restarted using homologous recombination (HR) or non-conservatively repaired using micro-homology mediated end joining (MMEJ). HR repair of stressed forks is initiated by 5’ end resection near the fork junction, which permits 3’ single strand invasion of a homologous template for fork restart. This 5’ end resection also prevents classical non-homologous end-joining (cNHEJ), a competing pathway for DNA double-strand break (DSB) repair. Unopposed NHEJ can cause genome instability during replication stress by abnormally fusing free double strand ends that occur as unstable replication fork repair intermediates. We show here that the previously uncharacterized Exonuclease/Endonuclease/Phosphatase Domain-1 (EEPD1) protein is required for initiating repair and restart of stalled forks. EEPD1 is recruited to stalled forks, enhances 5’ DNA end resection, and promotes restart of stalled forks. Interestingly, EEPD1 directs DSB repair away from cNHEJ, and also away from MMEJ, which requires limited end resection for initiation. EEPD1 is also required for proper ATR and CHK1 phosphorylation, and formation of gamma-H2AX, RAD51 and phospho-RPA32 foci. Consistent with a direct role in stalled replication fork cleavage, EEPD1 is a 5’ overhang nuclease in an obligate complex with the end resection nuclease Exo1 and BLM. EEPD1 depletion causes nuclear and cytogenetic defects, which are made worse by replication stress. Depleting 53BP1, which slows cNHEJ, fully rescues the nuclear and cytogenetic abnormalities seen with EEPD1 depletion. These data demonstrate that genome stability during replication stress is maintained by EEPD1, which initiates HR and inhibits cNHEJ and MMEJ.
| The cell itself damages its own DNA throughout the cell cycle as a result of oxidative metabolism, and this damage creates barriers for replication fork progression. Thus, DNA replication is not a smooth and continuous process, but rather one of stalls and restarts. Therefore, proper replication fork restart is crucial to maintain the integrity of the cell’s genome, and preventing its own death or immortalization. To restart after stalling, the replication fork subverts a DNA repair pathway termed homologous recombination. Using any other pathway for fork repair will result in an unstable genome. How the homologous recombination repair pathway is initiated at the replication fork is not well defined. In this study we demonstrate the previously uncharacterized EEPD1 protein is a novel gatekeeper for the initiation of this fork repair pathway. EEPD1 promotes 5’ end resection, the initial step of homologous recombination, which also prevents alternative fork repair pathways that lead to unstable chromosomes. Thus, EEPD1 protects the integrity of the cell genome by promoting the safe homologous recombination fork repair pathway.
| Maintaining genome stability depends on faithful DNA replication [1–3]. Since DNA damage from endogenous and exogenous sources creates barriers for the replication fork, replication is not a smooth, continuous process, but rather one of intermittent stress, with stops and restarts [4–6]. Replication fork reactivation after stalling at DNA damage is best characterized in E. coli, where forks are restarted by recombination-dependent or -independent pathways requiring RuvABC or the PriA/C complexes, respectively [5–7].
Eukaryotic replication fork restart is more complex and less understood, with the canonical repair pathway mediated by RAD51-dependent homologous recombination (HR) [1–3,8]. HR is best characterized for the repair of DNA double-strand breaks (DSBs). It is initiated by a litany of components mediating 5’ end resection to create 3’ single-stranded (SS) DNA, which then use BRCA2/RAD51 to create heteroduplexes with homologous sequences on sister chromatids [3,4,8–12]. After an invading strand re-initiates DNA synthesis, Holliday junctions may be resolved by either Gen1 or Mus81, with Slx4 serving as a scaffold [11–15]. End resection directs DSB repair toward HR, preventing the competing DSB repair pathway, classical non-homologous end-joining (cNHEJ) from occurring [16–19].
Similar to DSB repair, repair of stressed replication forks also requires 5’ end resection to initiate HR, but how this is regulated in fork repair and restart is less well defined [1–3,16,17]. End resection at a replication fork requires a free DNA double strand (DS) end structure to initiate 5’ exonuclease activity. This DNA DS end can be created at stressed forks in at least two ways: the fork can reverse into a chicken foot structure with a single DS DNA end [2,3,20], or a nuclease can cleave the fork, directly creating a free DS end [3,13,14,17]. If a stressed fork is not repaired in timely manner, it may convert into toxic structures that make fork restart difficult [1,13,14,19], leading to cell death or genome instability and neoplastic transformation [1,4,6].
Repair pathway choice at stalled forks is important for genome stability, because unopposed cNHEJ, as seen in malignancies with inherited deficiencies in HR proteins BRCA1 or BRCA2, results in fusion of these DNA DS ends at damaged replication forks [21–26]. These chromosomal fusions cause severe genome instability, resulting in catastrophic mitoses revealed as gross nuclear abnormalities including nuclear bridges and micronuclei [1,21,22,25,27]. The tumor suppressor p53-binding protein 1 (53BP1) promotes cNHEJ at least in part by preventing end-resection. Preventing cNHEJ by repressing 53BP1 rescues HR-deficient cells from these nuclear defects [21–23]
There is accumulating evidence that DSB pathway choice between cNHEJ and HR is mediated by 5’ end resection [16–18]. End resection appears to be a two-step process, with CtIP and Mre11 nucleases responsible for short end resection, and Dna2 and Exo1 catalyzing longer resection for HR [16,17,19,28,29]. It is thought that short end resection may lead to MMEJ and long range end resection to HR [17,19,30,31]. Although it is clear that end resection is important for regulating pathway choice at DSBs, key questions remain on how end resection is initiated at stressed forks.
In this study we identify a previously uncharacterized 5’ endonuclease, EEPD1 (endonuclease/exonuclease/phosphatase family domain-containing 1), by its up-regulation in embryonic stem cells after DNA damage. We found that EEPD1 initiates end resection, thereby enhancing HR at the expense of cNHEJ, and also of MMEJ. Consistent with an upstream role in end resection, EEPD1 depletion markedly reduces stress-induced ATR and Chk1 phosphorylation and the formation of RPA, gamma-H2Ax, and RAD51 foci, while NBS1, 53BP1, and BRCA1 foci are intact. Depletion of EEPD1 results in severe chromosomal abnormalities, made worse by replication stress. This places EEPD1 at the apex of pathway choice in repair of stressed replication forks, where it is required for maintenance of genome integrity.
In a survey of proteins induced by the topoisomerase IIα poison VP-16 in embryonic stem cells, we found that expression of EEPD1, an uncharacterized human protein (Uniprot Q7L989, AAH65518.1), was markedly increased. EEPD1 is a 569 aa protein with two amino terminal helix-hairpin-helix (HhH) DNA binding domains related to RuvA, a carboxy terminal DNase I-like domain that places it in the exonuclease-endonuclease-phosphatase (EEP) family, and a conserved D-D-N/D/E nuclease active site that overlaps with the HhH domain and the DNase I-like domain (S1 Fig) [32]. It is located at 7p14.2, but is not involved in any known neoplastic translocations (Catalogue of Somatic Mutations in Cancer). EEPD1 is evolutionarily conserved from some insects to humans and expressed at variable levels in a wide variety of primary human tissues and human cell lines (S1B Fig). It is more highly expressed in the testis, leukocytes, and brain, as are many other DNA DSB repair components [33,34]. EEPD1 depletion moderately altered cell cycle progression in asynchronous or synchronized cells (S2 Fig), increasing the fraction of cells in S and G2 phases in both situations.
EEPD1 alone is required for proper clonogenecity; plating efficiency is reduced by almost 50% from EEPD1 depletion alone (Fig 1A). EEPD1 deficiency also significantly slows cell growth (Fig 1B), and increases the fraction of cells expressing cyclin A, without an increase in the fraction of cells with phosphorylated histone H3 (Fig 1C and 1D). This suggested a potential role in DNA replication. To investigate whether EEPD1 is important for survival after exposure to agents that stress replication forks, we tested whether EEPD1 regulates sensitivity to VP-16, hydroxyurea (HU), camptothecin (CPT), UV light, cisplatin, and ionizing radiation (IR) (Fig 1E). EEPD1 depletion resulted in 3.5-fold less clonogenic survival after 18 h exposure to 10 uM VP-16, compared to controls. EEPD1 depletion also decreased survival to continuous 0.4 mM HU (12-fold), 18 h exposure to 10 uM CPT (6-fold), continuous 0.4 mM HU (10-fold), 18 h exposure to 10 uM CPT (6-fold), continuous 5 uM cisplatin (4-fold), 15 J/m2 UV (12-fold), and 4 Gy IR (4-fold).
To investigate the mechanism by which EEPD1 promotes cell survival during replication stress we used two techniques to measure replication fork restart after stalling. First, BrdU incorporation into nascent DNA after release from HU replication stress was measured by immunofluorescence ([35,36]. By 2 h after release from an 18 h HU exposure, when replication fork restart was maximal in control cells (as indicated by the number of BrdU foci), EEPD1-depleted cells restarting forks were reduced by 5-fold (Fig 2A). This is a specific EEPD1 effect, as the fork restart defect in EEPD1 depleted cells was rescued by expression of an siRNA-resistant version of EEPD1 (Fig 2A).
We next used DNA fiber analysis to measure replication fork restart after release from a 1 h HU treatment, as well as replication speed and replication fork symmetry [32,35]. We found that 20 min after HU release, EEPD1 depletion reduced replication fork restart by 2.3-fold (Fig 2B and 2C). Interestingly, over-expressing EEPD1 increased fork restart; however, by 30 min nearly all forks restarted even in EEPD1-depleted cells. New fork initiation is rare under these conditions, and EEPD1 depletion had no significant effect on this endpoint (Fig 2C). By measuring fiber lengths, we determined that EEPD1 depletion significantly reduces replication speed (Fig 2D). Consistent with EEPD1 promoting fork restart, EEPD1 depletion significantly reduced the percentage of bidirectional forks, reflecting restart at both ends of a replicon (Fig 2D). These results indicate that EEPD1 accelerates restart of stressed replication forks, and that it increases the speed of replication during recovery from stress, implying that EEPD1 also assists in normal fork progression.
Based on the above observations, we investigated the role of EEPD1 in the major DNA DSB repair pathways by using two previously described assays. EEPD1 depletion increased cNHEJ by 2.3-fold in the EJ5 cell reporter system (Fig 3A) [37,38], implying that EEPD1 inhibits cNHEJ. EEPD1 depletion reduced HR repair of I-SceI induced DSBs by 6.4-fold in the HT256 reporter system (Fig 3B) [39]. This reduction in HR raised the question of whether EEPD1 depletion increased gene conversion tract lengths. Cells with defects in HR components display longer gene conversion tracts among residual HR products [30,31,40–46]. Consistent with these prior studies, HR products from EEPD1-depleted cells had significantly longer conversion tracts compared to controls (Fig 3C and 3D). The longer gene conversion tracts are thought to reflect unstable heteroduplexes [40] and/or defective resection preventing efficient 5’ end-capture by the invaded template [30,46]. In the case of EEPD1 depletion, we hypothesize that defective end resection (shown below) results in less efficient 5’ end-capture. If true, then this implies that less efficient SS end-capture reactively stimulates synthesis along the invaded template, an idea supported by several published reports [30,31,42].
Cells with HR defects, such as those with BRCA1 or BRCA2 mutations, are hypersensitive to PARP1 inhibitors, due to an increase in unrepaired DSBs arising during replication [47–49]. We therefore repressed EEPD1 in BRCA1/2 proficient cells and assessed the effect of the PARP1 inhibitor olaparib on cell survival. EEPD1 repression markedly increased the cytotoxicity of olaparib (19-fold, Fig 3E), in the absence other genotoxins, consistent with EEPD1 playing a significant role in HR repair.
There are two DSB repair pathways that use 5’ end resection to initiate the repair cascade, microhomology-mediated end joining (MMEJ), and HR. The frequency of utilization of these two pathways can be compared at a single induced DSB in the EGFP-based MMEJ/HR-Mlu1 reporter (Fig 4A) [19]. Upon DSB induction with I-SceI transduction, repair by either MMEJ or HR results in loss of the I-SceI site and generation of EGFP, allowing repaired cells to be sorted by flow cytometry (Fig 4B). The repaired EGFP loci were PCR amplified, and analyzed for repair by HR versus MMEJ. Cells repaired by MMEJ have a 9 nt duplication containing a BssHII site, while cells repaired by HR have an MluI site (Fig 4C). The fraction of BssHII cleaved products among the total PCR products represent the fraction repaired by MMEJ, while the fraction cleaved by MluI represents HR repair.
Depletion of EEPD1 resulted in an average decrease of 2.5-fold in EGFP-positive cells in the MMEJ/HR-MluI reporter system. The EGFP locus was PCR amplified from EGFP-positive cells, and digested with BssHII or MluI. This revealed that EEPD1 depletion resulted in a 9-fold reduction in HR and a 50% increase in MMEJ (Fig 4D). This implies that when MMEJ and HR are competing at a single DSB site, EEPD1 pushes that repair decision towards HR, and away from MMEJ. This may mean that EEPD1 is important for initiating long range end resection used in HR, consistent with its interaction with Exo1/BLM as noted below.
Since the key step for determining DSB and replication fork repair pathway choice is 5’ end resection [16–18], we therefore assessed the role of EEPD1 in 5’ end resection after DSB formation using two approaches. First, we measured the generation of SS DNA at IR-induced DSBs by immunostaining newly incorporated BrdU in non-denatured SS DNA [50,51]. We found that depletion of EEPD1 reduced the number of cells with SS BrdU after IR by 5-fold (Fig 5A and 5B). In the second approach, we assessed resection around an induced I-SceI DSB [52,53]. Using this technique, we found that EEPD1 depletion reduced end resection by 3-fold, nearly the same extent as CtIP depletion (Fig 5C). Thus, EEPD1 is important for end resection after both a transduced restriction enzyme (I-SceI) and exogenous IR.
We next tested whether EEPD1 functions in the same end resection pathway as Exo1 or CtIP (Fig 5C). We depleted EEPD1 with or without co-depletion of Exo1 or CtIP. There is no significant difference in end resection when EEPD1 and Exo1 are co-depleted compared to individual depletion, suggesting that these enzymes function in the same resection pathway. Co-depletion of EEPD1 and CtIP also yielded similar results as individual depletions. These results suggest that EEPD1 functions in the same resection pathway(s) as Exo1 and CtIP.
When a replication fork collapses, SS DNA arises by end resection, or by uncoupling of the polymerase complex from the helicase [10,54,55]. Such ss DNA is coated by RPA, which recruits ATRIP, leading to ATR activation and phosphorylation of RPA, H2Ax, and Chk1 which mediate cell cycle arrest and replication fork repair [56–58]. To define the epistatic position of EEPD1 in HR, confocal immunofluorescence microscopic studies of fork repair components were performed in cells treated with HU for a prolonged period, which causes replication fork collapse. We found that EEPD1 depletion significantly decreased foci formation by RPA32 (2.7-fold), gamma-H2Ax (3.3-fold), and RAD51 (3.5-fold) (Fig 5D and 5E). The decrease in RPA32 foci was consistent with its decreased phosphorylation, detected by Western blotting (Fig 5F and 5G). Consistent with the decreased formation of RPA32, gamma-H2Ax, and RAD51 foci, and the requirement for SS DNA to trigger ATR and Chk1 signaling, EEPD1 repressed cells also showed decreased phosphorylation of ATR (3-fold) and Chk1 (11-fold) (Fig 5F and 5G). EEPD1 did not co-localize with gamma-H2Ax foci after damage (S3 Fig), not surprisingly, since EEPD1 appears to act upstream of gamma-H2Ax.
The Mre11-Rad50-NBS1 (MRN) complex is a first responder to DSB damage [59–61]. NBS1 recruits BRCA1 to stalled replication forks in an alternative pathway to the canonical gamma-H2Ax/MDC1/RNF8/BRCA1 recruitment pathway [62,63]. Confocal immunofluorescence studies were performed to investigate whether EEPD1 depletion impairs these early regulators of DSB repair. While HU-induced RPA32, gamma-H2Ax, and RAD51 foci were significantly decreased in EEPD1 depleted cells (Fig 5D and 5E), BRCA1, 53BP1, and NBS1 foci were unaffected (S3 Fig), indicating the initial 53BP1 recruitment step of cNHEJ was functional, and that NBS1 and BRCA1 have upstream roles in repair of stressed forks. These data imply that MRN/BRCA1 and EEPD1 act in distinct repair pathways.
For 5’ end resection to take place at a stalled replication fork, there must be a free DNA DSB end [3,4,8–12]. This can occur via fork reversal to form a chicken foot structure, but the majority of stalled forks do not reverse [20]. For those stalled forks that do not reverse the fork must be nicked to create this required free DNA DSB end [3,13,14,17]. If this is true, then DNA nicking should be increased after HU nucleotide depletion to stall replication forks. We assessed the occurrence of DNA nicking using alkaline single cell electrophoresis assays with and without EEPD1 depletion (Fig 5H). We found that after HU exposure, DNA nicking increases 3.5-fold. However, EEPD1-depletion completely abolishes this increase, implying that EEPD1 is directly or indirectly responsible for DNA nicking in response to HU-induced replication stress.
EEPD1 is expressed primarily in the nucleus, consistent with it functioning as a nuclease (S4 Fig). Since EEPD1 has homology to RuvA, which binds to heteroduplex chicken foot structures, and it has a nuclease domain, we examined whether EEPD1 has nucleolytic activity on chicken foot structures. We found that recombinant EEPD1 protein did not nick any of the four double-stranded regions of the regressed fork, but it does have specific 5’ overhang endonuclease activity (S5 Fig), cleaving at a single site at the joint of the overhang. Chicken foot structures with 5’ overhangs are difficult for Exo1 to process [64]; EEPD1 could promote further 5’ end resection by presenting Exo1 with a more amenable structure. These data demonstrate that EEPD1 is a 5’ endonuclease, consistent with the marked reduction in HU-induced nicks in EEPD1 depleted cells.
We next assessed the effect of EEPD1 depletion and co-depletion of three other resection components, CtIP, Exo1, and Dna2, on cell proliferation with or without HU-induced replication stress (S6 Fig). EEPD1 depletion alone suppressed cell proliferation, as did depletion of CtIP and Dna2. Co-depletion of EEPD1 with CtIP or Dna2 did not further affect proliferation, with or without replications stress. By contrast, Exo1 depletion only modestly suppressed proliferation, and only with replication stress. Again, there was no further effect on proliferation with co-depletion of EEPD1 and Exo1 than with EEPD1 depletion alone.
We also compared the effects of EEPD1, CtIP, Exo1, and Dna2 depletion singly and in pairs on the formation and resolution of after HU-induced replication stress (S7 Fig). As above (Fig 5D and 5E), EEPD1 depletion strongly suppressed the formation of gamma-H2Ax foci 4 and 24 h after HU, with similar or greater effects than CtIP depletion. Co-depletion of EEPD1 and CtiP did not further suppress gamma-H2Ax foci 4 h after HU. Exo1 depletion had similar effects as CtIP depletion, with or without co-depletion of EEPD1. Dna2 depletion enhanced gamma-H2Ax focus formation, even in the absence of replication stress, indicating that Dna2 plays a key role in preventing endogenous DNA damage. Co-depletion of Dna2 and EEPD1 did not significantly suppress gamma-H2Ax foci in untreated cells or after HU exposure. The enhanced gamma-H2Ax foci with co-depletion of EEPD1 and CtIP may also reflect enhanced or more persistent DNA damage caused by the genomic lesions, independently of induced replication stress.
Western analysis revealed that after replication stress with HU, EEPD1 is enriched in the nuclear chromatin fraction (Fig 6A and 6B), suggesting that EEPD1 is recruited to chromatin containing damaged replication forks. We next assessed EEPD1 recruitment to stalled replication forks using Isolation of Proteins on Nascent DNA (iPOND) [65,66]. iPOND showed that EEPD1 is recruited to replication forks within 30 min of HU treatment, coinciding with the appearance of gamma-H2Ax (Fig 6C) which marks DSBs at stalled/collapsed replication forks [65]. PCNA was absent from the HU-stalled forks, consistent with replisome unloading from collapsed fork Okazaki fragments [65]. A control iPOND assay using a thymidine chase confirmed the specificity of EEPD1 recruitment to stalled forks (Fig 6D). By using chromatin immunoprecipition [52] we also demonstrated that EEPD1 is recruited to an I-SceI induced DSB (Fig 6E). Interestingly, EEPD1 constitutively co-immunoprecipitates with Exo1, RPA32, and BLM in the presence of DNase, whether or not replication stress is present, indicating that these proteins reside in the same complex (Fig 6F and 6G). However, EEPD1 does not co-immunoprecipitate with Dna2, indicating that it is likely not in the RPA/Dna2/MRN end resection complex [28]. Significantly, depleting EEPD1 reduced Exo1 and BLM protein levels, suggesting that EEPD1 promotes stability of the complex in which EEPD1, Exo1, and BLM reside (Figs 6H and S6B), indicating that EEPD1 resides in an obligate complex, perhaps to prevent aberrant nuclease activation and improper DNA cleavage.
Proper replication stress responses are required to prevent gross chromosomal instability, which can be assessed by the formation of micronuclei and nuclear bridges that result from mis-segregation of fused chromosomes [27,67]. We found that EEPD1-depleted cells display severe nuclear anomalies, with 6- and 7-fold increases in nuclear bridges and micronuclei, respectively (Fig 7A–7D). Chromosome fusion events occur when collapsed forks are aberrantly repaired, as in BRCA1-deficient cells with unopposed 53BP1 [22,23]. 53BP1 depletion alone did not alter nuclear anomalies, but 53BP1 depletion largely suppressed both bridges and micronuclei associated with EEPD1 depletion.
Metaphase analysis further demonstrated EEPD1 repression causes genome instability, revealed as significant increases in chromatid breaks and radial chromosomes, both of which arise in S/G2 cells (Fig 7E and 7F). Interestingly, IR induced these S/G2-associated events, and at lower frequencies, G1-associated chromosome breaks and double minutes. EEPD1 depletion alone did not increase G1-associated events, nor did it affect the frequency of IR-induced these events (Fig 7F). Thus, EEPD1 specifically suppresses S/G2 events. Although more chromatid breaks were observed in EEPD1-depleted cells treated with HU than with HU alone, the difference was not significant (P = 0.17); EEPD1 depletion did significantly increase IR-induced chromatid breaks compared to IR alone. Interestingly, 53BP1 repression fully suppressed spontaneous and HU-induced chromatid breaks seen in EEPD1 depleted cells as well as HU-induced chromatid breaks in cells with normal EEPD1 expression (Fig 7F). These results indicate that EEPD1 plays a critical role in maintaining genome stability, under stressed and non-stressed conditions, and suggest that EEPD1 promotes genome stability by mediating accurate HR repair of stressed replication forks.
Cancer cells experience continuous replication stress due to metabolic alterations and checkpoint defects that permit DNA replication despite significant DNA damage. To manage this stress, it is reasonable to suppose that EEPD1 would be up-regulated in cancers. We tested this by analyzing mRNA expression of EEPD1 in newly resected colorectal cancer versus adjacent normal tissue. In the present study we analyzed EEPD1 expression in 181 new colorectal cancers, and found that EEPD1 was expressed an average of 2.3-fold higher than adjacent normal tissue in 171 of 181 cases (Fig 7G).
This study demonstrates that the uncharacterized EEPD1 nuclease plays a key role in repairing stressed replication forks via HR. Interestingly, while EEPD1 confers resistance to replication stress, it only appears to accelerate fork restart by 10 min. Thus, nearly all forks still restart within 30 min of release from stress in EEPD1-depleted cells compared to 20 min in wild-type cells (Fig 2B and 2C). These results imply that when fork repair is repressed, even a relatively brief delay in fork restart can be lethal, perhaps because toxic recombination intermediates form if stalled forks fail to restart in timely manner [12,32,35]. EEPD1 is also important under non-stress conditions, as EEPD1 depletion significantly slows cell growth rate. Thus, EEPD1 probably promotes restart of replication forks that encounter DNA lesions arising spontaneously during normal cellular metabolism.
There are numerous reports demonstrating that most cells do not tolerate long delays in the restart of stalled replication forks, if repair is impaired. BRCA1 deletion results in cell death after even a brief period of replication stress [68]. There is evidence that forks blocked by interstrand crosslinks are restarted via lesion by-pass long before the lesion is repaired [69]. Thus, restarting replication forks appears to be a higher priority for the cell than repair, at least in some situations. When an ATR inhibitor is combined with a fork stalling agent, cells in S-phase lose all ability to recover within 45 min [70]. Removing the replication stalling agent and ATR inhibition after that point does not restore cell viability. Reintroduction of the DNA damage checkpoint in yeast mutants after a brief period of replication stalling does not rescue cell viability [71]. Our data here show that the largest effect of EEPD1 on stalled fork repair and restart is 10 min after HU release, yet EEPD1 is required for survival to many replication stress agents. Thus, when replication fork repair is impaired, even a brief period of fork stalling can be lethal.
Interestingly, cells proliferate more slowly when EEPD1 is depleted. In addition, even in the absence of replication stress with HU, cells with depleted EEPD1 have increased gamma-H2Ax and decreased RPA foci. Thus, cells lacking EEPD1 appear to experience spontaneous replication stress. This may be due to EEPD1 having a role in normal replication fork progression, perhaps in nucleolytic processing of replication fork lagging strand intermediates [1–3,13,14,16,17]. A not mutually exclusive alternative is that EEPD1 promotes restart of forks stalled by spontaneous lesions, which may arise frequently in rapidly growing cultured cells. This would require constant repair of replication forks stalled by the continuously generated DNA lesions within their paths, which would be reliant on EEPD1.
EEPD1 represses cNHEJ and enhances HR rates significantly. This indicates that EEPD1 plays an important role in DNA DSB pathway choice. By promoting 5’ resection EEPD1 would direct repair away from cNHEJ and towards resection-dependent repair pathways, namely HR and MMEJ [3,8–10,12,16,17,19]. PARP1 competes with the Ku complex for free DNA DSB ends to promote MMEJ over cNHEJ [72]. PARP1 inhibition with olaparib was synthetic lethal with EEPD1 depletion. This implies that EEPD1 depletion is not epistatic with PARP1 in replication fork repair or in MMEJ repair [49,73], given that there is additional cytotoxicity when both are repressed.
Both HR and MMEJ require 5’ end resection to begin their repair cascades [17,19]. Both pathways can repair and restart replication forks after stalling [19]. Recent reports indicate that DNA polymerase (pol) theta may suppress HR and promote MMEJ repair of DNA breaks [74]. HR-deficient tumors rely on pol theta for DNA DSB repair [75]. Pol theta enhances MMEJ by tethering free DSB ends after short range end resection for the microhomology search [76]. EEPD1 would seem to have the opposite effect, promoting HR at the expense of MMEJ (Fig 4). This would be beneficial because unopposed pol theta−mediated MMEJ repair of replication forks would increase non-conservative repair, and more importantly, chromosomal fusions [74,75]. It is possible that EEPD1 promotes HR over MMEJ by enhancing long range end resection [17,19,30,31,76], perhaps via its interaction with Exo1. Exo1 seems to be important for long range end resection during HR [17,19,77].
End resection promoting HR at DSBs is thought to initiate when BRCA1/CtIP displaces Rif1/53BP1 at DSBs [23–26]. CtIP has an important non-nuclease role in initiating 5’ end resection at undamaged DSB ends. CtIP may also have nuclease activity important for resection of damaged DSB ends, but this is controversial [24,30,60,78]. Thus, in addition to its role in cleavage of stressed replication forks, EEPD1 may also be important for initiating end resection at undamaged DSB ends in HR, where CtIP may not have a role. While the competing activities of Rif1/53BP1 and BRCA1/CtIP determines cNHEJ vs HR repair pathway choice, less is known about the role of these components in the repair decisions at stalled replication forks. The data here implies that EEPD1 directs the cell away from cNHEJ towards HR (Fig 3A and 3B), probably by enhancing end resection. EEPD1 functions in a distinct end resection pathway from BRCA1/CtIP, or downstream of BRCA1/CtIP, since HU-induced BRCA1 foci are intact in EEPD1-depleted cells (S3 Fig).
Replication forks require a free DNA end with which to initiate 5’ end resection for repair by either HR or MMEJ [8,10,13,16,20]. For approximately one quarter of replication forks stalled with HU, this occurs via fork reversal, with Dna2 then mediating 5’ end resection [20,79]. However, the majority of replication forks require a nick in one of the parent strands at the replication fork itself to create a free DNA end. HU-induced replication stress results in rapid DNA nicking, which is mediated by EEPD1 (Fig 5H). Given that EEPD1 is rapidly recruited to stalled replication forks, promotes 5’ end resection, HR, and replication fork restart, it is possible that the lack of DNA nicking seen after EEPD1 depletion is due to the failure of replication fork cleavage. This failure to cleave the stressed fork may prevent 5’ end resection for HR-mediated fork repair and decrease fork restart. This implies that some stressed replication fork nicks, rather than contributing to cell death, may instead promote cell survival by accelerating fork restart, perhaps by preventing accumulation of toxic HR intermediates [4,6]. Since many stalled replication forks do not reverse to form a one sided DNA free end for end resection [20], such cleavage is often necessary to initiate end resection and HR [8,10,13,16].
Placing the various end resection nucleases epistatically within the context of 5’ end resection is challenging [13,14,17]. From a biochemical standpoint, there appear to be two end resection complexes, BLM-DNA2-RPA-MRN and EXO1-BLM-RPA [28], with EEPD1 as a component of the latter. However, there is functional overlap between these complexes, and both are likely essential for HR and cell survival in response to replication stress [16,18]. From the data presented here, there is little additional deficiency in end resection after replication stress when EEPD1 is doubly depleted with Exo1 or CtIP. Interestingly, Dna2 depletion increases gamma-H2Ax formation de novo and after replication stress, while Exo1 and EEPD1 depletion reduce gamma-H2Ax formation after replication stress. Thus, Dna2 appears to operate downstream of gamma-H2Ax signaling, while EEPD1/Exo1 are upstream, but perhaps both complexes are needed to repair distinct forms of stressed replication forks [20].
End resection creates SS DNA that can signal replication stress and cell cycle arrest. EEPD1 depletion abrogates gamma-H2Ax foci (Fig 5D and 5E), indicating resection promoted by EEPD1 precedes phosphorylation of H2Ax during replication fork repair. Similarly, ATR is activated by RPA/ATRIP loading onto SS DNA, ultimately activating Chk1. There are two possible explanations for the finding that EEPD1 is required for ATR/gamma-H2Ax/Chk1 phosphorylation after HU. First, the resection defect in EEPD1-depleted cells may account for a fraction of the ATR and Chk1 activation defects after replication stress. In this scenario the SS DNA created by end resection plays a key role in RPA/ATRIP activation of ATR. Second, the SS DNA that signals ATR activation may arise not from end resection but from the disassociation of the helicase from the polymerase complex [54], and in this case EEPD1 might play a role in RPA/ATRIP signaling to ATR. In either case, at least for replication stress induced by HU, EEPD1 is an important factor in the activation of ATR/Chk1.
Interestingly, EEPD1 is in a constitutive and obligate complex with Exo1 and BLM (Fig 6F and 6G). The long resection exonuclease Exo1 requires a free 5’ DNA end to initiate resection at a damaged fork. Since EEPD1 and Exo1/BLM constitutively co-immunoprecipitate, this implies that EEPD1 is a partner in the Exo1/BLM/RPA end resection complex [28]. It also implies that when EEPD1 is recruited to the damaged fork, it is accompanied by Exo1/BLM, which are needed for completion of end resection. The obligate nature of this complex is not surprising, since nuclease function must be exquisitely balanced to prevent wide-spread and unregulated genome incision. It is imperative the cell tightly control all nucleases to prevent inappropriate or untimely DNA cleavage to suppress translocations, and to maintain genome stability.
Our results also indicate that EEPD1 helps maintain genome stability. As proposed for cells with defects in other HR proteins like BRCA1 and BRCA2 [21–23], the genome instability seen in EEPD1 depleted cells is likely a direct consequence of unopposed cNHEJ causing aberrant ligation of DNA ends at distinct collapsed replication forks. This hypothesis is supported by the fact that genome instability observed in EEPD1 depleted cells is suppressed by depletion of the cNHEJ promoting factor 53BP1 (Fig 7A–7F). There is another potential reason why EEPD1 may prevent chromosomal instability- The MMEJ pathway mediates chromosomal translocation events when replication stress overcomes the ability of the cell to repair such stress [74,75,77] EEPD1 promotes HR over MMEJ (Fig 4) and this could suppress chromosomal instability during replication stress by directing fork repair toward HR which is less prone to chromosomal fusions. Thus, EEPD1 may prevent chromosomal translocations by promoting HR and suppressing MMEJ during repair of chromosomal DSBs. Our results indicate that EEPD1 is an important guardian of genome stability that functions by regulating replication fork repair pathway choice.
Although the mechanism by which EEPD1 depletion sensitizes cells to replication stress is not well defined by the present study, many prior studies that demonstrate HR defects increase sensitivity to replication stress correlate increased sensitivity with increased gamma-H2Ax [reviewed in refs.1,2,3,8]. While it is widely accepted that stressed replication forks can collapse into aberrant structures in cells lacking HR machinery to repair them [1,13,14,19], whether these aberrant structures actually cause cell death is not known. We demonstrate here that EEPD1 depletion increases cell death in the face of replication stressors, and that it reduces HR, slows replication fork restart, reduces DNA nicking, and creates cytogenetic and nuclear abnormalities. The marked increase in micronuclei and mitotic bridges in EEPD1 depleted cells is exacerbated by replication stress, and this suggests the following mechanism for cell death from EEPD1 deficiency during replication stress: collapsed replication forks end-join aberrantly, creating chromosomal fusions that are manifest as mitotic bridges and micronuclei. These gross chromosome abnormalities would intuitively be difficult for cells to recover from, and therefore may serve as a better correlation between HR and cell death than gamma-H2Ax.
One would predict that malignancies would require intact EEPD1 to proliferate, and that loss of function mutations would be rare in human cancers. This is indeed the case; there are only 31 coding changes in EEPD1 out of 8273 individual cancer genome sequences in the COSMIC database (http://cancer.sanger.ac.uk/cosmic), and the vast majority of these are conservative, and are not predicted to alter function. This is not surprising, as a malignancy with EEPD1 functional loss would have difficulty proliferating given that tumors often survive despite significant replication stress caused by oncogene activation, hypoxia, and/or nutrient deprivation [6]. Thus, even though loss of EEPD1 results in genomic instability, EEPD1 should not be viewed as a tumor suppressor in the same sense as BRCA1 and BRCA2, two HR components that show loss of function mutations in cancer. It is likely that loss of EEPD1 function would be too detrimental to replication fork restart and fork progression to be selected for during oncogenesis, because of the fundamental importance of DNA replication to malignant cells.
On the other hand, given its role in replication fork rescue, EEPD1 could be an excellent target for treatment of human malignancies. EEPD1 is over-expressed in nearly all colorectal cancers [80] (Fig 7G) and large cell lymphomas [81], cancers whose treatment is based on agents that create replication stress. Targeting EEPD1 could block proliferation of cancers that depend on EEPD1, or sensitize tumors to chemotherapeutics that cause replication stress. Such agents are the foundation for treating both of these types of malignancies [81–83].
A549, HEK-293, HEK-293T, HT256 reporter cells, and the various U2OS reporter cells (EJ5, MMEJ/HR), were cultured in D-MEM supplemented with 10% fetal bovine serum and 1% penicillin and streptomycin. HT256 cells were cultured in Alpha-MEM supplemented with 10% fetal bovine serum and 1% penicillin and streptomycin. EEPD1 was depleted using two mechanisms, shRNA and siRNA, to control for variation in the method of mRNA destruction. EEPD1 was depleted either by 1) EEPD1 lentivirus shRNAs produced from 293T cells (pLKO.1, Thermo Scientific, Pittsburgh, PA); or 2) SMARTpool ON-TARGETplus EEPD1 siRNA from Dharmacon RNAi Technologies (GGACUGACCUUCACCGCCA; CUGAGAAGCCCUCGAGUCA, GGAAGUUGACCUCGGGGUA; UGCGAGAGGUGGUGUGCAU) (Pittsburgh, PA). EEPD1 3’UTR On-Target plus siRNA also from Dharmacon (GGAAGUUGACCUCGGGUA). 53BP1siRNA(h) is a pool of 3 different siRNA duplexes from Santa Cruz Biotechnology (sc-37455). All other siRNAs were from Dhamarcon SMARTpools. All nucleic sequences are listed 5’ to 3’ in this Supplement.
Polyethylenimine (PEI) was used to perform plasmid transfections according to the manufacturer’s instructions (Thermo Scientific). Briefly, PEI was incubated with plasmid DNA at 3:1 ratio in Opti-MEM at RT for 20 min before addition to cells. After 6 h incubation, cells were washed and placed in fresh media. RNAiMAX (Invitrogen, Grand Island, NY) was used to transfect siRNA pools. Briefly, RNAiMAX was incubated with 50 nM of siRNA in Opti-MEM at RT for 20 min before addition to cells. After 24 h cells were washed and placed in fresh media. EEPD1 repression was confirmed by western blotting for every experiment. At least two clones were used for each Lentiviral shRNA experiment, to control for clonal variation in repression. There was no difference in phenotypes obtained between the shRNA and the siRNA repression of EEPD1.
Experiments were repeated using both techniques for EEPD1 depletion to control for off-target effects of the mechanism of repression. All experiments were performed at least three times, in at least two cell lines, to control for experimental and cell line variation. A549 lung cancer cells were used for most replicative experiments since this cell line has high EEPD1 expression. All studies in A549 were also repeated at least once in HEK-293 cells to control for cell lineage variability. Clonal survival after treatment with DNA damaging agents was determined by seeding 2,000 cells per 10 cm dish in either control media or media with varying concentrations of genotoxic chemicals, or exposure to varying doses of IR or UV light. Cells were exposed to etoposide or olaparib for 18 h, then washed and incubated in fresh media for 12 days. Colonies were stained with 0.1% crystal violet in methanol and counted. A colony greater than 50 cells was counted as a surviving clone. For HU, cells were treated continuously for 12 days before colonies were stained and scored. Plating efficiency was calculated as the number of colonies divided by the number of cells plated without genotoxin treatment. In all of these assays, survival was normalized to untreated cells transfected with control or EEPD1 si or shRNAs. Survival fractions were calculated as the number of colonies formed after exposure to a given genotoxin divided by the number of cells plated, then multiplied by the plating efficiency. Unpaired Student t tests were used for all statistical analysis, unless otherwise indicated. Each experiment was performed 6–9 times in triplicate.
EEPD1 expression was monitored by standard western blotting protocol [36] using a custom-produced rabbit polyclonal antibody to EEPD1 peptide (CAEFYTEKDWSKKDAPRNHS, Lampire Biological Laboratories, Pipersville, PA). Phosphorylated RPA32 (S4/S8) and total RPA32 antibodies were from Bethyl Laboratories (Montgomery, TX). Phosphorylated ATR (T1989) antibody was from Genetex (Irvine, CA, cat. Gtx128145). Total ATR, phosphorylated Chk1 (S345), and total Chk1 antibodies were from Cell Signaling Technology (Danvers, MA). 53BP1 and BLM antibodies were from Abcam (Cambridge, MA), Exo1 antibody was from Proteintech (Chicago, IL), and beta-actin antibody was from Sigma-Aldrich (St. Louis, MO). When protein levels were quantitated, each western analysis was performed at least 3 times, with densitometric measures of band intensities normalized to loading controls. Student t tests were used for statistical analysis of the protein intensity differences.
Immunoprecipitation was performed with the Pierce Crosslink Magnetic IP/Co-IP kit according to manufacturer’s instructions (Thermo Scientific Cat.88805) as we described [36]. Briefly, HEK-293 cells overexpressing V5-tagged EEPD1 were treated, harvested and washed by PBS before lysis using IP lysis/wash buffer, then 5 ug of V5 mouse antibody (Invitrogen) were coupled to protein A/G magnetic beads and cross-linked with 20 uM disuccinimidyl suberate. The antibody cross-linked beads were incubated with cell lysate (0.8–1.2 mg) in a 500 ul of diluted lysate solution for 1 h at RT on a rotator. Beads were collected, washed and incubated with 100 ul of elution buffer for 5 min at RT. Antigen recovery was achieved by collecting the supernatant on a magnetic stand. Protease and phosphotase inhibitors were present in all buffers. ChIP was performed in HT256 cells using the procedure and GAPDH primers as we described [52]. ChIP primers for neo in HT256 152 nt from the I-SceI DSB site: Neo671 Forward: GACGGGCGTTCCTTGCGCAGCTG; Neo830 Reverse: CCAGATCATCCTGATCGACAAGAC. Primers 650 nt distant from the I-SceI site: Neo1 Forward: AAGCTTCACGCTGCCGCAAGCAC; Neo152 Reverse: GAACCTACCTGCTTTCTCTTTGC. GAPDH Forward: TCGGTTCTTGCCTCTTGTC; GAPDH Reverse CTTCCATTCTGTCTTCCACTC. Each immunoprecipitation was performed at least 3 times. Real time PCR to quantify immunoprecipitated sequences was performed using the SYBR green reagent (Applied Biosystems, Thermo Scientific) with the ABI 7000 sequence detection system, normalized to GAPDH amplification.
Two methods were used for measuring stalled replication fork restart. In the first method, replication fork restart after arrest was measured by immunofluorescent detection of BrdU foci after DNA denaturation (BrdU in DS DNA), as we described previously [36]. Log phase A549 cells expressing normal or repressed levels of EEPD1, with or without expression of siRNA-resistant FLAG-tagged EEPD1 were incubated with 10 mM HU for 18 h and then released into media with 10 uM BrdU for 30 min. After washing, cells were fixed at different time points. Replication recovery was shown as percentage of cells with ≥ 3 BrdU foci 2h after release from HU. Cells without HU treatment served as controls for background staining from normal cell proliferation, which was used as threshold for measurement. Values are averages (± SEM) for 11–23 distinct determinations (>100 cells scored per condition).
The second method was DNA fiber analysis, as we previously described [32,35]. Both A549 and HEK-293 cells were tested to control for cell line differences. 600,000 cells were incubated overnight at 37°C in six-well plates. 20 mm IdU was added to growth medium and incubated for 20 min at 37°C. The IdU media was removed and cells washed in fresh medium, cells were treated with 5 mm HU for 60 min or mock-treated. The HU-containing medium was replaced with fresh medium containing 100 mm CldU. Cells were then incubated for varying times at 37°C. The CidU medium was removed, cells harvested, resuspended in PBS, and 1,000 cells were transferred to a positively charged microscope slide (Superfrost/Plus, Daigger), and processed for DNA fiber analysis as we described previously [32]. Slides were mounted in PermaFluor aqueous, self-sealing mounting medium (Thermo Scientific), and DNA fibers were visualized using a confocal microscope (Olympus, FV1000D, 63× oil immersion objective). Images were analyzed using the Olympus Fluoview software.
Confocal immunofluorescence foci assays were performed as we described [35] with minor modifications. In brief, cells were cultured on coverslips followed by siRNA transfection and HU treatment. Cells were pre-extracted with 0.5% Triton X-100 and fixed with 4% paraformaldehyde for 20 min. Coverslips were then blocked with 1% BSA for 1 h before incubating with primary antibodies overnight. After washing twice, coverslips were incubated with secondary antibodies conjugated with Alexa Fluor dye (Invitrogen), mounted in anti-fade solution containing DAPI and stored at 4°C. All samples were analyzed within 24 h with a laser confocal scanning microscope (TCS-SP5, Leica Microsystems, Exton, PA). Cells with >5 foci were counted as positive. Photomicrographs of distinct cell populations were taken at equal magnifications and equal fluorescence intensities. For NBS1 and BRCA1 foci, the cells were fixed in 100% methanol and incubated with 1% BSA in 0.1% PBS-Tween for 1 h before incubating with primary antibodies overnight. RAD51 antibody was obtained from Santa Cruz Biotechnology (Dallas, TX). BRCA1 and RPA32 antibodies were obtained from Bethyl Laboratories (Montgomery, TX); gamma-H2AX (S139) antibody from Millipore (Billerica, MA), phosphorylated NBS1 (S343) and BrdU antibodies from Cell Signaling (Danvers, MA), and 53BP1 antibody from Abcam (Cambridge, MA). To assess nuclear structural abnormalities (micronuclei and post-mitotic bridging), control or HU-treated cells, with or without EEPD1 depletion, were grown on coverslips and fixed as above, and stained with 300 nM DAPI (Beckman) in PBS for 5 min. After washing thrice with PBS, coverslips were mounted in anti-fade solution and analyzed within 24 h. Of note, EEPD1 was located in the nucleus, but did not form discrete foci before or after damage. Each immunofluorescence assay was performed at least 3 times in triplicate.
iPOND was performed as described by Sirbu and colleagues[65,66], with minor modifications to improve protein capture. In brief, HEK-293T cells over-expressing V5-tagged EEPD1 were seeded in three 150 mm plates/condition 24 h before the experiment. After 24 h incubation, 10 uM EdU (Invitrogen) was added to the medium for 10 min. EdU treatment was followed with 3 mM HU (Sigma, St. Louis, MO) at indicated times. The cells were then fixed with 1% formaldehyde (Sigma) for 10 min at RT, quenched by 0.125 mM Glycine (Sigma), and collected by scraping. The cells were permeabilized with 0.25% Triton X-100 for 30 min, and then subjected to click–iT reaction using Biotin azide (Invitrogen) for 90 min at room temperature. Lysis conditions were modified to reduce background: lysis was performed in 0.25% SDS lysis buffer for 10 min at RT, followed by sonication at 4°C using Bioruptor (Diagenode) for 10 min with 30 s on/off cycles at high intensity. This treatment consistently yielded fragments between 80–100 bp. Finally, EdU-labeled DNA was pulled down by incubation with Streptavidin-agarose beads (Millipore) overnight at 4°C. The beads were washed once with lysis buffer, once with 1 M NaCl, and twice with lysis buffer. Bound proteins were eluted in 2 x NuPAGE LDS sample buffer (Invitrogen) containing 1 x sample reducing agent (Invitrogen) at 95°C for 35 min before loading for western analysis. Protease and phosphatase inhibitors (Thermo Scientific) were added to all buffers. Each iPOND assay was performed 3 times.
End resection was analyzed using two methods. First, end resection following gamma-irradiation was assessed using a single strand BrdU assay as described [51]. To detect single strand DNA formation, A549 cells were transfected with control and EEPD1 siRNAs, and plated on coverslips at 24 h, then incubated with 30 μM BrdU for 42 h before treatment with 20 Gy IR. At various times after irradiation, cells with native (non-denatured) DNA were analyzed by immunofluorescent confocal microscopy to detect BrdU in SS DNA created by end resection.
Second, end resection was also measured adjacent to a specific I-SceI-induced DSB by quantitative PCR (qPCR) [50,53]. Genomic DNA (gDNA) was extracted from HT1904 cells [52] harvested 4 h after infection with adenovirus vectors that express I-SceI (Adv-I-SceI) or GFP (Adv-GFP) as control. Half of the gDNA was digested with XmaI (NEB), and the remaining half was mock-digested. PCR reactions included XmaI-digested or undigested gDNA as template, 0.5 uM of each primer, 0.2 uM TaqMan probe, and 1X TaqMan universal master mix (ABI). qPCR was performed on a 7900HT Fast Real-Time PCR System (ABI) under standard thermal cycling conditions. Results were analyzed with SDS2.3 (ABI) and Graph Pad 6. For each sample, a ΔCT was calculated by subtracting the CT value of the undigested sample from the CT value of the XmaI-digested sample. The percentage of SS DNA was calculated with the following equation: SS DNA% = 1/(2^(ΔCt-1)+0.5)*100 [50]. Primers and probes were: forward (CGACCTTCCATGACCGAGTACAA), reverse (TCCGGGTCGACGGTGTG), and probe (6FAMACCGCGACGACGTCCCCCGGGCC-TAMRA). All Ct values were corrected for different DNA concentrations, as determined by qPCR of a ‘No Cut’ amplicon on chromosome 22 that lacks XmaI sites: forward (ACATTGTCTCTGTCATGGGC), reverse (TGTGTCAGGGATTTGCTCAC), and probe (6FAM AGAGCATGGGTGGATCCTGGATATTCA-TAMRA). DSB induction by Adenoviral-I-SceI was measured by qPCR and calculated as described [52] using a primer set that flanked the I-SceI site, and primers to the chromosome 22 ‘No Cut’ amplicon as a negative control. The ‘No Cut’ amplicon was used to normalize the amount of genomic DNA in the qPCR reaction, and the percentage of DSBs in Adv-GFP treated cells was set to zero. Both end resection assays were performed three times in triplicate.
Pure recombinant human FLAG-tagged EEPD1 protein was generated in 293 cells and purified as we described [84]. Nuclease assays were performed as we described [32,84]. 3’ overhang reversed fork (“chicken foot”) structures were obtained by annealing SHL101, SHL108, SHL109, and SHL110, and then gel-purifying the annealed structure. 5’ overhang reversed fork structures were obtained by annealing SHL101, SHL108, SHL111, and SHL112 and then gel-purifying the intact annealed structure [32,84]:
SHL101 (60mer): 5’-CGATACTGAGCGTCACGGACTCTGCCTCAAGACGGTAGTCAACGTGTTACAGACTTGATG-3’
SHL108 (60mer): 5’-CTAGACTCGAGATGTCAAGCAGTCCTAACTTTGAGGCAGAGTCCGTGACGCTCAGTATCG-3’
SHL109 (60mer): 5’-CATCAAGTCTGTAACACGTTGACTACCGTCGATCCACTAG AGGTCTAAGCGACCTCATTC-3’
SHL110 (40mer): 5’-CTAGTGGATCAGTTAGGACTGCTTGACATCTCGAGTCTAG-3’
SHL111 (40mer): 5’-CATCAAGTCTGTAACACGTTGACTACCGTCGATCCACTAG-3’
SHL112 (60mer): 5’-AGGTCTAAGCGACCTCATTCCTAGTGGATCAGTTAGGACTGCTTGACATCTCGAGTCTAG
The HT256 reporter system was used to determine I-SceI-induced HR efficiency and gene conversion tract spectra as we described [39,85]. The EJ5-GFP U2OS system was used to assess NHEJ [37,38]. Both of these reporter systems have single, integrated copies of reporters with I-SceI target sites cleaved upon transfection of an I-SceI expression vector. Cells were transfected with either control or EEPD1 siRNAs or shRNAs, and then transfected 24 or 48 h later with pCBA-SceI or empty vector, using PEI. After 48 h incubation, EJ5 cells were trypsinized and washed with PBS and GFP-positive cells reflecting NHEJ frequencies were measured by FACSort (Becton-Dickinson, San Jose, CA) and analyzed with CellQuest (Becton-Dickinson) software. Productive HR in HT256 cells reconstitutes a functional neomycin phosphotransferase gene, generating G418 resistant colonies. Two thousand cells were plated in three 10-cm dishes per each condition, in non-selective media, 24 h after I-SceI vector transfection to establish plating efficiency. To assess HR, 500,000 cells were plated in media with G418 (325 ug/ml, 100% active) added 24 h after transfection. DSB-induced HR frequencies were calculated as the number of G418-resistant colonies per viable cell plated in G418 medium after 21 days, normalized for plating efficiency. HR assays were performed 15 times in triplicate and the NHEJ assays 12 times in triplicate.
Gene conversion tracts were analyzed as we described [39,40,85] on the above HR repaired neo-positive colonies. HT256 G418 resistant colonies were stained and counted, or expanded under continuous G418 selection for gDNA isolation and molecular analysis. Genomic DNA was extracted using the DNeasy Tissue Kit (Qiagen, Valencia, CA). Primers A (CCTTCACTTTCCAGAGGGTC) and B (GCGAAGAACTCCAGCATGAG) were used to amplify a 1.5 kb fragment comprising the recipient neo allele (MMTVneo) by using standard PCR conditions. The donor neo allele carries 12 silent single-base mutations at approximately 100 bp intervals that create restriction fragment length polymorphisms (RFLPs). These RFLP markers allow high-resolution analysis of gene conversion tract length, directionality, and continuity. The 12 silent RFLP markers and the natural BanII site were mapped in PCR fragments amplified from HR products.
The analysis of MMEJ versus HR competitive repair from a single DSB was performed in modified U2OS cells as we described [19]. To directly compare MMEJ with HR, an EGFP-based MMEJ and HR competition reporter system, termed EGFP-MMEJ/HR-MluI, was generated. This reporter had the EGFP (R-EGFP) cassette of EGFP-HR replaced with the EGFP-MMEJ cassette. A unique MluI site in the parent EGFP (D-EGFP) cassette was created via a silent mutation at the BssHII site. Upon I-SceI cleavage, restoration of a functional EGFP cassette results in loss of the I-SceI site after cells undergo repair by either MMEJ or HR. PCR analysis of the sorted green cells using primers specific for R-EGFP was performed. The primers were: EGFP MMEJ/HR Forward:5’-ACGGGGTCATTAGTTCATAGCCCA, EGFP MMEJ/HR Reverse: 5’-GGGATTTTGCCGATTTCGGCC. Repair of the I-SceI DSB by MMEJ would retain one copy of the 9-bp duplication with an intact BssHII site. The percentage of the BssHII-digestible bands within the total PCR amplified product reflects the MMEJ frequency. Repair of that I-SceI-induced DSB by HR transfers the MluI site from D-EGFP to R-EGFP, and thus the percentage of MluI-digestible bands of the total PCR product reflects the HR frequency.
To assess nuclear structural abnormalities, micronuclei and post-mitotic bridging from aberrant chromosomal segregation, control or HU-treated cells, with or without EEPD1 depletion, were grown on coverslips and fixed as above, and stained with 300 nM DAPI (Beckman) in PBS for 5 min [27]. After washing with PBS, coverslips were mounted in anti-fade solution and analyzed within 24 h. Of note, EEPD1 was located in the nucleus, but did not form discrete foci with or without DNA damage. At least ten distinct determinations (142–190 nuclei per determination) were performed for each treatment group.
Structural aberrations in metaphase chromosomes were scored by Solid Giemsa staining as described [86,87]. EEPD1 and/or 53BP1 were repressed using siRNA in log phase A549 cells, with or without 18 h treatment with 10 mM HU. Cells were washed with PBS and fresh media with colcemid (final concentration 0.25 ug/mL) was added, and cells were incubated for 24 h before harvest. Chromosome preparations were made according to the standard air drying procedure as we described [87]. Cells were harvested, washed with pre-warmed PBS twice, hypotonically treated (0.56% KCl, 20 min at 37°C) and subsequently fixed in freshly prepared acetic acid-methanol (1:3). At least three changes of fixative were performed before the cell suspension was dropped on to a pre-cleaned chilled glass slide and dried at RT at least for 1 day before staining. Structural translocations such as dicentric and ring chromosomes, and Robertsonian translocations, were scored under 63x magnification [87]. Statistics were calculated using Fisher exact tests. Cytogenetic spreads were performed three distinct times with a total of 102–374 metaphase spreads were analyzed per condition.
Alkaline single cell electrophoresis assays for SS nicking was performed as described [88] using the CometAssay kit (Trevigen, Gaithersburg, MD). Briefly, A549 cells were transfected with siRNAs and treated with 10 mM HU for 1 h or mock treated. Cells were harvested, washed with cold PBS and mixed with molten 1:10 (v/v) LMAgarose and immediately spread over the sample area of comet slides. Cells were immobilized at 4°C in the dark for 30 min and immersed in lysis solution overnight. For the alkaline comet assay, slides were treated with alkaline unwinding solution for 1 h at 4°C in the dark before electrophoresis in alkaline electrophoresis buffer. Samples were rinsed with water and immersed in 70% ethanol before drying at 37°C for 15 min. SYBR Gold was used to stain dried agarose for 30 min at RT before rinsing and drying again. Slides were viewed with a Leica inverted epifluorescence microscope and analyzed by software Image J with OpenComet plugin [89]. Alkaline comet assays were performed five times in triplicate, counting >100 slides per experiment.
Colorectal carcinoma biopsies were re-analyzed specifically for EEPD1 expression, compared to adjacent normal mucosa. Gene expression measurements were performed in 217 patients with colorectal carcinomas from pre-therapeutic biopsies as we described [80]. From 217 patients, tumor samples were extracted, and from 181 of these matched normal tissue (mucosa) samples were also obtained. Gene expression was measured on Agilent Human Microarrays. Microarray data was extracted as log2 intensities and quartile normalized. Gene expression of EEPD1 (Agilent Probe: A_23_P333498, Refseq: NM_030636, Chr. Coord: chr7:36340858–36340917, Probe: CAGCCTGTTCTTACTCCAGCTCAACCCATTGGGTGTTGGCTGTTTTTGGTTTTAGTTGTT) was obtained. Significance was computed from matched tumor vs. mucosa samples using a paired Wilcoxon test.
|
10.1371/journal.pgen.1003548 | Comparative Polygenic Analysis of Maximal Ethanol Accumulation Capacity and Tolerance to High Ethanol Levels of Cell Proliferation in Yeast | The yeast Saccharomyces cerevisiae is able to accumulate ≥17% ethanol (v/v) by fermentation in the absence of cell proliferation. The genetic basis of this unique capacity is unknown. Up to now, all research has focused on tolerance of yeast cell proliferation to high ethanol levels. Comparison of maximal ethanol accumulation capacity and ethanol tolerance of cell proliferation in 68 yeast strains showed a poor correlation, but higher ethanol tolerance of cell proliferation clearly increased the likelihood of superior maximal ethanol accumulation capacity. We have applied pooled-segregant whole-genome sequence analysis to identify the polygenic basis of these two complex traits using segregants from a cross of a haploid derivative of the sake strain CBS1585 and the lab strain BY. From a total of 301 segregants, 22 superior segregants accumulating ≥17% ethanol in small-scale fermentations and 32 superior segregants growing in the presence of 18% ethanol, were separately pooled and sequenced. Plotting SNP variant frequency against chromosomal position revealed eleven and eight Quantitative Trait Loci (QTLs) for the two traits, respectively, and showed that the genetic basis of the two traits is partially different. Fine-mapping and Reciprocal Hemizygosity Analysis identified ADE1, URA3, and KIN3, encoding a protein kinase involved in DNA damage repair, as specific causative genes for maximal ethanol accumulation capacity. These genes, as well as the previously identified MKT1 gene, were not linked in this genetic background to tolerance of cell proliferation to high ethanol levels. The superior KIN3 allele contained two SNPs, which are absent in all yeast strains sequenced up to now. This work provides the first insight in the genetic basis of maximal ethanol accumulation capacity in yeast and reveals for the first time the importance of DNA damage repair in yeast ethanol tolerance.
| The yeast Saccharomyces cerevisiae is unique in being the most ethanol tolerant organism known. This property lies at the basis of its ecological competitiveness in sugar-rich ecological niches and its use for the production of alcoholic beverages and bioethanol, both of which involve accumulation of high levels of ethanol. Up to now, all research on yeast ethanol tolerance has focused on tolerance of cell proliferation to high ethanol levels. However, the most ecologically and industrially relevant aspect is the capacity of fermenting yeast cells to accumulate high ethanol levels in the absence of cell proliferation. Using QTL mapping by pooled-segregant whole-genome sequence analysis, we show that maximal ethanol accumulation capacity and tolerance of cell proliferation to high ethanol levels have a partially different genetic basis. We identified three specific genes responsible for high ethanol accumulation capacity, of which one gene encodes a protein kinase involved in DNA damage repair. Our work provides the first insight in the genetic basis of maximal ethanol accumulation capacity, shows that it involves different genetic elements compared to tolerance of cell proliferation to high ethanol levels, and reveals for the first time the importance of DNA damage repair in ethanol tolerance.
| The capacity to produce high levels of ethanol is a very rare characteristic in nature. It is most prominent in the yeast Saccharomyces cerevisiae, which is able to accumulate in the absence of cell proliferation, ethanol concentrations in the medium of more than 17%, a level that kills virtually all competing microorganisms. As a result this property allows this yeast to outcompete all other microorganisms in environments rich enough in sugar to sustain the production of such high ethanol levels [1], [2]. Very few other microorganisms, e.g. the yeast Dekkera bruxellensis, have independently evolved a similar but less pronounced ethanol tolerance compared to S. cerevisiae [3]. The capacity to accumulate high ethanol levels lies at the basis of the production of nearly all alcoholic beverages as well as bioethanol in industrial fermentations by the yeast S. cerevisiae. Originally, all alcoholic beverages were produced with spontaneous fermentations in which S. cerevisiae gradually increases in abundance, in parallel with the increase in the ethanol level, to finally dominate the fermentation at the end.
The genetic basis of yeast ethanol tolerance has attracted much attention but until recently nearly all research was performed with laboratory yeast strains, which display much lower ethanol tolerance than the natural and industrial yeast strains. This research has pointed to properties like membrane lipid composition, chaperone protein expression and trehalose content, as major requirements for ethanol tolerance of laboratory strains [2], 4 but the role played by these factors in other genetic backgrounds and in establishing tolerance to very high ethanol levels has remained unknown. We have recently performed polygenic analysis of the high ethanol tolerance of a Brazilian bioethanol production strain VR1. This revealed the involvement of several genes previously never connected to ethanol tolerance and did not identify genes affecting properties classically considered to be required for ethanol tolerance in lab strains [5].
A second shortcoming of most previous studies is the assessment of ethanol tolerance solely by measuring growth on nutrient plates in the presence of increasing ethanol levels [2], [4]. This is a convenient assay, which allows hundreds of strains or segregants to be phenotyped simultaneously with little work and manpower. However, the real physiological and ecological relevance of ethanol tolerance in S. cerevisiae is its capacity to accumulate by fermentation high ethanol levels in the absence of cell proliferation. This generally happens in an environment with a large excess of sugar compared to other essential nutrients. As a result, a large part of the ethanol in a typical, natural or industrial, yeast fermentation is produced with stationary phase cells in the absence of any cell proliferation. The ethanol tolerance of the yeast under such conditions determines its maximal ethanol accumulation capacity, a specific property of high ecological and industrial importance. In industrial fermentations, a higher maximal ethanol accumulation capacity allows a better attenuation of the residual sugar and therefore results in a higher yield. A higher final ethanol titer reduces the distillation costs and also lowers the liquid volumes in the factory, which has multiple beneficial effects on costs of heating, cooling, pumping and transport of liquid residue. It also lowers microbial contamination and the higher ethanol tolerance of the yeast generally also enhances the rate of fermentation especially in the later stages of the fermentation process. Maximal ethanol accumulation capacity can only be determined in individual yeast fermentations, which are much more laborious to perform than growth tests on plates. In static industrial fermentations, maintenance of the yeast in suspension is due to the strong CO2 bubbling and this can only be mimicked in lab scale with a sufficient amount of cells in a sufficiently large volume.
The advent of high-throughput methods for genome sequencing has created a breakthrough also in the field of quantitative or complex trait analysis in yeast [6], [7]. The new methodology has allowed efficient QTL mapping of several complex traits [5], [8], [9] and reciprocal hemizygosity analysis [10] has facilitated identification of the causative genes. The efficiency of the new methodologies calls for new challenges to be addressed, such as comparison of the genetic basis of related complex properties. In addition, complex trait analysis in yeast has been applied up to now mainly to phenotypic properties that are easy to score in hundreds or even thousands of segregants [5], [8]–[16]. However, many phenotypic traits with high ecological or industrial relevance require more elaborate experimental protocols for assessment and it is not fully clear yet whether the low numbers of segregants that can be scored in these cases are adequate for genetic mapping with pooled-segregant whole-genome sequence analysis.
The aim of this work was to compare the genetic basis of the complex traits of maximal ethanol accumulation capacity and tolerance of cell proliferation to high ethanol levels. We show that both traits have a partially different genetic basis and we have identified for the first time specific genes involved in maximal ethanol accumulation capacity.
We have evaluated 68 different yeast strains in small-scale fermentations for maximal ethanol accumulation capacity under very high gravity (VHG) conditions [17], using 33% (w/v) glucose. The robust wine strain V1116 was used as reference in each series of fermentation experiments. Figure 1A shows the number of strains able to accumulate a certain maximal ethanol level expressed as percentage of the ethanol level accumulated by V1116 in the same experiment, which was 18.4±0.4% (v/v). There was no correlation between the final glycerol and ethanol levels produced but there was an inverse correlation between the final glycerol level and the ethanol yield. Table 1 shows the fermentation results for a number of representative strains ranked according to the maximal ethanol level produced in comparison with the reference V1116.
The fermentation of the reference strain, V1116, took 9.4±1.1 days to complete. The ethanol productivity was 0.65 g.L−1.h−1 (or 0.83 g.L−1.h−1 when we omit the last two days where the fermentation had slowed down very much). The productivity was highest during the first three days (1.17 g.L−1.h−1). The yield was 0.446 g ethanol/g glucose (87.4%). There was 2.20±0.57% (w/v) glucose leftover. Glycerol production was 10.34±0.47 g/L. The final pH was 4.5±0.2 for all strains evaluated. The best ethanol producer was the sake strain, CBS1585, that accumulated 103.4% of the amount of ethanol accumulated by V1116. The relative ethanol production (% compared to V1116), the final ethanol % (v/v), the glycerol yield (g/L) and ethanol yield (% of maximum theoretical yield) for all 68 strains are listed in Table S1.
The laboratory strains BY4741 (Mata his3Δ1 leu2Δ0 ura3Δ0 met15Δ0) and S288c (prototrophic) produced only 64% and 80%, respectively, of the ethanol level accumulated by V1116. This is in accordance with previous studies that showed the prototrophic laboratory strain (S288c) to be generally more stress tolerant than its auxotrophic counterpart (BY4741) [18], although this has not yet been documented for ethanol tolerance. The eight beer strains tested all produced less than 80% of the ethanol produced by V1116, in agreement with the relatively low ethanol levels generally present in beers. On the other hand, strains used for the production of bioethanol and sake were among the best for maximal ethanol accumulation, which fits with the high level of ethanol produced in these industrial fermentations [19], [20].
Cell viability at the end of the fermentation was lower than 10%, and usually only 1–5%, for all strains tested, except for Ethanol Red and CBS1585. The bioethanol production strain Ethanol Red retained 22.1%±4.1% viable cells and the sake strain, CBS1585, even 31.5%±5.1%. The latter strain also showed the highest ethanol accumulation among all strains evaluated. High ethanol production is a well-known trait of sake strains [21]. The high residual viability is remarkable in view of the 18–19% of ethanol accumulated. The ethanol level could be enhanced further by applying continuous stirring (200 rpm) and raising the glucose concentration to 35%. In this case, ethanol levels between 20 and 20.5% (v/v) were routinely obtained, with an absolute maximum of 20.9% (v/v). In six consecutive fermentations with the same cells under these conditions, 20.5% ethanol was accumulated in the first fermentation and 16.5–19.5% ethanol (v/v) in the subsequent fermentations, demonstrating the persistent viability of strain CBS1585 under high ethanol conditions.
We have compared the maximal ethanol accumulation capacity with the ethanol tolerance of cell proliferation in the 68 strains. The results are summarized in Figure 1B and all original data are provided in Table S1. The results show that most strains with a low ethanol tolerance of cell proliferation also displayed poor maximal ethanol accumulation and that none of these strains reached a final ethanol titer of more than 18% (v/v). Strains with a higher ethanol tolerance of cell proliferation tended to produce higher maximal ethanol levels. This was most pronounced in the strains able to grow in the presence of 20% ethanol on plates. All of these strains showed high maximal ethanol accumulation and 50% produced a final ethanol level higher than 18% (v/v). On the other hand, the general correlation between the two traits showed only weak significance (Spearman one-tailed test: 90% confidence interval, P-value = 0.0984). This suggested that the genetic basis of the two traits was at least partially different.
The diploid sake strain CBS1585 was sporulated and stable mating type a and α segregants were obtained indicating heterothallism of the parent strain. Ten segregants were phenotyped in small-scale VHG semi-static fermentations. A segregant, Seg5 (MATa), was identified, which showed the same fermentation profile (Figure 2A) and maximal ethanol accumulation capacity as its parent strain, CBS1585 (Figure 2B). The laboratory strain BY710 (derived from BY4742; same genotype: Matα his3Δ1 leu2Δ0 ura3Δ0 lys2Δ0) showed a lower fermentation rate and also a much lower maximal ethanol accumulation capacity, which was only around 12% (v/v) (Figure 2A and 2B). The a mating type of the Seg5 strain was stable and FACS analysis confirmed that its DNA content was half that of its diploid parent CBS1585 (data not shown). We have crossed Seg5 with BY710 to obtain the diploid Seg5/BY710, which showed a similar high fermentation rate (Figure 2A) and high ethanol accumulation capacity (Figure 2B) as the original CBS1585 diploid strain. Growth assays on solid media, with or without glucose, and containing different levels of ethanol, showed that CBS1585, Seg5 and Seg5/BY710 had a similar ethanol tolerance of cell proliferation whereas the laboratory strain (BY710) was much more sensitive (Figure 2C). These results indicate that the two ethanol tolerance traits are dominant characteristics in the strain backgrounds used.
We have investigated whether ethanol tolerance as determined by the classical assays of cell proliferation on solid nutrient plates containing different levels of ethanol, correlates with maximal ethanol accumulation capacity in fermenting cells in the absence of cell proliferation. For that purpose, Seg5 was crossed with BY710, the Seg5/BY710 diploid sporulated and the segregants were first plated on solid media containing glucose and/or ethanol (18% to 20% v/v). Figure 3A shows a representative result. The haploid parent Seg5 showed high tolerance of cell proliferation to ethanol whereas the laboratory strain BY710 was much more ethanol sensitive. Among the segregants we could observe some with very high ethanol tolerance (e.g. Seg 11C), some with intermediate tolerance (e.g. Seg 10A) and others that were as ethanol sensitive as the laboratory strain (e.g. Seg11D). Out of 301 segregants evaluated in this way, 101 segregants showed moderate to high ethanol tolerance, whereas about half of the segregants (48.8%) could not grow at all on plates containing 18 or 20% ethanol (v/v). In the first category, 32 segregants showed an ethanol tolerance level as high as Seg5. Hence, about 1 in 9 segregants showed the same high ethanol tolerance as the superior parent. If we suppose random segregation of the loci and no epistasis, this ratio predicts three independent loci as being involved in determining the high ethanol tolerance of Seg5 compared to the laboratory strain BY710.
Subsequently, we tested 15 ethanol sensitive segregants (similar to Seg11D of Figure 3A) by fermentation in 250 mL of YP+33% (w/v) glucose. All 15 segregants clearly showed poor fermentation performance, with a low ethanol accumulation capacity (<14% v/v) (not shown). This suggests that there is a correlation between ethanol tolerance as measured by the cell proliferation assays on solid nutrient plates and maximal ethanol accumulation capacity in VHG fermentation, at least for the ethanol sensitive strains. Hence, to reduce the high workload required for phenotyping all segregants in fermentations, we tested in the small-scale fermentations only the 101 segregants that showed moderate to high ethanol tolerance in the growth assays on solid nutrient plates. We are aware that the strains with poor ethanol tolerance of cell proliferation may contain mutant genes that compromise maximal ethanol accumulation capacity or that when these strains show relatively high maximal ethanol accumulation capacity, they may contain (in part) different mutant alleles than the strains with high ethanol tolerance of cell proliferation. The main purpose of this work, however, was to identify the first set of major causative genes determining maximal ethanol accumulation capacity and this is the main reason why we continued first with the strains preselected for medium to high ethanol tolerance of growth.
The distribution of maximal ethanol accumulation capacity among the 101 segregants, as tested in semi-static small-scale fermentations in 250 mL of YP+33% (w/v) glucose, is shown in Figure 3B. We have also compared ethanol tolerance of cell proliferation and maximal ethanol accumulation capacity for the 101 segregants. The results are shown in Figure 3C. They are similar to the results obtained for the 68 natural and industrial yeast strains (Figure 1B) in two aspects. First, irrespective of the ethanol tolerance of cell proliferation, the segregants show a wide range of ethanol accumulation capacities. This confirms that the correlation between the two properties is weak. Second, the segregants with a higher ethanol tolerance of cell proliferation show a tendency towards higher ethanol accumulation capacity. The latter effect is less pronounced than with the selection of strains in Figure 1B, but this can be due to the fact that the poorest segregants for ethanol tolerance of cell proliferation have already been eliminated for the high-gravity fermentation experiments.
Only 22 segregants produced ethanol titres higher than 17% (v/v), similar to the ethanol production of Seg5 and Seg5/BY710. If we assume that all ethanol sensitive segregants, as determined by growth assays on solid nutrient plates, also display poor maximal ethanol accumulation, we have a ratio of one superior strain in ±14 segregants (301/22 = 13.7). Assuming random segregation of the QTLs and no epistasis, this ratio is consistent with four independent loci being responsible for the superior ethanol accumulation capacity of Seg5 compared to the BY710 control strain. We constructed several diploids by crossing the four best performing segregants but none of those showed higher ethanol accumulation capacity than the original CBS1585 diploid strain (data not shown).
We have performed genetic mapping of the two polygenic traits: on the one hand, high ethanol accumulation capacity in fermenting cells in the absence of cell proliferation, using the 22 best-performing segregants (pool 1) as determined in semi-static VHG fermentations, and on the other hand, tolerance of cell proliferation to high ethanol levels, using the 32 segregants (pool 2) that showed the best growth on solid nutrient media containing 18 to 20% (v/v) ethanol. The two pools had 12 segregants in common. Identification of the QTLs was performed by pooled-segregant whole genome sequence analysis [5], [6], [8], [9]. Genomic DNA was sent to two independent companies (GATC Biotech, Konstanz, and BGI, Hong Kong) for custom whole-genome sequence analysis with an average depth of ∼38 by the Illumina platform. Other sequencing parameters are summarized in the Methods section.
Sequence analysis of the genome of the superior parent Seg5 and comparison to S288c, allowed us to select 48,512 high-quality SNPs after filtering for sufficient coverage (≥20 times) and ratio (≥80%) [5], [22]. The coverage of at least 20 times was based on previous findings that a 20-fold sequencing coverage is sufficient to compensate for errors by the number of correct reads [23]. The ratio of at least 80% was chosen based on the plots of the SNPs between the two parent strains [5]. We also mapped the reads to the assembled sequence for the Kyokai n°7 strain available in the Saccharomyces genome database [24]. We were able to map about 20,000 additional reads to this sequence and 93% of the total read pairs aligned with proper distance and orientation to the Kyokai n°7 assembly, while only 87% of the read pairs mapped in the same way to S288c. We also identified the sake strain specific genes AWA1 and BIO6 [24], which further confirmed that CBS1585 belongs to the sake cluster of S. cerevisiae strains.
Genomic DNA was extracted from the two selected pools, containing 22 and 32 segregants, respectively, and also from an unselected pool, composed of 237 segregants (pool 3) in order to assess proper segregation of all chromosomes and possible links to inadvertently selected traits, such as sporulation capacity or spore viability. After sequence analysis, the SNP variant frequency was plotted against the chromosomal position (Figure 4). Upward deviations from the mean of 0.5 identify QTLs linked to the superior parent Seg5, while downward deviations identify QTLs linked to the inferior parent BY710. In most areas of the genome, and especially in the QTL areas, the independent sequence analysis by the two companies matched well, which confirms the robustness of the pooled-segregant whole-genome sequencing technology. Only in some selected areas the matching was poorer, which may be due to the low pool sizes. The SNP variant frequencies were smoothed using a Linear Mixed Model (LMM) framework [5] and the putative QTLs were identified by applying a Hidden Markov Model (HMM) similar to the one implemented in the FastPHASE package [25]. For each polymorphism, the HMM had three possible states: (i) a link with the superior parent (Seg5), (ii) a link with the inferior parent (BY710) and (iii) no link (background level). The SNP frequencies for each pool of segregants, analysed with the HMM, were assigned probability scores, that indicated to which state (Seg5, BY710 or background) they belonged and hence identified the QTLs, linked to either the superior parent (Seg5) or to the inferior parent (BY710).
The smoothed data of the SNP variant frequency and the Probability of linkage values obtained by HMM analysis with the selected pools 1 and 2 and the unselected pool 3, are shown in Figure 4. The QTLs identified with the HMM approach are listed in Tables 2 and 3 for pools 1 and 2, respectively. SNPs were considered significantly linked to the superior or inferior parent strain when the Probability of linkage was higher than 0.95 or lower than −0.95, respectively. The QTLs were numbered according to their position in the genome starting from chromosome I, independently of the trait (Tables 2 and 3).
The unselected pool 3 (237 segregants) showed ±50% SNP variant frequency in most of the genome and thus no evidence of any QTLs (Figure 4). The only exception was the right arm of chromosome V which was preferentially inherited from the BY parent strain. Comparison with the data of the selected pools, suggested some weak linkage with the genome of the BY parent strain in this part of chromosome V. Because of the weak linkage this was not retained for further analysis. Crosses of Seg5 with other BY strains did not show aberrant segregation of the right arm of chromosome V (results not shown). The results obtained with the unselected pool show that the QTLs identified for the two ethanol tolerance traits were not due to linkage with inadvertently selected traits, such as sporulation capacity or spore viability.
The QTLs identified with the selected pools 1 and 2 showed two common QTLs (on chr XIII and chr XV). They were called 12.1 and 17.1 for pool 1 and 12.2 and 17.2 for pool 2. It has to be emphasized that the ‘common’ character of these QTLs is only based on their common location in the genome. In principle, they could be located in the same place on a chromosome but caused by a different causative gene. Moreover, the QTLs 15 and 16 (pool 2) were also present in pool 1 as minor putative QTL of which the significance could not be demonstrated with the current number of segregants (Probability of linkage <0.95). Other minor putative QTLs of which the significance could not be demonstrated with the current number of segregants (Probability of linkage <0.95) were present in pool 1 and pool 2. They were also seen with the smoothed data and the HMM analysis (Figure 4) (e.g. on chromosome VII). There was no indication for linkage of the areas with the sake strain specific genes AWA1 and BIO6 to one or both of the ethanol tolerance traits.
We have analysed in detail two QTLs (2 and 3) involved in high ethanol accumulation capacity (pool 1) because this trait is more relevant in industrial fermentations and because these two QTLs were among those with the strongest linkage. QTL2 is located on chromosome I and was fine-mapped by scoring selected markers in the 22 individual segregants. This reduced the length of the QTL to the area between chromosomal positions 151 kb and 178 kb (P-value<0.05) (Figure 5A). The association percentage of the markers, their genomic positions, the respective P-values and the genes located in the putative QTL 1 are shown in Figure 5A.
Nearly all genes present in the centre of the QTL had at least on polymorphism either in the ORF, promotor or terminator. Hence, it was not possible to exclude on this basis a significant number of genes as candidate causative genes. Because of the large number of candidate genes and the high workload of the phenotyping for maximal ethanol accumulation capacity, we have introduced a modification of the Reciprocal Hemizygosity Analysis (RHA) methodology, which has been used previously for identification of causative genes [10]. Instead of testing one candidate gene at a time, we first evaluated a series of adjacent genes by ‘bulk RHA’. For that purpose a set of adjacent genes was deleted directly in the heterozygous diploid background (Seg5/BY710) so as to obtain the two reciprocally deleted hemizygous diploids of which the phenotype was subsequently compared. The first block of genes (bRHA 1.1) deleted, consisted of NUP60, ERP1, SWD1, RFA1 and SEN34. The two reciprocally deleted diploid strains were tested by fermentation in YP+33% (w/v) glucose, to address the effect of the Seg5 and BY710 alleles on ethanol accumulation capacity. The results showed no difference in the fermentation profile and maximal ethanol accumulation (Figure 5B), suggesting that none of these five genes were causative genes. There was also no difference in fermentation profile and maximal ethanol accumulation with the hybrid parent strain Seg5/BY710, further supporting that these genes did not influence these phenotypes.
The second block of genes tested consisted of YARCdelta3/4/5, YARCTy1-1, YAR009c, YAR010c, tA(UGC)A, BUD14, ADE1, KIN3 and CDC15 (bRHA 1.2) (Figure 5A). In this case there was a clear reduction of the fermentation rate and maximal ethanol accumulation when the alleles of the Seg5 strain were absent compared to absence of the BY710 alleles (Figure 5C). Glucose leftover correlated inversely with final ethanol titer. This suggested the presence of one or more causative genes in this region. Moreover, the fermentation rate was higher in the hemizygous strain where the BY710 alleles were absent compared to the hybrid parent strain Seg5/BY710, indicating that one or more of the BY710 alleles had a negative effect on this phenotype.
YARCdelta3/4/5, YARCTy1-1, YAR009c and YAR010c are transposable elements, while tA(UGC)A encodes one of the sixteen tRNAs for the amino acid alanine. BUD14 is involved in bud-site selection [26], ADE1 is involved in de novo purine biosynthesis [27], KIN3 encodes a non-essential serine/threonine protein kinase involved in a.o. DNA damage repair [28] and CDC15 encodes a protein kinase involved in control of the cell division cycle [29]. In order to identify the genes(s) involved in ethanol accumulation capacity, we investigated the most likely candidate genes individually with the classical one-gene RHA [10]. Involvement of the transposable elements appeared unlikely and was not evaluated by RHA. The other genes, BUD14, ADE1, KIN3 and CDC15, have polymorphisms (SNPs and/or indels) within their ORFs and/or promoter regions. RHA with the genes ADE1 and KIN3 showed that deletion of the Seg5 alleles resulted in strains with clearly lower ethanol accumulation capacity and higher glucose leftover compared to the strain with deletion of the respective BY allele, indicating that ADE1 and KIN3 are causative genes for high ethanol accumulation capacity in Seg5 (Figure 6A). For both genes, the hybrid parent strain Seg5/BY710 behaved in a similar way as the strain with the deleted BY710 allele. For CDC15 and BUD14 there was no difference in the performance of the two reciprocally deleted diploid strains (not shown). Deletion of ADE1 and KIN3 in the Seg5 and BY backgrounds caused a more pronounced effect in the Seg5 sake genetic background (Figure 6B).
The causative genes ADE1 and KIN3 were located in QTL2, which was not linked with ethanol tolerance of cell proliferation. When we tested the hybrid diploid strains previously used in RHA for maximal ethanol accumulation for determination of ethanol tolerance of cell proliferation, we could indeed not observe any significant difference between the two strains (Figure 6C). This confirms that these causative genes are specific for maximal ethanol accumulation capacity and that the genetic basis of the two ethanol tolerance traits is indeed partially different.
We also analysed in more detail QTL3, located on chromosome V. In the same chromosomal region, Swinnen et al. [5] previously identified URA3 as a causative gene in tolerance of cell proliferation to high ethanol levels of VR1, a Brazilian bioethanol production strain, in comparison with BY4741 as inferior parent strain. Since we crossed Seg5 with an ura3 auxotrophic laboratory strain (BY710), we first tested whether deletion of URA3 in Seg5 affected maximal ethanol accumulation in this genetic background. The fermentation profile and maximal ethanol accumulation of the strain Seg5-ura3Δ/BY710-ura3Δ (which is thus homozygous for ura3Δ) compared with the Seg5/BY710-ura3Δ diploid (which is heterozygous for ura3Δ) are shown in Figure 7A. Double deletion of URA3 resulted in a strain with a reduced ethanol fermentation rate, lower maximal ethanol accumulation and higher glucose leftover. We have also tested the effect of introducing URA3 in the ura3 auxotrophic strain BY4741, which accumulates only low amounts of ethanol under VHG conditions (±12% v/v). Introduction of URA3 enhanced the fermentation rate in the later stages of the fermentation and resulted in a clearly higher maximal ethanol titer and lower glucose leftover (Figure 7B). These results show that URA3 positively affects maximal ethanol accumulation capacity. The URA3 gene was located in QTL3, which was not significantly linked with ethanol tolerance of cell proliferation. When we tested the hybrid diploid strains previously used in RHA for maximal ethanol accumulation for determination of ethanol tolerance of cell proliferation, we observed slightly better growth for the strain with the URA3 allele from Seg5 (Figure 7C). This confirms that URA3 has only a minor contribution to this phenotype in this genetic background and suggests that the very weak upward deviation in the SNP variant frequency plot observed in this position for ethanol tolerance of cell proliferation might have been due to the URA3 gene.
Comparison of the sequence of ADE1 and KIN3 in Seg5 and BY710 (S288c background) revealed a C to T transition in the promoter of ADE1 and a C to T transition in the promoter of KIN3 as well as three synonymous transition mutations in the ORF of KIN3. We have checked the presence of these SNPs in the ADE1 and KIN3 genes of 36 yeast strains of which the whole genome sequence has been published. The results are shown in Table 4. (Among the 36 strains there were additional SNPs compared to S288c, which were not present in Seg5. These SNPs are not shown). The C to T change at position 169227 in ADE1 is present only in two other strains, Kyokai nr. 7 and UC5. Both strains are sake strains and these strains are known to have superior maximal ethanol accumulation capacity. Sake fermentation produces the highest ethanol level of all yeast fermentations for production of alcoholic beverages [21]. The SNPs in KIN3 of Seg5 at positions 170564 and 170945 are present in many other strains. Interestingly, however, the two other SNPs in KIN3 of Seg5, at positions 170852 (in the ORF) and 171947 (in the promoter) are not present in KIN3 of any one of the 36 sequenced strains and therefore may be rather unique.
Tolerance to high ethanol levels is an exquisite characteristic of the yeast Saccharomyces cerevisiae and no other microorganism has ever been reported to show higher ethanol tolerance. This unique property of yeast lies at the basis of the production of most alcoholic beverages and of ethanol as biofuel. In most studies, ethanol tolerance has been assayed by measuring cell proliferation in the presence of increasing ethanol levels. Although this assay is convenient for routine measurement and large-scale screenings, its true relevance for ethanol tolerance in yeast fermentation is unclear. Industrial yeast fermentations always start with an excess of fermentable sugar compared to other essential nutrients. As a result, the ethanol production rate in the second phase of the fermentation, the extent of attenuation of the residual sugar and the final ethanol titer reached are always achieved by stationary phase cells. In this work we have compared for the first time the genetic basis of maximal ethanol accumulation capacity in fermenting cells in the absence of cell proliferation with that of ethanol tolerance of cell proliferation. To avoid interference by the genetic background of the strain, we have used the same pool of segregants derived from one hybrid parent. The results of the QTL mapping by pooled-segregant whole-genome sequence analysis reveal a partial overlap between the genetic basis of the two traits. Although only two significant QTLs, 12.1/12.2 on Chr. XIII and 17.1/17.2 on Chr. XV appear identical, there were minor QTLs in pool 1 of which the significance could not be demonstrated with the current number of segregants (e.g. on Chr. VII and XV), which are likely overlapping with significant QTLs in the same position in pool 2. However, because of the lower number of segregants in pool 1, the P-value of these QTLs is not low enough for significance. It is also important in this respect to recall that the two pools had 12 segregants in common. A stronger argument for partial overlap between the genetic basis of the two traits could be made if two pools would be assembled not only with different segregants but containing in each pool only segregants that would not fit phenotypically in the other pool. This would have required, however, a large amount of additional experimental work.
Our work has shown that successful QTL mapping using pooled-segregant whole-genome sequence analysis can be performed with relatively low numbers of segregants. This is particularly important for elucidation of the genetic basis of complex traits of industrial importance, like maximal ethanol accumulation capacity, which require laborious experimental protocols for scoring. It has also shown that resorting to seemingly similar traits, like ethanol tolerance of cell proliferation, which can be scored easily with simple growth tests on plates, is not a valid alternative. On the other hand, there were several minor QTLs detected for the trait of maximal ethanol accumulation capacity, for which the significance could not be demonstrated with the number of segregants used. The ability to detect QTLs depends on the importance of the causative allele for establishing the trait and on the number of QTLs/causative alleles involved. Higher numbers of segregants will therefore always be useful to map minor QTLs and identify their causative alleles.
Detailed analysis of QTL 2 on Chr. I and QTL 3 on Chr. V identified three genes specifically linked to maximal ethanol accumulation capacity, which indicates that ethanol tolerance as relevant for maximal ethanol accumulation in fermentations cannot be fully assessed in a reliable way by simple growth tests on solid nutrient plates in the presence of ethanol. The identification of KIN3 as a causative gene is striking because it reveals for the first time a role for DNA damage repair in ethanol tolerance as required for maximal ethanol accumulation. Moreover, the superior KIN3 allele of Seg5 contained two SNPs, which were absent in the KIN3 gene of all yeast strains of which the genome has been fully sequenced up to now, suggesting that they may be important for the exceptional ethanol accumulation capacity of the Seg5 strain and its diploid parent CBS1585. KIN3 encodes a serine-threonine protein kinase, required for arrest at the G2/M-phase checkpoint in response to the DNA damage inducing agents MMS, cisplatin, doxorubicin and nitrogen mustard [28]. Involvement of Kin3 in the DNA damage response may be consistent with its requirement for tolerance to high ethanol levels. Ethanol was reported to be mutagenic and to induce single-strand DNA breaks in repair-deficient but not in repair-proficient yeast cells [30]. It was also reported to trigger chromatin condensation, fragmentation, and DNA cleavage in yeast, features suggestive of induction of apoptosis [31]. Mitochondrial DNA loss in yeast is induced by ethanol and mitochondrial DNA from more ethanol tolerant flor yeasts enhanced ethanol tolerance when transferred into a laboratory strain [32]. Also in mammalian cells, ethanol was shown to induce DNA damage and is a known carcinogen [33]. A role for DNA repair in protecting mammalian cells from ethanol-induced damage has been proposed [34]. It will be interesting to investigate to what extent maximal ethanol accumulation in yeast can be enhanced by further strengthening DNA damage repair capacity.
The case of URA3 is remarkable. It encodes one of the most active enzymes, oritidine 5-phosphate decarboxylase (OCDase), that catalyzes the decarboxylation of oritidine 5-phosphate (OMP) to uridylic acid (UMP) [35], [36]. This is the sixth enzymatic step in the de novo biosynthesis of pyrimidines. Yeast strains lacking URA3 need supplementation with uracil in the medium. Our previous work identified ura3 [5] and several other auxotrophic mutations (unpublished results) as causative mutations for ethanol tolerance of cell proliferation in a cross of a Brazilian bioethanol production strain VR1 and the BY laboratory strain. We have now identified ura3 as causative gene for maximal ethanol accumulation capacity in the cross of the sake strain CBS1585 and BY. However, in this genetic background ura3 was not significantly linked to ethanol tolerance of cell proliferation. This indicates that the genetic basis of the latter property is dependent on the genetic background of the strain. A stronger capacity to generate the electrochemical potential required for symport, may for instance offset the ethanol sensitivity of the uptake of auxotrophic supplements.
Lower expression of auxotrophic genes, like URA3, or lower activity of the gene product, forces the yeast cells to take up most uracil using the uracil permease, Fur4, which is an active proton symporter [37]. Stress conditions, including nutrient starvation, can trigger degradation of Fur4 [38]. Hence, the requirement of URA3 for maximal ethanol production capacity might be linked to nutrient starvation towards the end of the semi-anaerobic, high-gravity fermentation process, which can take up to 21 days. Uracil is likely depleted and/or its transporter Fur4 may be degraded because of the nutrient starvation conditions at the end of the fermentation. In addition, ethanol toxicity may also compromise the proton gradient, which is required for uptake by symport of uracil and protons from the medium. This type of inhibition was reported for amino acid uptake by the proton symporter Gap1 [39]. The reduction of maximal ethanol accumulation in ura3 auxotrophic strains suggests that in general the active uptake of nutrients may be compromised by the increasing ethanol level at the end of the fermentation. Yeast cells have only one permease to transport uracil, Fur4, which may make this system more sensitive to ethanol inhibition compared to for instance amino acid transport, for which many transporters exist.
Another relevant factor may be the general fitness problem of URA3 deleted strains. URA3 auxotrophic strains (BY710-ura3Δ, BY4741-ura3Δ and Seg5-ura3Δ/BY710-ura3Δ) showed much less biomass production in the pre-cultures performed in YPD, YP+5% glucose, YP+10% glucose and during the fermentations in YP+33–35% glucose (OD600 around 12.4±2.68) whereas Seg5/BY710-ura3Δ (prototrophic) for example, had much higher cell densities (32.6±3.42 in stirred fermentations). Low cell densities contribute to a slow fermentation phenotype that is also associated with lower final ethanol levels. The importance of uracil supplementation and fitness problems related to uracil auxotrophic strains have been reported recently by Basso et al. [19].
We identified the ADE1 allele in Seg5 by RHA as a superior allele for maximal ethanol accumulation capacity in high-gravity fermentation. As in the case of URA3, there was no link between ADE1 and tolerance of cell proliferation to high ethanol levels. ADE1 encodes a N-succinyl-5-aminoimidazole-4-carboxamide ribotide (SAICAR) synthetase, that is required for de novo purine biosynthesis [27]. ADE genes have not been connected previously to ethanol tolerance, but they have been linked to high sugar tolerance. In a genome-wide screen with the deletion strain collection, Ando et al. [40] identified three adenine biosynthetic genes (ADE5,7, ADE6 and ADE8) as being required for tolerance to 30% (w/v) sucrose. These genes were not required for tolerance to high sorbitol and NaCl, indicating a specific role in high sugar tolerance. The ADE genes are involved in biosynthesis of purine and derived metabolites, such as ATP. Measurements of the ATP level revealed a reduction with two-fold in the ade mutants, indicating that inability to synthesize sufficient ATP could be related to the high sucrose stress sensitivity. Alternatively, in the ade mutants the STRE-controlled stress response gene, HSP12, which encodes a plasma membrane chaperone protein, was not induced under high-sucrose stress, as opposed to sorbitol and salt stress [40]. This suggests a possible defect in induction of stress protection factors as cause for the high-sucrose sensitivity and once more a specific role of ADE genes in high sugar stress. Osmotic stress is known to trigger the HOG-pathway [41]. Phosphorylation of Hog1, the central component of the HOG pathway, however, was normal under all three osmotic stress conditions in all ade mutant strains, suggesting that deficiency of the HOG pathway, or at least the osmosensing systems, was not involved in the sensitivity of the ade mutants [40]. Because we measured maximal ethanol accumulation in fermentations with a very high sugar level (33%, w/v, glucose), the link with the superior allele of ADE1 in QTL2 (chr I) may be due to its importance for tolerance to high sugar stress. If this would be the reason why the superior ADE1 allele of Seg5 supports higher ethanol accumulation under VHG conditions, it would explain why the ADE1 gene was not linked to ethanol tolerance of cell proliferation as measured with pool 2, since the solid nutrient plates contain a low sugar level and a high ethanol level. The ADE1 gene from the superior parent Seg5 did not have any mutation in the ORF compared to the sequence in the laboratory strain BY. However, one SNP was located in the promoter region of the Seg5 allele (Chr I: 169.228 bp - C/T). The promoter of ADE1 is known to bear a hexanucleotide (5′ TGACTC 3′) element that is under amino acid control [27]. Although the mutation is not within that regulatory element, it is possible that it is affecting ADE1 expression and thereby also high sugar tolerance.
In conclusion, our work has shown that successful QTL mapping with pooled-segregant whole-genome sequence analysis can be performed for traits of industrial importance, which require elaborate experiments to score the phenotype, using a relatively low number of segregants. We have identified for the first time genes required for maximal ethanol accumulation capacity in the absence of cell proliferation in fermenting yeast cells and have shown that the genetic basis of this trait is partially different from that of tolerance of cell proliferation to high ethanol levels. The superior alleles identified can be used for improvement of maximal ethanol accumulation capacity in industrial yeast strains for bioethanol production and for the production of alcoholic beverages. This improves attenuation of the sugar at the end of the fermentation, which enhances yield in industrial bioethanol production and reduces residual sugar levels in alcoholic beverages. A higher final ethanol level in bioethanol production reduces distillation costs and lowers the liquid volumes in the plant, which in turn reduces costs associated with cooling, heating, pumping and transport of liquid residue.
The S. cerevisiae strains utilized in this study are listed in Table S2. Yeast cells were grown with orbital agitation (200 rpm) at 30°C in YPD medium containing 1% (w/v) yeast extract, 2% (w/v) Bacto peptone and 2% (w/v) glucose.
VHG fermentations were performed in which the glucose concentration was raised to such an extent (33% w/v) that a maximal final ethanol level (17–18%) was obtained with only minimal residual sugar left [17]. A further increase in glucose concentration above this level reduced the maximal ethanol level again. Cells were first pre-grown in 3 mL of YPD medium for 24 h (200 rpm, 30°C), after which 0.5 mL was transferred to 5 mL of YP+5% (w/v) glucose and the culture incubated for 24 h (200 rpm, 30°C). Cells of the last pre-culture were inoculated in 100 mL of YP+10% (w/v) glucose with initial OD600 of 1.0. The cells were grown for 2 days (200 rpm, 30°C) until stationary phase. 12.5×109 cells, based on cell counting, were harvested. The cells were centrifuged (3000 rpm, 5 min, 4°C), the pellet was resuspended in 3 mL of YP and inoculated into 250 mL of YP+33% (semi-static) or 35% (continuous stirring) (w/v) glucose. The fermentations were performed at 25°C. Agitation was performed with a magnetic rod (30×6 mm) at 120 rpm (semi-static, 4 h) or 200 rpm (continuous stirring). The fermentation was followed by weighing the tubes and from the weight loss the glucose leftover was calculated. Samples were taken at the end of the fermentation for HPLC analysis and cell viability determination. The metabolites quantified by HPLC were glucose, glycerol and acetic acid. The HPLC system utilized (Waters Breeze) consisted of an ion-exclusion column (WAT010290) at 75°C and detection was performed by refractive index (model 2414). The eluent used was H2SO4 (5 mM) at a flow rate of 1.0 mL/min. Samples of 10 µL were automatically injected and processed for 20 min. Ethanol was quantified by near infrared spectroscopy (Alcolyzer, Anton Paar). Cell viability was assessed by oxonol staining followed by flow cytometry analysis [42]. The ethanol yield (g of ethanol produced per g of glucose consumed) was calculated by dividing the ethanol produced with the glucose consumed (initial glucose concentration minus glucose leftover).
The cells were pre-grown in YPD for 2 days (200 rpm, 30°C). The OD600 was measured in triplicate and the cells were diluted to an initial OD600 of 0.5. Four serial dilutions were made (10−1, 10−2, 10−3 and 10−4). A volume of 4 µL was spotted on plates: YPD (control), YPD+16% (v/v) ethanol, YP+16% (v/v) ethanol, YPD+18% (v/v) ethanol, YP+18% (v/v) ethanol and YPD+20% (v/v) ethanol. The plates were incubated at 30°C for up to 11 days and growth was scored from the second day on. The ethanol levels indicated are initial ethanol levels. During the preparation and incubation of the plates some ethanol may evaporate. Therefore, sample and control strains were always put together on the same plates.
General procedures for sporulation and tetrad dissection were used [43].
A small amount of cells (1.5 mg) was incubated with 10 µL of NaOH (0.02N) for 1 h (RT). The determination of the mating type was done by PCR with the primers for the MAT locus and MATa and MATα (alpha) DNA [44]. The 3 primers were used together.
Preparation of the DNA pools from the segregants was done either by (1) individual genomic DNA extraction and pooling of the DNA in equimolar concentrations; (2) mixing of the cells, based on dry weight, prior to DNA extraction, or (3) mixing of the cells based on OD600, prior to DNA extraction. For all preparations, the genomic DNA was extracted according to Johnston [45]. At least 3 µg of DNA per pool was provided for whole-genome sequencing to both GATC Biotech GA (Konstanz, Germany) and Beijing Genomics Institute (BGI, Hong Kong, China). In both cases the sequencing was performed with the Illumina platform and gave for most of the genome, and especially in the QTL areas, very similar results. For both pools and at both companies the sequencing depth was ∼38 and the read length was 75 at GATC Biotech and 90 at BGI.
Assembly and mapping were done with DNAstar Lasergene software. Smoothing of the sequencing data was performed with a Linearized Mixed Model (LMM) framework [5], [22]. We implemented a Hidden Markov Model (HMM) to identify regions related with the phenotypes similar to the one implemented in the FastPHASE package [25]. For each variant, the HMM has three possible states: (i) relation with the superior parent, (ii) relation with the control parent and (iii) no relation (background). To capture the effect of recombination, the transition between two states of the same type is the probability of no recombination and the probability of the transition between two states of different type is the probability of recombination divided by two. We estimated the probability of recombination for each pair of neighbor variants using a negative exponential relation with the physical distance as in [25]. The emission of each state is the number of calls of the alternative allele which is an integer between zero and ni, where ni is the total number of allele calls for the variant i. We used beta-binomial distributions for all states to take into account the fact that given the finite number of segregants, the contribution of each parent to the pool is not exactly half. For the superior parent states we setup α = 10 and β = 1. For the control parent states we set α = 1 and β = 10. For the background states we estimated α and β using the alternative allele frequencies in all sites. We checked that for the background distribution α≈β>1, which makes the background distribution to be close to a binomial with probability 0.5 (as expected). We used the forward-backward algorithm to calculate the posterior probability of each state given the allele counts for each dataset. A manuscript with a complete explanation of the algorithm and comparisons with currently available methods is in preparation. The QTLs detected were further analyzed by scoring SNPs in the segregants individually using allele-specific primer sets, which were rigorously tested for reliability with the two variants of each SNP in the parent strains and all segregants. Statistically significant QTLs were confirmed by multiple testing using a false discovery rate (FDR) control [46].
Yeast cells were transformed with the LiAc/SS-DNA/PEG method [47]. Genomic DNA was extracted with PCI [phenol/chlroform/isoamyl-alcohol (25∶24∶1)] [48]. Polymerase chain reaction (PCR) was performed with Accuprime polymerase (Invitrogen) for sequencing purposes and ExTaq (Takara) for diagnostic purposes. Sanger sequencing was performed by the Genetic Service Facility of the VIB. The detection of SNPs by PCR was performed as previously described [5].
RHA was performed as described previously [5], [10] in the diploid Seg5/BY710 genetic background. In addition to single gene deletions we also performed large deletions (bulk RHA) of regions up to 27 kb long. The selection marker utilized was the amidase gene (AMD1), which was amplified from the vector pF6a-AMD1-MX6. The gene AMD1 was cloned from Z. rouxii [49]. The primers utilized in the AMD1 amplification had at least 80 extra bases that corresponded to the flanking regions of the area to be deleted. The transformants were selected on solid YCB + acetamide 10 mM (yeast carbon base 11.7 g/L; sodium phosphate buffer 0.03 M; agar 20 g/L). The correct integration of the constructs was checked by PCR, using one primer that annealed within AMD1 and two other primers that annealed either downstream or upstream of the deleted region. The PCR products were sequenced and the polymorphisms (SNPs and indels) present in the regions flanking the selection marker were identified when the Seg5 allele was replaced by AMD1. On the other hand, when the laboratory allele was deleted, no polymorphism was detected by Sanger sequencing. Double allele deletion was not observed during the bulk RHA because the deleted regions contained at least one essential gene.
The fermentations with different yeast strains were done with the reference strain V1116 as a control in duplicate. The most interesting strains were repeated at least once. The fermentations with different meiotic segregants were done with the reference strains Seg5, BY710 and Seg5/BY710. The segregants showing more than 16.5% (v/v) ethanol production were evaluated by fermentation at least once more. The fermentations for RHA were done in triplicate. The results were analyzed with a paired t-test (p<0.01, except for the comparison of V1116 and CBS1585 for which p<0.05 was used).
All sequence data have been deposited in the Sequence Read Archive (SRA) at the National Center for Biotechnology Information (NCBI) and can be accessed with account number SRA056812.
|
10.1371/journal.pcbi.1000466 | Investigating CTL Mediated Killing with a 3D Cellular Automaton | Cytotoxic T lymphocytes (CTLs) are important immune effectors against intra-cellular pathogens. These cells search for infected cells and kill them. Recently developed experimental methods in combination with mathematical models allow for the quantification of the efficacy of CTL killing in vivo and, hence, for the estimation of parameters that characterize the effect of CTL killing on the target cell populations. It is not known how these population-level parameters relate to single-cell properties. To address this question, we developed a three-dimensional cellular automaton model of the region of the spleen where CTL killing takes place. The cellular automaton model describes the movement of different cell populations and their interactions. Cell movement patterns in our cellular automaton model agree with observations from two-photon microscopy. We find that, despite the strong spatial nature of the kinetics in our cellular automaton model, the killing of target cells by CTLs can be described by a term which is linear in the target cell frequency and saturates with respect to the CTL levels. Further, we find that the parameters describing CTL killing on the population level are most strongly impacted by the time a CTL needs to kill a target cell. This suggests that the killing of target cells, rather than their localization, is the limiting step in CTL killing dynamics given reasonable frequencies of CTL. Our analysis identifies additional experimental directions which are of particular importance to interpret estimates of killing rates and could advance our quantitative understanding of CTL killing.
| The immune response mediated by cytotoxic T lymphocytes (CTLs), which kill infected cells, is thought to be essential to control viral infections. Experiments offer data which allow one to address the efficacy of this cell population in vivo and to estimate characterizing parameters. However, it is unclear which mathematical description reflects the experimental situation best and leads to reliable parameter estimates that quantify CTL efficacy. We simulate the spatial interaction of CTLs and infected cells in a 3-dimensional computer model to examine different mathematical descriptions of the experimental situation, independently of experimental data. Thereby we find an appropriate mathematical term to describe the killing process. Estimates obtained so far describe CTL efficacy on a population level. By varying the individual properties of simulated CTLs, such as the velocity, we find that the time a CTL needs to kill an infected cell is probably the key factor limiting CTL killing efficacy. Our analysis identifies additional experimental directions which could advance our quantitative understanding of CTL killing for different diseases.
| Cytotoxic T lymphocytes (CTL) are some of the most important cells of our immune system. They are particularly important against viral infections or tumours. They recognize infected cells by scanning their surfaces for peptide-MHC-I complexes which present peptide fragments sampled from the cytoplasm. These complexes can tell the CTL if the cell is infected or not. Once activated and primed for a specific peptide-MHC-I complex, CD8+ T cells differentiate into effector CTL, which are able to lyse infected cells. After an infection is cleared, some specific CTL may persist as memory cells.
Immunologists are interested in quantifying the efficacy of CTL in vivo. An appropriate measure of CTL efficacy would allow us to disentangle quantitative from qualitative aspects of the CTL response: For example, such a measure should tell us whether a memory CTL response is less efficacious than an effector CTL response because there are fewer cells, or because individual memory CTL do not perform as well as effector cells. A measure of CTL efficacy represents the first step in predicting if CTL responses will be able to control an infection, and in quantifying the selection pressure CTL responses exert on the pathogen population. This selection pressure may lead to immune escape where the virus evolves to become mainly undetected by the actual immune response [1],[2]. Rates which determine how fast CTL lyse infected cells are already estimated for HIV-I in vitro [3] and indirectly via the selective advantage of escape variants in vivo [4].
The best experimental data for the estimation of the CTL efficacy in vivo so far originate from the in vivo CTL killing assay [5],[6]. In this assay, cells are prepared to display LCMV-peptides on their MHC-I molecules. The cells are then transferred into mice which harbour CTL specific for these LCMV-peptides. It is known that the transferred cells migrate to the spleen where they are targeted by CTL. These cells are mostly located either in the red pulp or in the T cell-zones (perioarteriolar lymphoid sheaths (PALS)) depending on the stage of infection [7]. While effector CTL preferentially accumulate in the red pulp, memory CTL are mostly located in the PALS. Some time after the transfer, the levels of target cells are determined in the spleen. To estimate CTL efficacy, Regoes et al. [8] and Yates et al. [9] proposed a mathematical model that takes into account the migration of target cells into the spleen, and their subsequent killing by CTL. Fitting this model to in vivo CTL killing data, we obtained a killing rate constant , and proposed this constant as a measure for CTL efficacy. We found differences between killing rate constants of effector and memory CTL, as well as for immunodominant and -subdominant epitopes (see Table 2 in [9]).
In these previous studies, we intended to compare the efficacy of distinct CTL populations whose levels differ. Therefore we assumed a mass-action killing term to disentangle quantitative from qualitative aspects of CTL killing. However, the validity of the mass-action assumption is uncertain. Furthermore, it is unclear how the killing rate constant in our mathematical model, which describes CTL killing on the level of the cell populations involved, is related to properties of individual CTL. For example, how does CTL velocity or the time needed to kill a cell influence the estimate of the killing rate constant?
To address these questions we simulate the dynamic inside the spleen or PALS, respectively with a three-dimensional cellular automaton (CA) [10],[11]. A CA is an individual-based computer simulation of a dynamical system on a lattice (in our case a three dimensional one). This method allows us to identify a more appropriate mathematical description of CTL killing than the simple mass-action term. Additionally, by generating in vivo CTL killing data for different scenarios, we are able to relate the properties of individual cells to the population dynamics of the system.
We find that there is a parameter regime in which the behaviour of our CA model of CTL killing in the spleen is consistent with data obtained by two-photon microscopy [12]–[15]. Further, we find that the most appropriate mathematical description of CTL killing is linear in the target cell levels, and a saturating function of the CTL levels. However, fitting a mathematical model with such a saturating killing term does not improve the fit to the original in vivo CTL killing data consistently.
Studying the influence of single cell properties on our killing rate estimates we find that one specific experimental detail, which concerns the fate of CTL-target cell conjugates after splenectomy, is of particular importance to be able to interpret the population-level killing rate constants in terms of single cell efficacy. Nevertheless, given the CTL frequencies observed experimentally, the killing rate constant is mainly determined by the time a CTL needs to kill its target, and not the CTL's velocity.
The spleen is the secondary lymphoid organ which surveys the blood for foreign antigen. It consists of red pulp, which is a site of red blood cell destruction and comprises roughly 80% of the splenic volume, interspersed with lymphoid regions (white pulp). While most of the blood will bypass the lymphoid regions and remain in the direct circulation, around ten percent of the cells will diffuse through the T cell-zone (PALS) [16]. During this passage the cells are under constant surveillance by T lymphocytes.
We simulate the population dynamics of the cells in the PALS as a cellular automaton. A cellular automaton allows us to investigate the impact of individual cell properties and spatial aspects on the dynamics. Into our simulation model, we incorporate target cells, target-cell-specific CTL, splenocytes, and a limited number of large cells which correspond to dendritic cells or macrophages. In addition, we include the reticular network (RN), which defines the anatomical structure of the spleen, as well as some free space (see Fig. 1). For a detailed description of the automaton see Materials and Methods.
We first simulate the specific CTL without target cell interaction to characterize their behaviour with regard to experimental observations. The simulated CTL perform a random walk (see Fig. 2B, Video S1 in the Supporting Material) consistent with observations made in lymph nodes and the spleen based on in vivo imaging techniques [17]–[21].
We are able to manipulate the motility of our simulated CTL through the rules of movement. We adjusted these rules of movement such that they display a mean velocity, velocity fluctuations, and a motility coefficient largly consistent with observations in vivo.
Miller et al. [14],[21] measure an average velocity for T cells of about in lymph nodes. It is thought that the velocity is in the range of [22]. For the spleen, T cell velocities are observed which are slightly slower, even correcting for differences in the observation method [20], but comparable to those found in lymph nodes [19]. We mainly use a parametrization where the simulated CTL migrate with an average velocity of .
In Fig. 2A we show the velocity fluctuations of six simulated CTL for this parametrization chosen at random for a time period of 100 min. The amplitude of the velocity fluctuation of the simulated CTL agrees with experimental observations (e.g. [12]). However, the velocity fluctuations are less rugged than those observed in experiments (compare to Fig. 5 in [12]). This is due to the discreteness of space in our cellular automaton, i.e. cells can not be arbitrarily displaced but have to occupy a node in the lattice.
As stated earlier, the simulated CTL perform a random walk. This can be seen from a projection of their normalized tracks (Fig. 2B) as well as from the relation between their mean displacement and the square root of time (Fig. 2C). The discreteness of space also affects the random walk characteristics of simulated CTL. As CTL have to “move” on given edges, there is only a discrete number of turning angles available. As cell movement involves the restructuring of the actin-filament network in the cytoskeleton [23], cells will prefer small turning angles. Therefore, in the simulation, they are programmed to preferentially choose per move. In the absence of killing and given a mean velocity of , the simulated CTL show a mean turning angle of where the turning angle was measured every minute. This slightly increases if we include killing activity (). The distribution of (see Fig. S1) differs from those observed experimentally [13]. This is provoked by the fact that the motility of simulated CTL is only affected by environmental conditions. CTL do not change their moving direction as frequently which leads to a low mean turning angle.
A second variable to characterize cell movement is the motility coefficient, . Given standard parametrization, the motility coefficient of simulated CTL is approximately (see Fig. 2C, blue line), which is slightly above the range of observed experimentally for T cell movement [12]–[15],[22]. For other parametrizations, which we consider in this paper, the motility coefficient is in the range .
The mean velocity as well as the motility of CTL decreases in the presence of CTL-target cell interaction (Fig. 3) which is also observed experimentally in the spleen [19]. For CTL-target cell interaction, target cells will appear in the cellular automaton with a certain rate and are killed after encountered by CTL (see Material and Methods for details). In our simulations, the mean CTL velocity decreases exponentially in the killing duration with a rate constant of approximately −0.012 min−1, given a fixed CTL concentration. In Fig. 3B we show the change in the mean displacement per square root of time for different values of in comparison to the mean displacement of simulated CTL in the absence of killing. In all cases, CTL velocity was fixed to . Given a killing duration of , the motility coefficient decreases from to . The motility coefficient includes only 15% of the value measured in the absence of killing, if we assume a killing duration of . The decrease of is linear in .
In previous studies [8],[9], it is assumed that the rate at which target cells are killed depends linearly on the frequency of the CTL, , and the frequency of the targets, , in the spleen. Such a dependence is commonly referred to as mass-action hypothesis. However, the mass-action assumption may be inaccurate if the system is not well-mixed and the dynamics is spatially confined. In addition, the fact that CTL cannot seek for target cells while bound in a conjugate may lead to deviations from a mass-action killing term.
To address the question whether the mass-action hypothesis appropriately describes the killing dynamics given spatial confinements, we initialized the cellular automaton with different combinations of and the starting target cell frequency . The CTL frequency ranges from 0–20% of simulated cells, which covers the frequencies observed for dominant and subdominant effector and memory responses (Tab. 1). The average velocity of CTL, , was fixed and the killing duration was defined by , in agreement to experimental observations [24]. The loss of target cell frequency, , was calculated by(1)Fig. 4A shows linearity of in for different levels of . In contrast, is not linear in (Fig. 4B), but saturates for high levels of .
As an improvement over the mass-action killing term, we therefore propose the following relationship between and :(2)Hereby, denotes the maximum killing rate at high levels of , and denotes the CTL frequency at which the killing rate is at half of the maximum. Such terms have been suggested previously [25]–[27]. The saturation in the CTL frequency was observed independent of the density of the reticular network (varying the volume occupied by reticular network from 0–50% of the simulated space, data not shown). Fitting Eq. (2) to the data generated in our simulations yield and . The estimate of is slightly above the frequency of a subdominant effector or memory CTL response (Tab. 1, see or , respectively for GP276). This estimate of suggests that — if the parametrization of our cellular automaton agrees with the situation in vivo — the saturating term Eq. (2) should be preferred over a mass-action killing term for in vivo killing data with CTL frequencies .
The result that a killing term which saturates in the CTL frequency is more appropriate stays valid if we include multiple time points at which we calculate the loss in target cell frequency . By this, we additionally include the search process and relate to the fact that CTL which are bound in conjugates are prevented from hunting other target cells.
To account for the non-linear relationship between the loss of target cells and CTL frequencies, we substituted the mass-action killing term in our basic model (see Materials and Methods) by one that saturates in . In this case our equation is extended as follows:(3)Eq. (3) was fitted to the in vivo CTL killing data from Barber et al. [5] as described previously [8].
We obtain a reduction in the residual sum of squares compared to the previous model with a mass-action killing term, Eq. (8). This reduction is significant (as assessed by an ) for the effector and memory CTL response against the NP396-LCMV epitope given a significance-level of (see Tab. 1). However, even though a saturating killing term does not significantly improve the fit of our killing model to all the data, our simulations strongly suggest that a saturating term is more appropriate to describe the killing dynamics.
We therefore use a saturating term in the following analyses. Estimates for and as well as for the are given in Table 1.
The saturating term and the term respectively, mathematically describe the reduction in the target cell population due to their interaction with the CTL population. How does this population-level description of the killing dynamics relate to properties at the individual-cell level, such as the velocity of CTL or the time it takes to kill a target cell?
We find that one specific experimental detail is of particular importance for the interpretation of population-level parameters in terms of the single cell properties. This experimental detail concerns the fate of target cells in conjugates after the splenectomy. It is unknown if conjugates are simply broken up by the preparation of the spleen for the cell sorter, or if, during this preparation, killing of target cells that are bound to CTL continues. In any case, conjugates are not observed when the splenocytes are analysed by fluorescent-activated cell sorting (FACS).
We perform simulations with varying CTL velocities and killing durations . The CTL velocity is indirectly manipulated via the rules of cell movement in our cellular automaton.
We generate data with the same structure as Barber et al. [5] (see Materials and Methods). We choose a frequency of target-specific CTL with which is orientated at the CTL frequencies observed ([5] and Tab. 1). While the population of target-specific CTL is kept constant according to the assumptions made in the analysis [8],[9], target cells as well as control cells appear in the cellular automaton with a certain rate out of a restricted pool of cells (see Fig. 5 and Video S2 for a time course of one simulation). Each simulation represents the dynamics of CTL killing in a single mouse followed over 300 minutes. We perform 36 simulations for each combination of CTL velocity and killing duration. A bootstrap analysis with 1000 replicates is performed by sampling the number of free and bound target cells, control cells and CTL in 6 randomly chosen “mice” per indicated time-point (at either 15, 30, 60, 90, 120 or 240 minutes) per replicate.
We consider two measures of CTL velocity: (i) , the mean “hunting velocity” based only on CTL that are not bound to target cells, and (ii) , the mean velocity of all CTL regardless of whether they are bound to target cells or not.
In Fig. 6A, we plot the killing parameter versus the hunting velocity . Hereby, we assume that target cells which are bound to CTL remain alive and are counted by the cell sorter. Estimates of Spearman's rank correlation coefficient () show a weak correlation between and (see Tab. 2). For the mean velocity and , we observe a correlation coefficient of (see Fig. 6B). The finding of a weak influence of the CTL velocity on is further corroborated if we analyze the lifespan, , of a single target after it appears in the cellular automaton. For a killing duration of , the average lifespan , and 80% of all target cells are killed 20 min after their appearance regardless of the CTL velocity. Similar observations are made for the other levels of . Most target cells are recognized immediately after their appearance in the cellular automaton.
Assuming that all target cells bound to CTL are killed before the FACS analysis leads to slightly different conclusions regarding the influence of CTL velocity on the parameter estimates that describe the killing dynamics. We find a stronger correlation between and (), and () (see Fig. 6D–E).
The correlation between the CTL velocities and are analogous but inverse to those observed in .
Quantifying CTL efficacy based on data of an in vivo CTL killing assay requires models which capture the anatomical complexities inside the mice and reflect experimental conditions. Previously, we proposed a population-level model [8],[9], in which the killing dynamics was seperated into migration to and killing in the spleen. However, uncertainties remained concerning the most appropriate mathematical description of the killing term and how to interpret the population-level killing parameter. To address these questions, we constructed a 3-dimensional cellular automaton model of a T cell zone in the spleen.
Unlike previous studies [12],[30], we do not model the biophysical processes involved in cell movement. Rather, we impose simple rules of directed movement and position swapping. We find that our cellular automaton model can recapitulate experimentally observed CTL motility. This has recently been established in a more general context by Bogle and Dunbar [31].
We find that the most appropriate description of the killing term is linear in the frequency of target cells and saturates in the frequency of CTL. This saturation is observed irrespective of the various densities of the reticular network we considered (0%–50%). This is both expected and surprising: Expected, because it is well established that killing terms will saturate in CTL frequency [25],[32] as CTL, while bound to a target cell, are prevented from hunting and killing others. Surprising, because our cellular automaton is spatially structured, and we expected this spatial structure, which was not considered in previous work [25],[32], to percolate into the most appropriate killing term. Additionally, the conditions, under which the saturating term is theoretically derived, namely that there conjugates are in a quasi steady state, are not fulfilled in our simulations.
The main aim of our study was to identify the most appropriate killing term, rather than to obtain new estimates for killing. Nevertheless, we re-analyzed the data by Barber et al. [5] using a model with a saturating killing term. Using a saturating killing term improves the fit significantly for the immunodominant NP396-epitopes in the effector and memory response. This suggests that the CTL levels specific for NP396 are in the saturating regime. The improvement of the fit in the case of the effector NP396 response is consistent with the critical CTL frequency which we derived from our simulations. (Above this critical CTL frequency the killing rate is saturated.) The improvement of the fit in the case of the memory NP396 response is also consistent with the critical CTL frequency if we factor in the finding that memory CTL are mostly located in and around the T cell zones [7]. That means that, in the case of memory responses, it is more appropriate to compare the critical CTL frequency with the proportion of epitope-specific CTL in the pool of CD8+ T cells ( in Tab. 1), rather than the entire spleen ( in Tab. 1).
The killing term allows us to estimate populational-level parameters which quantify CTL efficacy. By varying the velocity of CTL and the time a CTL needs to kill a target cell, we were able to determine the influence of these single-cell properties on the population-level killing parameters. We based our analysis on a CTL frequency of which is in the range of the frequencies observed experimentally ([5] and Tab. 1). We find that the population-level parameters are mostly affected by the killing duration . The longer , the lower the killing rate constant or . The impact of the CTL velocity on the killing rate constant or varies depending on our assumptions regarding the fate of target cells in conjugates after the splenectomy. The impact of CTL velocity is weak if we assume that target cells in conjugates are still alive and counted by the cell sorter. If we assume that killing in conjugates continues and, therefore, target cells bound in conjugates are not detected by the cell sorter, the impact of CTL velocity is stronger. To clearly separate the relative effects of killing duration and CTL velocity it is necessary to determine the fate of target cells in conjugates during the in vivo killing assay.
Yates et al. [9] showed a significant difference between estimates of the killing rate constant in effector and memory responses. Effector CTL are more efficacious than memory CTL indicated by a higher value of . Our analysis suggests that the difference between effector and memory CTL can be explained by a difference in the time which a single CTL needs to kill a target cell. This hypothesis is in line with the observation that memory CTL store intermediate or low level of perforin and granzyme in comparison to effector CTL, which could prolong the killing process [33]. The difference between the killing rate constants for NP396- and GP276-specific CTL could be explained by different binding rates between the T cell receptor of the CTL and the peptide-MHC complex of the target cells. It is known that the binding of the T cell receptor specific for NP396 to NP396-MHC is stronger than that of the T cell receptor specific for GP276 to GP276-MHC [34]. We find that lower probabilities of recognition, which correspond to low binding rates, lead to lower estimates of in our simulations (see Fig. S4). To test these hypotheses about the killing process experimentally, one could combine in vivo CTL killing assays with two-photon microscopy as it is performed for the analysis of T cell activation [12]–[15],[22],[24].
In a recent study, Ganusov and De Boer [35] calculated target cell half-lives using a mathematical model that did not control for differences in CTL levels. The reason that these authors neglected CTL levels was that the exact form of the killing term is unknown. However, to predict the protection afforded by CD8+ T cell responses it is necessary to extrapolate the efficacy of a CTL population of varying size. Further, to decide if effector CTL are more efficacious than memory CTL, or CTL in acute infections are more efficacious than CTL in persistent infections, it is necessary to disentangle quantitative from qualitative aspects. Therefore, the dependence of the killing rate on the level of CTL can, in the long run, not be ignored.
The point of the present study was to derive a more appropriate mathematical description of the killing term from a model that incorporates more of the spatial complexities of the spleen as previous population level descriptions. Based on our analysis, we showed that a killing term which saturates in the CTL frequency would be more appropriate to describe the experimental situation.
There is no clear answer to the question which of the different killing rates for LCMV epitopes presented in this study and published so far [8],[9] should be preferred. Much more detail about the killing process is required to clearly favour one estimate. Our study is a first step to improve the estimation of per-capita killing rates based on a population-level and to enhance their interpretation in terms of single cell properties.
We use a three-dimensional lattice of nodes and edges to simulate the T cell zone of the spleen. Recent analysis showed the suitability of a lattice based approach to simulate T cell movement [31],[36]. We define periodic boundary conditions in which a cell leaving the simulated space on the one side of the lattice will reappear at the opposite side. Each node of the lattice represents a cell or a part of a cell. We consider target cells, target-specific CTL and splenocytes, which occupy a single node. Macrophages and dendritic cells are larger than CTL and have an average diameter of [12],[37]. These cells are modelled as occupying four nodes connected in no regular shape. The shapes are not stable, we only require that each cell-part has at least one other part of the cell as its direct neighbour. Some nodes are occupied by reticular network which does not change position over time and represents spatial obstacles to moving cells. Lastly, a few nodes are left unoccupied and define free space.
Each node has 26 neighbours. As cell movement requires a complex restructuring of the actin cytoskeleton [23], each cell in our cellular automaton is assumed to have a preferred moving direction . The direction can change upon encounter of another CTL, a target cell, or reticular network (see below). A cell will only have a moving direction of while it is bound in a conjugate. The cellular automaton was implemented in the C++ programming language.
We consider a lattice of 30×30×30 nodes, which makes 27000 nodes in total. To set the spatial scale of the simulation we assume that each edge of the lattice has a size of , the average diameter of a T cell [38],[39]. The reticular network occupies ≈4500 nodes (≈17% of the space). To achieve the actual structure of a network, we seed the lattice at random nodes from which the network grows until the assigned volume is occupied. As the spleen is a densely packed organ, free space is set to ≈1300 nodes (≈5%). We consider only a small number of large cells (macrophages and dendritic cells) (≈24–40 nodes, <0.1% of space). To determine the appropriateness of the mass-action term, target cells and target-specific CTL are randomly positioned in the lattice according to their assigned frequency. The rest of the lattice is filled with unspecified splenocytes.
For the analysis of the influence of the CTL velocity and killing duration on our estimates, we simulate the migration of target cells into the spleen, in addition to target cell killing in the spleen, in accordance with the events in an in vivo CTL killing assay [5]. In these simulations, 450 CTL (≈2% of the total number of cells without reticular network) are randomly positioned in the lattice. Target and control cells ( each) appear in the lattice at a rate [8]. They can either appear on a free node or replace an unspecified splenocyte, which, in turn, is simply deleted.
At each time-step of the simulation the position and other properties of each cell are updated. A time-step corresponds to 30 seconds of real time. In the simulations in which we increase or decrease the velocity of CTL, we update them more or less often than the other cells, respectively. In these simulations, a time-step corresponds to 12–40 s real time. We initialize our simulations by a burn-in phase of 20 minutes real time before target cells are allowed to migrate into the spleen, and target cells and CTL are allowed to interact. After the burn-in phase a simulation is run for 300 min real time.
The 27000 nodes of the cellular automaton comprise approximately 21000 (biological) cells. As the total number of splenocytes of a mouse spleen is estimated to be around 2×107–108 cells [5],[40], the modelled compartment comprises roughly 0.01%–0.1% of the white pulp of the spleen.
Each cell is able to move. We distinguish between two types of movement. The first is movement into free space. A cell can move into a neighbouring unoccupied node if it has the appropriate moving direction . If several cells are able to move into the free spot, one cell is chosen at random. The second type of movement is defined as neighbour swapping. As we are not interested in knowing if an unspecified splenocyte changes its place with another unspecified splenocyte (and to speed up computation), neighbour swapping is performed by target cells and target-specific CTL only. Hereby, such a cell will swap its place with a splenocyte irrespectively of the moving direction of the splenocyte while two CTL or target cells only swap their places if they move towards each other.
Movement of cells consisting of several nodes involves the restructuring of their shape. Such a cell is simulated to “diffuse” into an unoccupied node by placing its cell-part ( = node) farthest from the unoccupied node into this node. The node which was occupied by the moved cell-part becomes free space. The same procedure is performed for neighbour-swapping with a CTL or target cell. Position changes become effective after all cells updated their position.
If a CTL or target cell is not able to move, it randomly chooses a new moving direction . This new direction becomes effective in the next round of updating. The new moving direction is sampled from the set of the 26 possible directions defined by the next neighbours according to the following method. The former moving direction is translated into cartesian coordinates , with . One coordinate is chosen at random and updated dependent on the former value. As cells prefer small changes in their direction, with probability and otherwise if (analogous for ). If , at random. If the cells hit the reticular network, there is a higher chance to make larger turns in our simulations, as there is a high chance that a node in a direction similar to the previous moving direction will also be occupied by reticular network (). By controling the number of changes and moves per time step, we are able to regulate the velocity of cells. With these rules, CTL will perform random walks as described above (see Fig. 2B).
Before each update of the lattice, all the target-specific CTL scan their direct neighbourhood for target cells. If a CTL encounters a target cell it recognizes it with a certain probability (the probability of recognition). Upon recognition both cells will form a conjugate. Unless it is stated otherwise, we assume the probability of recognition to be one. It is observed, that conjugate-formation is followed by a period where T cells and bound target cells migrate together before they finally stop [24],[41]. However, it is not clear how the direction of the conjugate is determined and what happens if several CTL are bound to one target or vice versa. We assume that conjugates will immediately stop migrating after conjugate formation and stay immobile during the time of the killing, . Allowing conjugates to migrate together for a certain time does not generally affect our results.
We allow multiple killing of CTL which is in agreement with observations in vitro where CTL kill multiple targets simultaneously [28]. When the target cell is killed, the CTL chooses a new moving direction at random and proceeds.
The average velocity of a CTL in a simulation with CTL is defined by , with as the average velocity of CTL over time. We distinguish between two different types of velocities in the presence of killing. The “hunting” velocity is calculated based on all CTL that are not bound to target cells. The second velocity, , describes the average velocity over all CTL regardless of them being bound to target cells or not.
The motility coefficient measures the temporal displacement of a cell. If denotes the position of a cell at time and its displacement during this time, the motility coefficient, , in three dimensions is estimated according to: . For a graphical representation, we plot the mean displacement against the square root of , which denotes the time interval on which the calculation of the displacement is based. The motility coefficient can then be calculated from the slope of the curve (Fig. 2C) (see e.g. [18]).
Our research was motivated by the in vivo CTL killing assay presented in Barber et al. [5]. The experimental details are comprehensively described in this paper. Briefly, mice are infected by LCMV to generate CTL responses. Eight days after infection the mice harbour effector CTL, whereas 30 days after infection or later the mice harbour memory CTL. A mixture of fluorescently labelled cells is then injected intravenously into the tail vene of the mice. This mixture consists of equal proportions of target cells expressing either of the two LCMV epitopes (NP396 and GP276) and control cells, which do not express LCMV peptides and are therefore assumed to be unaffected by the CTL response. The frequencies of CTL, target and control cells are measured in the spleen after sacrificing the mice at different time points up to 270 min after the transfer of the target cells.
According to previously published mathematical models [8],[9], the data obtained by an in vivo CTL killing assay are analysed in two steps. First, we consider the migration of target cells into the spleen after injection. Second, we analyse the killing of target cells in the spleen. The model assumes that killing only occurs in the spleen and that the frequency of target-specific CTL, , is constant during the short time period of the experiment.
Estimates of migration parameters are based on absolute numbers of control cells in the blood, , and in the spleen, . The dynamics is described by:(4)(5)This leads to(6)where refers to the number of control cells transferred at the start of the experiment. The parameter defines the migration rate of cells into the spleen and the natural loss rate of cells in the blood. To estimate the actual killing rate , we assume that target and control cells migrate into the spleen following the same rate .
If denotes the frequency of target cells in the spleen then the basic model is given by:(7)The solution of the above differential equation is:(8)Here represents the total number of splenocytes.
To fit experimental data to the model, Regoes et al. [8] used Eq. (6) and Eq. (8). Assuming that most of the experimental error arises from different number of cells injected into the mice, this method can be refined [9]. In Yates et al. [9] we used the proportion of target cells that have been killed, , to estimate the killing rate constant . The proportion of target cells that have been killed, , is given by:(9)Here .
We showed in Yates et al. [9] that Eq. (9) provides a less biased estimator based on simulated data if there are large variations in the number of injected cells, . As we control in our simulations, both methods lead to the same results for given a mass-action assumption in the killing term (see Fig. 7 and Fig. S3). However, the latter method seems to be less robust if we assume a killing term, which is linear in the target cell frequency and saturates in the CTL frequency. This is surprising as we expect to reduce variation in the estimates with the revised method [9]. We do not understand the lower robustness of the fitting method yet. However, we mainly show results for using Eq. (8) because the estimates are more robust for our simulated data.
To perform the statistical analysis we used the language of statistical computing [42].
|
10.1371/journal.pntd.0004902 | Post-exposure Treatment with Anti-rabies VHH and Vaccine Significantly Improves Protection of Mice from Lethal Rabies Infection | Post-exposure prophylaxis (PEP) against rabies infection consists of a combination of passive immunisation with plasma-derived human or equine immune globulins and active immunisation with vaccine delivered shortly after exposure. Since anti-rabies immune globulins are expensive and scarce, there is a need for cheaper alternatives that can be produced more consistently. Previously, we generated potent virus-neutralising VHH, also called Nanobodies, against the rabies glycoprotein that are effectively preventing lethal disease in an in vivo mouse model. The VHH domain is the smallest antigen-binding functional fragment of camelid heavy chain-only antibodies that can be manufactured in microbial expression systems. In the current study we evaluated the efficacy of half-life extended anti-rabies VHH in combination with vaccine for PEP in an intranasal rabies infection model in mice. The PEP combination therapy of systemic anti-rabies VHH and intramuscular vaccine significantly delayed the onset of disease compared to treatment with anti-rabies VHH alone, prolonged median survival time (35 versus 14 days) and decreased mortality (60% versus 19% survival rate), when treated 24 hours after rabies virus challenge. Vaccine alone was unable to rescue mice from lethal disease. As reported also for immune globulins, some interference of anti-rabies VHH with the antigenicity of the vaccine was observed, but this did not impede the synergistic effect. Post exposure treatment with vaccine and human anti-rabies immune globulins was unable to protect mice from lethal challenge. Anti-rabies VHH and vaccine act synergistically to protect mice after rabies virus exposure, which further validates the possible use of anti-rabies VHH for rabies PEP.
| Rabies is an infectious disease causing 59,000 deaths and millions are exposed each year worldwide. Post-exposure prophylaxis (PEP) against rabies consists of a combination of passive (immune globulins) and active immunisation (vaccine) directly after viral exposure. Currently used plasma-derived anti-rabies immune globulins are expensive and scarce, urging the development of alternatives. Nanobodies or VHH are the smallest antigen-binding fragments of camelid heavy chain antibodies and are easy to produce with intrinsic good thermal stability and solubility. Combined treatment with anti-rabies VHH and vaccine gave significantly better protection than either compound alone in an intranasal rabies challenge model in mice, which validates the potential use of anti-rabies VHH as replacement of immune globulins in PEP.
| Rabies virus ultimately causes an aggressive and lethal infection in the brain of humans and other mammals. Rabies virus is a model neurotropic RNA virus that belongs to the family Rhabdoviridae, Genus Lyssavirus [1;2]. The virus is transmitted through the saliva of an infected animal by biting or scratching. Once the virus enters peripheral nerves or neurons, it quickly replicates in the neuronal cytoplasm and progeny virus is transported through the neuronal network by crossing tight interneuronal synapses, eventually giving rise to encephalitis [3;4]. Each year, an estimated 59000 people die from rabies and about 29 million receive post-exposure prophylaxis (PEP) after close contact with a suspected animal [5].
Passive antibody therapy with anti-rabies immune globulins (RIG) plays a major role in rabies post-exposure prophylaxis after high risk exposure [6]. Together with thorough wound cleansing, it is the first line of defence against the virus, and prophylaxis without RIG is associated with treatment failure [7;8]. Pioneering studies on the effects of anti-rabies serum date back to the late 1800s and early 1900s, and since 1954 the World Health Organisation (WHO) recommends the use of RIG in combination with vaccination for rabies post-exposure prophylaxis [9]. Treatment with RIG and vaccine should be initiated as soon as possible after potential infection, with additional vaccine administrations in the following weeks to activate a full-blown and lasting immune response. Passive immunization with RIG serves to immediately neutralize the virus and close the gap between viral exposure and the vaccine-induced immune response [7]. In this regime, initial protection is offered by RIG, which is then gradually replaced by vaccine-induced antibodies mounted between day 0 and 7–14, providing continued protection to patients [10].
Rabies antibodies can be either from equine (ERIG) or human (HRIG) origin. Due to adverse effects, such as serum sickness, equine antibodies are now used under the form of pepsin-digested Fab fragments, but if available, HRIG is still preferred over ERIG [9]. The production of HRIG, however, requires sufficient numbers of immune donors and gives rise to the typical problems associated with biological products of human origin, such as the transmission of infectious agents [9]. The worldwide shortage and the high costs makes these products poorly available to developing countries, where rabies is endemic [7;9], the reason why the WHO recommends to develop alternatives [11].
VHH or Nanobodies (a trade-name by Ablynx) are the smallest functional fragments (15 kDa) of heavy chain-only antibodies naturally occurring in Camelidae, and represent the antigen-binding variable domain. By nature VHH are hydrophilic and do not require hydrophobic interactions with a light chain, which allows high solubility, physicochemical stability and high-yield production in Escherichia coli, yeast or mammalian expression systems. The single domain nature and the small size of VHH also allows for easy formatting by genetic fusion into multimeric and multispecific constructs [12–14].
Previously, we generated a potent neutralizing anti-rabies VHH recognising two epitopes on the rabies glycoprotein, fused to an anti-albumin VHH to extend its serum half-life (HLE). The Rab-E8/H7-ALB11 was able to neutralize the virus at picomolar doses [15]. Post exposure treatment with anti-rabies VHH at 24 hours after intranasal virus challenge could significantly delay disease onset in mice, and depending on the dose, could rescue part of the mice from lethal disease [15].
The main aim of this study was to examine whether the combined treatment with anti-rabies VHH and vaccine (Rabipur, Novartis) after exposure to rabies virus has added value compared to single treatment with either compound in the intranasal rabies virus challenge model, which is very well suited to study intervention strategies for prevention and prophylaxis [16]. Via the intranasal route the virus can directly access the brain via the olfactory epithelium, which results in a highly reproducible infection [17]. First disease signs appear at 7 days, which rapidly progress the following 2 days, requiring euthanasia at 8–9 days post inoculation (DPI). This model was recently also proposed as a valuable alternative to intracranial inoculation for rabies vaccine potency testing [18]. The typically short incubation period of this model (6.07 ± 0.59 days) is ideal to study the potentially beneficial effect of the combined passive (VHH) and active (vaccine) immunisation on disease outcome. Our results show that anti-rabies VHH and vaccine act synergistically to protect mice after rabies virus exposure, which further validates the possible use of anti-rabies VHH for rabies PEP.
VHH directed against the rabies virus glycoprotein G were described previously [19]. Briefly, llamas were vaccinated using the inactivated rabies Human Diploid Cell Vaccine (HDCV, Sanofi, France) and RNA was extracted from peripheral blood lymphocytes. VHH genes were amplified from a cDNA library. Anti-rabies VHH were selected by panning phage libraries on plates coated with the native G protein. Multivalent VHH constructs were generated by the fusion of monovalent VHH into multimeric VHH constructs using flexible glycine-serine (GS) linkers [20]. In this study, we used the half-life extended VHH (HLE Rab-E8/H7-ALB11), containing two different VHH against the rabies virus spike protein and an anti-albumin VHH (ALB11) for half-life extension, and the non-HLE Rab-E8/H7 [15]. VHH was produced and kindly provided by Ablynx (Zwijnaarde, Belgium).
Human rabies immune globulins (HRIG) (Berirab, CSL Behring GmbH, Germany) are gammaglobulins purified from plasma of vaccinated human donors.
Rabipur (Purified Chicken Embryo Cell Vaccine, Novartis, Belgium) was reconstituted according to the manufacturer’s instructions and was administered via intraperitoneal or intramuscular injection. The vaccine contains at least 2.5 antigenic units (AU)/ml. It contains the inactivated Flury LEP strain produced on purified chick embryo cells.
Challenge Virus Standard (CVS)-11 is a virulent classical rabies virus obtained from the American Type Culture Collection (ATCC reference VR959) and was grown in baby hamster kidney (BHK)-21 cells. For virus inoculation in mice, a dose of 102.5 50% cell culture infectious doses (CCID50) was used.
Six-to-eight weeks old female Swiss outbred mice (Charles River, France) were used. Mice were kept in filter top cages, water and feed provided ad libitum and exposed to a natural day/night light cycle. Intranasal (IN) inoculation procedures are described in detail by Rosseels et al. [16]. The intranasal inoculation of rabies virus is an excellent technique to study antiviral treatment in the brain, since it leaves the brain mechanically intact, in contrast to intracranial inoculation, and yields a highly reproducible brain infection and disease outcome with little variation in the median survival time. This inoculation route has been used before for the evaluation of post exposure prophylaxis of rabies in mice [21]. For intraperitoneal (IP) or intramuscular (IM) injections maximum volumes of respectively 1000 and 100 μl were respected (50 μl per site in case of IM injections). Prior to intramuscular or intranasal administrations, mice were briefly anesthetized using isoflurane gas (IsoFlo, Abbott laboratories Ltd., United Kingdom), as described by Rosseels et al. (2011) [16].
Three retro-orbital bleedings were performed under isoflurane anaesthesia during the 28 day immunization period.
Mice were observed daily for signs of disease until 35 days post virus inoculation. Mice develop a typical disease pattern, which progresses as follows: isolation from the group (score 1), slow/less vivid movement (score 2), paresis in paws (score 3), uncoordinated movement (score 4), absence of spontaneous movement (score 5), no response to stimuli (score 6) and the end-stage characterized by mice burying their heads in cage bedding and slow breathing (score 7). The score per mouse ranges thus from 0 (no disease) to 7 (severe nervous disease). Disease progression was represented by plotting the daily score in function of the days post inoculation (DPI). The incubation period was defined as the period between virus inoculation and the first appearance of disease signs. In our experience, mice with a disease score of 6 or more die within 24 hours. Therefore, mice were euthanized by cervical dislocation when they reached a score of ≥ 6. Results were expressed with Kaplan-Meier survival curves. Rabies virus infection in the brain was confirmed using real-time reverse transcriptase polymerase chain reaction (RT-qPCR) as described by Suin et al. [22], and by the fluorescent antigen test (FAT), performed according to the Manual of Diagnostic Tests and Vaccines for Terrestrial Animals (Office International des Epizooties, 2008).
The viral RNA load in the brain of mice was determined using RT-qPCR, as previously described [15;22]. Briefly, the brain was homogenized and RNA was extracted according to the manufacturer’s instructions (RNeasy kit, Qiagen, Hilden, Germany). Ribosomal 18S was used as a reference gene for standardization and delta cycle thresholds (Δ Cq) values were calculated using the following formula: Δ Cq = Cqref−Cq, with Cqref equal to 45, the number of cycles in this program.
The virus-neutralizing titer of serum, antibody and VHH preparations was determined with the Rapid Fluorescent Focus Inhibition Test (RFFIT), according to the Manual of Diagnostic Tests and Vaccine for Terrestrial Animals (Office International des Epizooties, 2008). The neutralizing potency is expressed in international units (IU)/ml in reference to "The Second International Standard for Anti-Rabies Immunoglobulin", purchased from the United Kingdom National Institute for Biological Standards and Control.
GraphPad Prism was used for statistical analyses of in vivo data. Differences in survival times were tested using the Log-Rank test with a Bonferroni post-test, differences in Δ Cq values were tested using a Student’s t-test after normalization to the house-keeping gene. Differences in antibody titers were also tested using a Student’s t-test.
All experimental procedures were approved by the Ethical Commission of the WIV-ISP and CODA-CERVA (advice number 070515–05) and were performed according to the EU Directive 2010/63/EU for animal experiments.
To validate the protective effect of rabies vaccine (Rabipur, Novartis) in the intranasal rabies mouse model, mice were vaccinated with two intramuscular vaccine doses (0.25 AU/mouse), with a 3-day interval, following the schedule also used later on for PEP. This vaccination schedule is schematically represented in Table 1. Mice received a viral challenge 25 days after the last vaccine, allowing sufficient time for the development of an immune response.
The mounting of the humoral immune response in the blood after vaccination was monitored by assessing the rabies neutralization activity in vitro (RFFIT) in blood collected at different time points. In Fig 1, it is shown that mice that received the vaccine had detectable antibody titers eight days after the first dose (day -20), (mean 7.22 ± 3.28 IU/ml, range 3.73–12.62 IU/ml), which were well above the generally accepted protective threshold of 0.5 IU/ml. Antibody titers continued to increase until 28 days later (day 0, mean 11.47 ± 4.77 IU/ml, range 6.01–14.81 IU/ml).
To verify if the efficacy of the vaccine would be affected by the simultaneous administration of anti-rabies VHH, an interference phenomenon which is well known for anti-rabies immune globulins [23], a group of mice received besides the vaccine also a single dose of anti-rabies VHH in the same pre-exposure setting. In this regime, the first vaccination (day-28) was accompanied by anti-rabies VHH (Rab-E8/H7-ALB11) at a dose of 1.5 mg/mouse (corresponding to 60 mg/kg, 392600 IU/kg), at the moment of the first vaccination (day -28). Vaccine (IM) and VHH (IP) were administered at separate sites. As reference groups mice were treated with anti-rabies VHH alone, or left untreated. In mice, the half-life of the anti-albumin VHH is approximately 1.5 days, hence the anti-rabies VHH will be removed from the circulation at the moment of viral challenge.
Fig 1 shows that the rabies neutralization titers of mice that were injected with anti-rabies VHH, whether or not in combination with vaccination, were high 3 days after VHH administration (day -25, mean 88.28 ± 58.05 IU/ml, range 0.61–149.18 IU/ml). As expected, anti-rabies VHH titers rapidly declined over time with the clearance of the VHH from the blood (day -20, mean 9.43 ± 6.04 IU/ml, range 0.16–15.57 IU/ml) and no detectable titers (< 0.5 IU/ml) on day 0. Mice that received both vaccine and anti-rabies VHH had a mean titer of 5.69 ± 3.03 IU/ml (range 1.73–9.37 IU/ml) at day -20, similar to mice that received vaccine alone, while at day 0, antibody titers were significantly (p<0.005) lower in the vaccine + VHH group (mean 5.15 ± 3.38 IU/ml, range 0.37–10.03 IU/ml), compared to the vaccine only group.
Mice were challenged by intranasal virus inoculation 4 weeks after the start of the vaccination (day 0). Fig 2 shows the survival curves of the vaccinated and control mice. Despite the fact that all mice had high neutralizing antibody titers at the time of challenge, only 50% was protected from disease and survived the challenge. In the remaining mice disease progression was delayed (median survival time 27 days versus 9 days in control group). Disease signs in vaccinated mice were different compared to control mice, which typically develop signs of depression, such as unresponsiveness to stimuli and isolation from the group (S1 Video). The vaccinated animals remained responsive to stimuli and aware of the environment, while developing ascending paresis, starting at the hind limbs, that gradually evolved into paralysis. Eventually, mice had to be euthanized because of severe paresis and paralysis (S2 Video).
The survival of mice that received the combination regime 4 weeks before viral challenge was substantially reduced compared to the mice that received only the vaccine (11% versus 50%). The median survival time of these mice was not significantly different from the control group (10 days versus 9.5 days), despite the presence of relatively high neutralizing antibody titers at the moment of challenge. As expected, survival rates of mice that received anti-rabies VHH were comparable to the control group. The presence of the anti-rabies VHH in the circulation hence seems to reduce the vaccine efficacy. This may indicate that in absence of the virus, the binding of the anti-rabies VHH to the vaccine may interfere with the induction of an effective humoral immune response.
In previous in vivo studies, post-exposure treatment with the anti-rabies VHH one day after virus challenge was shown to provide protection from disease and death in a dose-dependent manner [15]. The same set-up was used to examine the efficacy of the combination of vaccine with a single anti-rabies VHH dose after exposure to the virus, which is the main indication for the use of vaccine in humans. Two different experiments were conducted. In a first experiment mice were treated with IP administered anti-rabies VHH (Rab-E8/H7-ALB11, 1,5 mg = 7852 IU/mouse) and IM administered vaccine (0.25 AU/mouse), twenty-four hours after challenge with a lethal rabies dose. A second vaccine dose was administered 3 days after the first. This treatment was than compared to treatment with anti-rabies VHH at the same dose or the vaccine regimen alone. The anti-rabies VHH dose was the lowest effective dose in post-exposure treatment in previous studies [15]. Similar to the pre-exposure set-up, vaccinated mice received a second vaccination 3 days after the first dose.
In the second experiment, the same vaccination schedule was applied, but instead of anti-rabies VHH, mice were treated with human rabies immune globulins (HRIG, Berirab, IP, 1 ml/mouse = 121.50 IU/mouse) at 24h after virus challenge. This is the highest volume and dose of the commercial HRIG product which could be administered to mice. Control mice were treated with HRIG alone. A schematic overview of both experiments can be found in Table 2.
The survival curves of the different treatment groups in the post-exposure prophylaxis setting are depicted in Figs 3 and 4. In the post-exposure setting, the combination of vaccination with anti-rabies VHH rescued 60% of mice (Fig 3), significantly better than the treatment with anti-rabies VHH alone which rescued only 19% of mice. The vaccine by itself in the post-exposure setting did not provide any protection, and disease was similar to the control group. The median survival time was significantly longer after the combined treatment (>35 days), compared to treatment with anti-rabies VHH (14 days, p<0.01) only, vaccine only (7 days, p<0.001) or the control group (8 days, p<0.001). Mice that were treated with the combination of vaccine and HRIG did not survive challenge, similar to mice treated with HRIG alone. The median survival time of mice treated with vaccine and HRIG was 9 days and treatment with HRIG alone resulted in a median survival time of 10 days.
The viral RNA load in the brain of mice was also assessed (Fig 5). Mice that received the PEP with vaccine and anti-rabies VHH had significantly lower viral RNA loads than control mice or mice treated with anti-rabies VHH only (Fig 5).
Together these data show that in the post-exposure setting anti-rabies VHH acts synergistically with a standard vaccination regime to protect mice from disease after virus exposure.
Post exposure prophylaxis (PEP) for rabies consists of a combination of passive (human or equine immune globulins) and active immunisation (vaccine) soon after exposure. Anti-rabies immune globulins are expensive, scarce and often not available or affordable for people in developing countries, that are typically most at risk [24;25]. Also in Western countries, RIG are increasingly difficult to procure [26]. Cheaper and easier-to-produce alternatives are needed. Previously, we developed anti-rabies VHH (Nanobody) capable of neutralizing virus at picomolar doses in vitro [15]. We also showed that post-exposure treatment with anti-rabies VHH only is capable of prolonging the incubation period of the disease in a dose dependent manner. In the current study, we evaluated whether post exposure treatment with the combination of anti-rabies VHH (half-life extended Rab E8/H7-ALB11) and vaccine (Rabipur, Novartis) is better than single treatment with anti-rabies VHH or vaccine only. The combined treatment was tested using an intranasal challenge model of mice. Treatment was initiated at 24 hours after challenge.
In humans, rabies can have incubation periods as short as 4–6 days, especially if the virus is deposited in highly innervated facial tissues, as is often the case in children [27]. Failure of classic PEP is described for several cases, often with short incubation periods or when highly innervated tissues were infected, which allows quick entry of the virus in nerves [28–30]. In order for PEP to be effective, it is believed that the virus needs to be intercepted by passive or active immune effectors before invasion of the central nervous system [28]. In case of a short incubation period, with rapid invasion of the nervous system, PEP cannot intercept the virus in time to prevent brain infection.
Compared to anti-rabies VHH or vaccine alone, the combination therapy in a post-exposure setting significantly delayed the onset of disease, prolonged median survival time and decreased mortality. Sixty per cent of mice treated with anti-rabies VHH and vaccine survived the infection, in contrast to 0% with vaccine only and 19% with anti-rabies VHH only. This is in agreement with the observations from Servat et al., who also showed that PEP with vaccine only was unable to prevent lethal disease [31]. Post-exposure treatment with anti-rabies VHH only proved more effective than vaccine only. This partial protection is in line with studies previously described by our group [15]. We assume that the synergy between vaccine and VHH lies in the fact that anti-rabies VHH can immediately delay the spread of the virus and prolong the incubation period, which allows more (sufficient) time for the active immune response to mount and control the infection in part of the mice. Indeed, treatment with VHH prolongs the incubation period from six to ten days, and the earliest antibody and cellular immune response can be expected as soon as seven days after intramuscular vaccination with an inactivated rabies vaccine [32]. This hypothesis also explains the limited efficacy of the combined treatment with vaccine and HRIG. Indeed, in the current and a previous study [15], we found that administration of HRIG to mice after lethal challenge merely prolongs the median survival time by one or two days. This limited prolongation of the incubation period is probably not long enough to mount an effective immune response, able to control the virus infection before it becomes lethal. Our results indicate that an active antibody response was induced in all survivor mice, corresponding to low residual levels of viral RNA (ΔCq ≤10) in the brain at the endpoint measurement (35 DPI).
Pre-exposure treatment with vaccine (IM) and VHH (IP) seemed to partially reduce the immunogenicity of the vaccine, a phenomenon that is also described for the combination of RIG and vaccine [23;33;34]. Mice which received anti-rabies VHH in conjunction with vaccine developed significantly lower antibody titers 4 weeks later and were significantly less well protected against virus challenge. Indeed, whereas mice receiving vaccine only had a 50% survival rate and a delayed disease progression, only 11% of the mice treated with vaccine and anti-rabies VHH survived infection and no delay could be observed. These results were confirmed in independent experiments in which a pre-incubated mix of rabies virus and VHH was administered simultaneously at the same site (S1 Fig). Antibodies can interfere with active immunization via different mechanisms. Most of the described mechanisms are Fc dependent, like inhibition of the B-cell responses by binding to the Fc-receptor, cross-linking of the B-cell receptor and the complement system, or antigen removal by macrophages [35]. Only humoral, and not cellular, immune responses seem to be affected by the presence of specific antibodies [36]. Since the used anti-rabies VHH is not a full antibody and lacks the Fc domain, it is unlikely that these mechanisms are involved [37]. The half-life extended anti-rabies VHH can interact with the neonatal Fc receptor through the intermediate of albumin, but it remains an unlikely mechanism since the non-HLE anti-rabies VHH, lacking an albumin-binding VHH component showed similar reduction of the vaccine efficacy (S1 Fig). Therefore a likely mechanism could be epitope masking. By binding to the surface glycoproteins of the inactivated vaccine virus, the anti-rabies VHH might shield recognition of the epitopes by the immune system [36]. The fact that the combination of anti-rabies VHH with vaccine still proved superior in PEP, argues for the relative importance of immediate passive immunisation in PEP, especially when the virus has easy access to nerves or neuronal cells.
Pre-exposure vaccination offered only partial protection upon intranasal virus challenge (50% survivors). Half of the mice that were actively immunized with (inactivated) vaccine, both at 28 and 25 days before challenge, still developed lethal brain infection. This incomplete protection, even with high antigenic doses (2 x 0.25 AU/mouse), is also described by other researchers, using similar models [18]. Nevertheless, the applied vaccine schedule resulted in clear seroconversion of all mice, with virus-neutralizing serum titers well above the protective threshold of 0.5 IU/ml (range 6.01–18.04 IU/ml) at the moment of challenge. Moreover, the challenge occurred at four weeks after the first vaccine administration, at the moment when the peak serological response can be expected [38;39]. The height of the neutralizing antibody titer in vaccinated mice did not correspond to the level of protection upon challenge. Some mice with titers up to 20 IU/ml still developed lethal disease.
The incomplete protection in the post-exposure setting may be explained by the aggressive nature of the used intranasal challenge model, in which virus is inoculated directly on a site that contains a high concentration of olfactory neuronal cells, providing a direct portal of entry to the central nervous system. In earlier studies we found spread of the virus in the olfactory bulbs of the brain already at the first day after inoculation [15]. Once inside the central nervous system, the virus is protected from several systemic immune effectors, which may limit the protection by the vaccine [40;41]. We therefore assume that the mice that survived the challenge after preventive vaccination or PEP with anti-rabies VHH and vaccine were able to develop a cellular immune response, capable of controlling the infection in the brain.
The intranasal challenge model is our preferred experimental model because of the high reproducibility, practicability, safety and animal wellbeing issues [16]. It may be that in an infection model with a longer incubation period and a more pronounced phase of peripheral virus replication in non-neuronal cells, preventive vaccination would be more effective, since vaccine-induced antibodies might be more effective to intercept virus spread between non-neuronal and neuronal cells. In our hands, intramuscular inoculation of rabies virus requires unnaturally high levels of virus in the inoculum (>105−6 CCID50) and yields variable inter-assay results, limiting its use for experimental comparison of intervention strategies [16].
Another remarkable finding was the different clinical picture observed depending on the vaccination status of the mouse prior to virus challenge. Naïve mice typically showed signs of depression, such as isolation from the group, inactivity and unresponsiveness to stimulation. In contrast, pre-immunised mice remained alert and vivid, but developed ascending paresis, resulting in paralysis of all limbs, requiring euthanasia. Vaccinated mice developed disease after a longer incubation period (13.7 instead of 9 days) and had a longer morbidity period (3 instead of 1.5 days), which resulted in a longer median survival time (27 instead of 9 days), compared to naïve mice. They also had lower viral loads in the brain at the peak of disease. The vaccine-induced immune response thus had a clear effect on pathogenesis and symptomatology. Iwasaki et al. also found that the host immune response has a clear impact on the development of, what they refer to as, either “encephalitic” or “paralytic” disease in mice. Rabies virus challenge in immunocompetent mice resulted in “paralytic disease”, with relatively low viral loads and a high extent of inflammation and damage in the brain. The same challenge in cyclophosphamide-treated mice resulted in the absence of an immune response and “encephalitic disease”, with severe general depression, only minor paralysis, high viral loads, and less neuronal cell damage [42]. In our study, the pre-immunized mice developed a disease pattern similar to the immunocompetent mice of Iwasaki et al., whereas the naïve mice evolved comparably to the cyclophosphamide-treated mice. In human cases, the average survival time of paralytic rabies is twice as long, compared to the encephalitic (furious) form [42]. Patients with paralytic rabies typically remain fully conscious, while developing ascending motor weakness [43]. Also in dogs, paralytic rabies is associated with reduced viral load and more prominent inflammation [44]. Our observations further add to the evidence that paralytic rabies may be caused by an immuno(patho)logical response of the host to the virus infection.
In humans, passive immunisation with anti-rabies antibodies is expected to bridge the immunity gap between virus exposure and onset of the active antibody production induced by the vaccine. The half-life extension of the anti-rabies VHH is based on the addition an anti-albumin VHH component. In mice, addition of anti-albumin VHH extends the half-life to 0.5–1.9 days [15], while in humans it is extended up to 10–20 days [45]. It would therefore be feasible to formulate and dose anti-rabies VHH for humans to obtain protective levels (> 0.50 IU/ml) in the blood for 14 days, which would be sufficient for the active immune response to take over. Compared to (human) rabies immune globulins (150 IU/ml), VHH can be produced and formulated at very high potencies (>6000 IU/ml). WHO recommends that rabies immune globulins are administered locally into the wound, however, due to the limited potency per ml of the rabies immune globulins, this is not possible for small wounds or injuries to nose, fingers or toes as it can cause compartment syndrome. VHH formulations containing high potencies per ml could overcome this problem and would be more suited for infiltration of the whole dose into small body parts.
These results provide evidence for the possible use of anti-rabies VHH together with vaccine for post exposure prophylaxis of rabies. Early treatment with anti-rabies VHH can delay the incubation period of the disease, which allows more time for the vaccine-induced immunity to control the infection. The ease of production and high thermal stability of VHH are important advantages over the currently used anti-rabies immune globulins.
|
10.1371/journal.pbio.0060288 | The Making of a Compound Inflorescence in Tomato and Related Nightshades | Variation in the branching of plant inflorescences determines flower number and, consequently, reproductive success and crop yield. Nightshade (Solanaceae) species are models for a widespread, yet poorly understood, program of eudicot growth, where short side branches are initiated upon floral termination. This “sympodial” program produces the few-flowered tomato inflorescence, but the classical mutants compound inflorescence (s) and anantha (an) are highly branched, and s bears hundreds of flowers. Here we show that S and AN, which encode a homeobox transcription factor and an F-box protein, respectively, control inflorescence architecture by promoting successive stages in the progression of an inflorescence meristem to floral specification. S and AN are sequentially expressed during this gradual phase transition, and the loss of either gene delays flower formation, resulting in additional branching. Independently arisen alleles of s account for inflorescence variation among domesticated tomatoes, and an stimulates branching in pepper plants that normally have solitary flowers. Our results suggest that variation of Solanaceae inflorescences is modulated through temporal changes in the acquisition of floral fate, providing a flexible evolutionary mechanism to elaborate sympodial inflorescence shoots.
| Among the most distinguishing features of plants are the flower-bearing shoots, called inflorescences. Despite a solid understanding of flower development, the molecular mechanisms that control inflorescence architecture remain obscure. We have explored this question in tomato, where mutations in two genes, ANANTHA (AN) and COMPOUND INFLORESCENCE (S), transform the well-known tomato “vine” into a highly branched structure with hundreds of flowers. We find that AN encodes an F-box protein ortholog of a gene called UNUSUAL FLORAL ORGANS that controls the identity of floral organs (petals, sepals, and so on), whereas S encodes a transcription factor related to a gene called WUSCHEL HOMEOBOX 9 that is involved in patterning the embryo within the plant seed. (F-box proteins are known for marking other proteins for degradation, but they can also function in hormone regulation and transcriptional activation) Interestingly, these genes have little or no effect on branching in inflorescences that grow continuously (so-called “indeterminate” shoots), as in Arabidopsis. However, we find that transient sequential expression of S followed by AN promotes branch termination and flower formation in plants where meristem growth ends with inflorescence and flower production (“determinate” shoots). We show that mutant alleles of s dramatically increase branch and flower number and have probably been selected for by breeders during modern cultivation. Moreover, the single-flower inflorescence of pepper (a species related to tomato, within the same Solanaceae family) can be converted to a compound inflorescence upon mutating its AN ortholog. Our results suggest a new developmental mechanism whereby inflorescence elaboration can be controlled through temporal regulation of floral fate.
| A striking manifestation of plant evolution is observed in the diverse branching and patterning of inflorescences, which are the shoots that bear flowers [1,2] Inflorescences are derived from the growth of dome-shaped groups of pluripotent cells called apical meristems. Apical meristems first produce leaves, and upon flowering induction, they produce inflorescence meristems that transition to floral meristems, which produce flowers. Extensive variation in inflorescence complexity is found in the nightshade (Solanaceae) family, where flowering marks the end of main shoot growth, and vegetative aerial growth is renewed from axillary meristems in a perennial growth system known as “sympodial” [3–5]. The simplest Solanaceae inflorescence is a solitary flower, represented by pepper (Capsicum annum) in Figure 1A. Tomato (Solanum lycopersicum), on the other hand, generates a few-flowered inflorescence organized in a zigzag branch (Figure 1B), but there are three classical mutants called compound inflorescence (s) (Figure 1D and 1E), anantha (an) (Figure 1F and 1G), and falsiflora (fa) (Figure 1H) that bear highly branched inflorescences resembling wild Solanaceae species like S. crispum (Figure 1C) [6–8] These similarities suggest that branching complexity may arise from tuning a common underlying developmental program. We set out to begin to unravel the basis of Solanaceae inflorescence diversity using these mutants whose variation ranges from branched inflorescences that produce hundreds of fertile flowers as seen in s [6], to the branching shoots of an that terminate in cauliflower-like tissue [7], to the leafy inflorescences of fa, which is defective in the tomato ortholog of LEAFY (LFY) [9].
The tomato plant is a compound shoot formed from reiterated sympodial shoot units (SYM) that arise from vegetative meristems that produce three leaves before terminating with an inflorescence [10]. The tomato inflorescence is also a compound shoot, which is condensed, consisting of sequential one-nodal inflorescence sympodial units (ISUs) each terminated by a single flower [11]. During early inflorescence development, individual ISUs developed in a progression of two phases. In the first phase, a sympodial inflorescence meristem (SIM), which was distinct from a SYM because it formed within the inflorescence itself, arose and produced a new SIM on its side before differentiating into a floral meristem (FM) in a second phase. These events created the first ISU and the SIM of the second ISU (Figure 2 and Figure S1). This pattern reiterated as subsequent SIMs developed perpendicular to one another, producing a zigzag pattern of flower initiation (Figure 2A). In an and fa mutants, the primary meristems failed to become flowers, remained indeterminate, and repeatedly initiated secondary SIMs that, themselves, repeatedly produced SIMs (Figure 2 and Figure S1). s was more asynchronous, as SIMs eventually transitioned to flowers after producing 2–4 axillary SIMs in a variable, environment-dependent manner (Figure 2 and Figure S1).
Although the branching effects were similar between the three mutants, floral phenotypes were not. Mutants of fa were primarily vegetative, producing numerous leaves that developed early as primordia coming off the flanks of meristems (Figure 1H and Figure S1K). Mutants of an, on the other hand, produced leaf primordia mixed with other tissue that at maturity resembled modified sepals or bracts (Figure 1F and 1G). It is interesting that s mutants maintained the capacity to produce normal flowers, indicating a reduced role in the flower relative to the SIM, although on occasion we observed some leaf-like primordia (Figure S1F). Thus, beyond distinctions in controlling floral organ identity (Figure 1), s, an, and fa mutants exhibit delayed ISU maturation, resulting in additional branches through the ongoing initiation of lateral SIMs. Notably, SIM branching in diverse Solanaceae is based on an s-like program, as seen in early inflorescence development of S. crispum (Figure 2D and Figure S2). This suggests that delays in floral termination (perhaps mediated by S, or the genetic pathway that S defines) provide a developmental framework for the modulation of sympodial branching in the Solanaceae.
To identify the genes responsible for these phenotypes, s and an were localized to linked regions of chromosome 2, and s was positionally cloned using a remarkable level of multi-genome synteny between the eudicot species poplar (Populus trichocarpa), Barrel Medic (Medicago truncatula), and grape (Vitis vinifera) (Figure 3). Several genes were shared in a short chromosomal segment ranging from 105–140 kb, and aligning these regions revealed three transcription factors: two AP2-like genes and a WUSCHEL-homeobox (WOX) that each co-segregated with s. Sequencing of all three genes revealed independent point mutations in the WOX gene from two alleles of s (s-classic and s-multiflora), and Southern blot analysis showed chromosomal changes in an additional allele (s-n5568), demonstrating that s is mutated in this gene (Figure 3 and Figure S3B and S4) (GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) accessions FJ190663 and FJ190664). To our knowledge, this is the first example of gene identification using multi-genome synteny among four eudicot species. Our data suggest that an even greater level of synteny remains to be discovered, and that non-model species will realize similar benefits as more genomes are sequenced.
WOX proteins share homology with the meristem maintenance gene WUSCHEL and are plant-specific transcription factors [12]. Among 14 WOX genes in Arabidopsis, S is most similar to WOX9/STIMPY (STIP) that functions with WOX8/STIPL to regulate embryonic patterning [13]. In contrast, we have found that S is a major determinant of inflorescence architecture in tomato. In the context of a European Solanaceae project (Eu-Sol), we established and phenotyped a collection of more than 6000 domesticated tomato varieties for various traits (Materials and Methods; https://www.eu-sol.wur.nl). The power of such a large germplasm resource resides in the fact that extensive natural allelic variation with both qualitative and quantitative effects has been selected and maintained since tomato was first domesticated [14]. Thus, these varieties provide a complement to the stronger, mostly deleterious effects, of alleles derived from artificial mutagenesis [8]. Among the 6,000 tomato lines, we identified 23 accessions with highly branched inflorescences and all were allelic to s Surprisingly, 22 of these lines carried the s-classic allele of the original mutant described 100 years ago, indicating that early breeders were positively selecting this mutant, probably for aesthetic value and fruit production (Figure 4) [6]. However, it was also possible that one or more of these lines arose independently, generating a mutation in the same nucleotide as s-classic. To address this question, we sequenced the coding region of all 22 lines and un-branched controls and found that all were identical except CC5721. Interestingly, this line carried four single-nucleotide polymorphisms (SNPs) that were shared with at least one un-branched variety, indicating that the s mutation in CC5721 may have arisen independently from a genetically distinct progenitor line (GenBank accessions FJ190665, FJ190666, and FJ190667). Two pieces of evidence lend support to this claim. Firstly, the four SNPs were distributed close (all within 1,000 bp) to the s lesion. Secondly, we sequenced a short segment of DNA from the tightly linked bacterial artificial chromosome (BAC) 298N3 and found that CC5721 had six SNPs and a 9–bp insertion-deletion (indel) that distinguished it from the other 21 domesticated types carrying s-classic (Figure 3b) (GenBank accessions FJ215691 and FJ215692). Still, in the absence of a geographic distribution of haplotypes, we cannot exclude a remote possibility that CC5721 arose as a result of an intra-genic recombination between s-classic and an unbranched variety. Regardless, at least three independently arisen alleles of s (s-classic, s-multiflora, and Rose Quartz Multiflora) are responsible for a major portion of the diversity in tomato inflorescence architecture.
The similarity between the phenotypically strong allele s-multiflora and strong an mutants suggested a functional link in regulating an underlying inflorescence branching program (Figure 1). Furthermore, we created double mutant plants of weak an alleles and s and found they were phenotypically enhanced to resemble strong an (Figure S5). Interestingly, stronger phenotypes were observed for both inflorescence branching and floral identity. Specifically, we found that the sepal and carpelloid tissue of weak an mutants became much more meristematic with less organ identity (Figure S5B). In some double mutants, additional leaves formed in the inflorescence, resembling fa mutants (unpublished data). This suggests that S and AN have overlapping roles in inflorescence architecture as well as floral identity. We noted that an resembled a Lotus japonicus mutant called proliferating floral organs (pfo) (Figure 1e) [15]. PFO encodes an F-box protein orthologous to Antirrhinum FIMBRIATA (FIM) and Arabidopsis UNUSUAL FLORAL ORGANS (UFO) [16], and the tomato ortholog of this gene co-segregated with an. Six alleles had mutations in the coding region, revealing that an is mutated in the tomato ortholog of FIM/UFO (Figures 2C, 3E, and Figure S3A and S6) (GenBank accession FJ190668). The similar inflorescence and floral phenotypes found in an and fa mutants [17] may, therefore, stem from conserved functional associations of their gene products as described in Arabidopsis [18]. However, the relationship between S and AN was less clear, and their expression patterns were therefore explored.
S was expressed to varying degrees in all tissues except roots, whereas most AN expression was restricted to floral buds, indicating a primary function in inflorescence and flower development. FA accumulated predominantly in shoot apices (Figure 5A)[9]. We explored further the expression of S and AN using in situ hybridization, which revealed temporally distinct patterns during inflorescence development. S was expressed in a wedge shape radiating outward from 2–3 cells from the center of immature SIMs (Figure 5B and 5C). This expression initiated shortly after lateral bulging of the SIM and was transient, because it disappeared before floral termination. AN expression initiated in incipient FMs shortly after down-regulation of S. AN expression was less intense than S, and was limited to the upper layers of the rapidly maturing SIM (Figure 5D and 5E). Both genes were reactivated in flower primordia in a ring of cells that marked a boundary domain, first between sepal and petal primordia and later between petals and stamens. These floral expression patterns are consistent with the failure of an mutants to initiate normal flowers, suggest a role for S in the flower, and likely explain the enhanced developmental and molecular phenotypes that s imposes on floral organ identity in weak alleles of an (Figure S5). Indeed, double mutants show little or no an expression similar to strong an mutants alone (Figure S5D).
The expression pattern of S suggested that it functioned early in SIM maturation to promote the transition to FM, whereas AN operated soon after to provide early FM identity. To test these hypotheses, we examined the expression of S and AN in s, an, and fa mutants. S was expressed in all mutant backgrounds, and, as in wild-type, was detected in younger lateral SIMs of an inflorescences (Figure 5F and 5G). This indicates that an meristems still reach a pre-floral SIM state. AN expression, on the other hand, was undetectable by RT-PCR (reverse-transcriptase PCR) in fa mutants, consistent with the proposal that FA functions upstream of AN [17](Figure 5F). Initial expression of AN in s mutants was delayed, and subsequently detected in only a small subset of SIMs compared to wild-type. In those meristems expressing AN, the signal was deeper and more intense than normal (Figure 5H). In situ hybridization from older inflorescences revealed some meristems lacking S and AN activity altogether, which we verified by whole-mount in situ hybridization (Figure 5H and unpublished data) This indicates that different meristems are at different phases of ISU maturation, and may also reflect the frequent observation of modified leaves or bracts in older an inflorescences if some meristems retain a more vegetative state. Taken together, these expression patterns support a mechanism where S and AN promote successive stages in the progression of an inflorescence meristem to floral specification through sequential transient activities that gradually promote maturation of SIMs (expressing S) to early FMs (expressing AN) (Figure 5I). Loss of either gene provides SIMs with an extended period of indeterminacy that facilitates ISU elaboration according to an underlying program of sympodial growth (Figure 5J). Furthermore, the observation that S expression is maintained in an mutants and vice-versa, and that their expression is restricted to temporally distinct domains, supports the notion that these genes have separate but overlapping functions in the maturation of individual ISUs, consistent with the enhancement of weak an alleles by s (Figure S5).
The expression patterns of S and AN along with their mutant phenotypes lead to a model in which temporal differences in the maturation of a SIM to an FM can regulate the duration of sympodial inflorescence branching. In other words, a slower transition enables more inflorescence branching and vice-versa. This suggests that the SIM phase and the early FM phase of a single flower can each provide a developmental window in which a compound inflorescence can form. We tested this hypothesis genetically by taking advantage of mutants of single flower truss (the tomato ortholog of FT, which is a major component of florigen), whose inflorescences are indeterminate vegetative shoots with single flowers separated in space by leaves [19] (Figure 6A). In sft:an double mutants, we observed that individual flowers became branched inflorescences, though less so than in an mutants alone (2–4 versus 20–25 branches at the same age, Figures 1F, 6B, and 6C). By contrast, branching in sft:s double mutants resulted in elaboration of the vegetative inflorescence, but normal flowers still formed (Figure S7). Taken together, these results support the proposal that S acts earlier within a single inflorescence meristem to regulate sympodial branching, whereas AN acts later as FM identity is reached.
Our model suggests that Solanaceae inflorescences with only single flowers may result from rapid termination of the FM and hence elimination of the SIM stage, but that single flower species can still produce branched inflorescences. We addressed this by mutagenizing pepper (C anuum), which identified one mutant (called Ca-an) that produced an indeterminate shoot instead of a flower. This structure lacked petals and stamens and branched more extensively in a mixed genetic background, resembling tomato an mutants (Figure 6E and Figure S8). We sequenced pepper AN from Ca-an and found a missense mutation from the wild-type progenitor sequence causing a nucleotide change just prior to one of our tomato an alleles (an-e1444) that co-segregated with the mutant phenotype (Figure S6), indicating that Ca-an is mutated in the pepper ortholog of FIM/UFO (GenBank accession FJ190669). Like tomato, Ca-AN was expressed in a ring of cells flanking developing petals and stamens (Figure S8). Interestingly, Ca-AN could not be detected in an earlier inflorescence meristem, which lends support to the idea that pepper has a short SIM phase and progresses rapidly to floral termination. Yet, Ca-an mutants revealed a latent potential to branch, indicating that Solanaceae AN shares a conserved role in promoting FM determinacy with its orthologs in other species [15,16,20,21]. Of all other known UFO mutants, the pfo mutant from L. japonicus is most similar to Ca-an, with a compact branched structure described as a reiteration of sepals and FMs. Normal L. japonicus produces pairs of flowers in the axils of leaves, and so loss of UFO function provides an extended period of indeterminacy to each pair of inflorescence meristems. By contrast, the stp mutant of pea generates similar organ defects but produces secondary FMs within the primary flower. Thus, UFO has a highly conserved role in floral identity, but its control of inflorescence branching is more species-specific and likely reflects differences in mechanisms of inflorescence meristem initiation. Notably, branching of tomato an mutants was more extreme than in Ca-an mutants (Figures 1 and 6). This indicates that underlying the tomato SIM phase is a program promoting branching and that the foundation for more complex branching is an inflorescence composed of reiterated SIMs. These data suggest that highly branched species like S. crispum evolved from an ancestral form that resembled tomato, as opposed to pepper (Figure S2).
Our results reveal a genetic foundation for the Solanaceae inflorescence and provide evidence for a possible mechanism that modulates simple and complex inflorescence structures known as “cymes” [1,2] While the generation of a cymose inflorescence through sympodial growth is likely a complex process involving many unknown genetic and environmental factors, we provide a major advance in understanding how cymes may be modified into more complex structures based on elaboration of the ubiquitous ISU shoot system (Figure 2) [5]. This mechanism uses conserved machinery (AN/UFO and FA/LFY) that regulates inflorescence and flower development in other species [15,16,20–26]. Interestingly, the effects of UFO on inflorescence architecture vary considerably, ranging from infrequent replacement of single flowers with secondary inflorescence shoots in Arabidopsis ufo mutants [27], to the production of ectopic flowers in the inflorescences of pea stp mutants [20], to the large mass of inflorescence/floral tissue in pfo mutants of L japonicus [15], and as shown here, the an mutant of tomato and pepper. Furthermore, we describe S/WOX9 as a novel component in the control of inflorescence architecture—a role that was not detected for its Arabidopsis ortholog. We also find that the tomato ortholog of TERMINAL FLOWER1 (TFL1) called SELF PRUNING (SP), which has a major effect on Arabidopsis inflorescences [2], is neutral on sympodial inflorescence branching in normal tomato inflorescences (unpublished data), and exhibits indirect effects on s inflorescence branching (Table S1). These differences may originate from the evolution of distinct growth habits. Branching complexity in sympodial species relies on termination of inflorescence meristems through the transition of a SIM to an FM. We suggest that a transient expression of S followed by AN was co-opted in Solanaceae sympodial development to boost two phases of sympodial meristem growth in this specialized shoot, both of which can potentiate branching (Figure 3I). In monopodial dicot species such as Arabidopsis or Antirrhinum, the inflorescence meristem produces no comparable SIMs, being indeterminate and generating lateral single flowers. This indeterminacy may explain why WOX9, by itself, is dispensable for inflorescence development [28,29]. Indeed, inflorescence ramification in monopodial dicot plants is more often stimulated through identity change [16,30,31], which could also explain some of the branching effects observed in Ca-an (Figure 6)
Thus, while the evolutionary diversification of plant inflorescence architecture is united under a common developmental theme [2], plants with different growth habits use related as well as distinct developmental modules to regulate branching [32]. We propose that Solanaceae inflorescence variation is based on controlling sympodial branching through temporal changes in the acquisition of floral fate, which is most flexible within the SIM phase. Short delays in the activation of genes like S (or as-yet-undiscovered other genes in the S pathway) followed by an abrupt switch to floral termination may explain the evolution and quantitative variation of compound inflorescences in the genus Solanum (Figure S9), as well as in other sympodial species, like trees [5]. Such a mechanism would provide a flexible way to guarantee the production and simultaneous maturation of large numbers of flowers, thereby ensuring a crucial aspect of reproductive success and perhaps providing a new tool for the manipulation of crop yields.
Classic alleles of s (s-classic LA3094), an (LA0536), and fa (LA0854), and those of representative wild tomato species were gifts from the C. M. Rick Center (Davis, California; http://tgrc.ucdavis.edu). An additional allele of s (LA0560; s-multiflora, C. M Rick Center) was verified by complementation test. A third s allele and six additional an alleles were identified as inflorescence mutants in a screen of a tomato mutant library [8]. Wild tomato species were gifts from the C. M. Rick Center. Wild tomato species, such as S. lycopersicoides, can be difficult to grow and maintain until flowering and only two representative plants were available for phenotypic analyses, but inflorescence complexity within each plant was uniform throughout. More distantly related Solanum species were gifts from the Botanical and Experimental Garden at Nijmegen, The Netherlands. Up to three representative plants were used for phenotypic analyses. The ∼6,000 domesticated tomato varieties were collected from various public and private germplasm sources. All plants were grown in greenhouses under natural light or in agricultural field conditions in Israel using standard irrigation and fertilization regimes.
The s mutant was originally mapped on the long arm of chromosome 2, and verified using 20 mutants selected from an F2 population derived from a cross with the wild species S. pimpinellifolium (LA1589). This positioned s in the region overlapping introgression lines IL2–3/2–4 on the tomato introgression line map [33]. A larger mapping population was generated by crossing s-n5568 with the wild tomato species S. pennellii (LA0716). F1 hybrid plants were self-fertilized to produce a mapping population of 5,000 F2 plants. Five hundred individual s mutant plants were scored with CAPS-PCR markers from the most current tomato genetic map (Solanaceae Genomics Network at http://www.sgn.cornell.edu), focusing on the region of IL2–3/2–4. Additional markers surrounding the tightly linked CNR locus were provided by K. Manning [34]. Marker density was improved using conserved synteny identified between seven markers in a 15-cM window in tomato and a 500-kb segment of Arabidopsis chromosome 1 (marker information available upon request). Two co-segregating markers (0 recombinant chromosomes out of 1,000 gametes) were used to isolate a BAC from a S. lycopersicum HindIII library kindly provided by J. J. Giovannoni and J. VanEck at Cornell University (Ithaca, New York). DNA fragments from three independent restriction enzyme digestions of a single BAC clone were sub-cloned into TOPO TA cloning vectors (Invitrogen) for shotgun sequencing. Fragments containing genes were annotated using BLASTX against the Arabidopsis protein database and used to search other genomes for additional synteny. Sequences from four tightly linked markers (Figure 3) were used in a BLASTN or TBLASTX search against genomes of P trichocarpa, M truncatula, and V vinifera. Genes in syntenic regions ranging from 110–140 kb were aligned manually and searched for candidate genes, which identified the Apetala-2 (AP2) and Wuschel-homoebox (WOX) transcription factors. A tomato-specific WOX marker was produced by degenerate PCR based on conserved regions in the WOX from these three species and an EST from Petunia hybrida (accession number EB174485). Transcript ends were determined by rapid amplification of cDNA ends (RACE) (Sambrook) using total RNA isolated from young inflorescences with TriReagent (Sigma-Genosys). DNA from s-like varieties with compound inflorescences from the Core Collection was PCR amplified with gene-specific primers and used in a CAPS-PCR assay diagnostic of s-classic.
The expressivity of the s phenotype is affected by genetic background, which became evident when phenotyping 22 domesticated varieties each carrying the s-classic allele, but varying in many phenotypic characters, including branching. Modifiers are responsible for these differences, which may or may not have a functional relationship to S. Furthermore, it is well-documented that sympodial shoot growth in tomato is highly sensitive to light intensity, which could also contribute to quantitative variation between accessions. On occasion, modestly branched accessions were observed that produced only 2–4 additional branches compared to normal, which, if not allelic to s, could potentially modify (enhance) the s phenotype. Yet, the majority of extreme branching variation was due solely to changes in S function, indicated by normal segregation of families segregating for each s allele in a common genetic background (cv. M82). Thus, differences in phenotypic strength, as seen in s-multiflora, result from modifier loci, but these are much weaker in their effects compared to s mutations.
The an mutant was originally mapped to the long arm of chromosome 2, and subsequently positioned in the region overlapping IL2–3/2–4/2–5. The phenotype of the pfo mutant in L japonicus resembled weak alleles of an and led us to search for the tomato ortholog of FIM/UFO. A single EST (SGN-U341425) with homology to FIM/UFO was used to generate a CAPS-PCR marker that mapped to the same region as an (http://www.sgn.cornell.edu). DNA from six EMS alleles was amplified using gene-specific primers and sequenced directly, which identified five independent mutations. The central portion of coding sequence of the an-classic allele could not be amplified by PCR, suggesting a structural change or large insertion (unpublished data). This rearrangement in the an-classic allele was verified using DNA Southern blot hybridizations (Figure S3) according to established protocols.
An EMS mutagenesis of the pepper variety Maor was performed according to a protocol for tomato seeds [8]. Among 1,500 M2 families of pepper, one inflorescence mutant was identified based on phenotypic similarity to weak alleles of tomato an This new mutant was first mapped by restriction fragment length polymorphism (RFLP) analysis to a region of chromosome 2 in pepper that is syntenic with tomato chromosome 2 where anantha was positioned previously. DNA from the mutant was isolated and sequenced using primers designed from the tomato gene. Co-segregation of the mutation with the pepper an phenotype was verified in an F2 population of approximately 100 plants, and the mutation was found to be derived from the Maor variety progenitor sequence. Our allele changes a nearly invariant glycine among F-box proteins into a charged amino acid, glutamic acid (see Figure S6). This glycine is the second amino acid in a short stretch of ∼10 highly conserved amino acids in UFO orthologs for which at least one mutant allele is available in Arabidopsis, pea, tomato, and now pepper. Thus, this region is a hot spot for mutations that give very similar floral phenotypes in multiple species.
Developmental and morphological analyses on single and double mutants were performed on alleles originating from the tomato cultivar M82. M82 lines were either mutant or wild type for the gene SELF PRUNING (SP), which had only modest effects on inflorescence phenotypes in s or an that could be attributed to changes in the length of sympodial units—a phenotype regulated by SP. Single and double mutants of sft were the allele sft-7187 [19]. Mutants of fa were in the background of Rheinlands Ruhm. Branching events were counted on two independent inflorescences from Core Collection varieties containing s-classic and non-s controls. For SEM, immature inflorescences from sympodial shoots were dissected and processed through an EtOH series, critical-point dried, and coated with gold particles for microscope analysis on a Philips XL30 ESEM FEG. RT-PCR was performed using a One-Step RT-PCR kit (Qiagen) on total RNA isolated by TriReagent (Sigma-Genosys) according to the manufacturers' protocols. Primer sequences are available upon request. Tissues for in situ hybridization were dissected and fixed according to standard protocols [35]. In vitro transcribed RNA probes were generated from 5′ partial (S) or full-length (AN) cDNA clones and transcripts were detected using standard in situ hybridization techniques. Whole-mount in situ hybridization was performed as described [36], using the same probes.
|
10.1371/journal.pcbi.1005385 | Classification and adaptive behavior prediction of children with autism spectrum disorder based upon multivariate data analysis of markers of oxidative stress and DNA methylation | The number of diagnosed cases of Autism Spectrum Disorders (ASD) has increased dramatically over the last four decades; however, there is still considerable debate regarding the underlying pathophysiology of ASD. This lack of biological knowledge restricts diagnoses to be made based on behavioral observations and psychometric tools. However, physiological measurements should support these behavioral diagnoses in the future in order to enable earlier and more accurate diagnoses. Stepping towards this goal of incorporating biochemical data into ASD diagnosis, this paper analyzes measurements of metabolite concentrations of the folate-dependent one-carbon metabolism and transulfuration pathways taken from blood samples of 83 participants with ASD and 76 age-matched neurotypical peers. Fisher Discriminant Analysis enables multivariate classification of the participants as on the spectrum or neurotypical which results in 96.1% of all neurotypical participants being correctly identified as such while still correctly identifying 97.6% of the ASD cohort. Furthermore, kernel partial least squares is used to predict adaptive behavior, as measured by the Vineland Adaptive Behavior Composite score, where measurement of five metabolites of the pathways was sufficient to predict the Vineland score with an R2 of 0.45 after cross-validation. This level of accuracy for classification as well as severity prediction far exceeds any other approach in this field and is a strong indicator that the metabolites under consideration are strongly correlated with an ASD diagnosis but also that the statistical analysis used here offers tremendous potential for extracting important information from complex biochemical data sets.
| Autism spectrum disorder (ASD) encompasses a family of neurological disorders characterized by limited social interaction and restricted repetitive behaviors. The number of children diagnosed with ASD has grown exponentially over the last four decades and is now estimated to affect ∼1.5% of children. Although ASD is currently diagnosed and treated based solely on psychometric tools, a biochemical view applicable to at least a subset of ASD cases is emerging. Abnormalities in folate-dependent one carbon metabolism and transsulfuration pathways can summarize a large number of observations of genetic and environmental effects that increase ASD predisposition. However, these complex, highly interconnected pathways require more advanced statistical models than the typical univariate models presented in the literature. Therefore, we developed multivariate statistical models that classify participants based on their neurological status and predict adaptive behavior in ASD. We emphasize that these models are cross-validated, helping to ensure that the results will generalize to new samples. The models developed herein have much stronger predictability than any existing approaches from the scientific literature.
| Autism Spectrum Disorder (ASD) encompasses a large group of early-onset neurological diseases characterized by difficulties with social communication/interaction and expression of restricted repetitive behaviors and interests [1]. In addition to these defining behavioral symptoms, individuals with ASD frequently have one or more co-occurring conditions, including intellectual disability, ADHD, speech and language delays, psychiatric diagnoses, epilepsy, sleep disorders, and gastrointestinal problems [2–5]. ASD affects ∼1.5% of the population and affects males disproportionately [6–8]. It is associated with an impaired quality of life [9] and the lifetime cost of supporting an individual with ASD amounts to $1.4–2.4MM, depending on co-existing disorders [10].
It is generally acknowledged that ASD has a strong genetic component, but environmental effects have also recently emerged as important contributors to the etiology and pathophysiology of ASD in at least a subpopulation of cases. Early twin studies suggested that the heritability of ASD was 80–90% [11]; however, twin studies since 2010 suggest a lower heritability of only 37–55% [12, 13]. Despite this high genetic association, only 15% of ASD cases have a known genetic source [1]. Although genetic studies continue to provide new evidence for contributing factors to ASD etiology [14], environmental effects such as maternal/paternal age, toxic chemical exposure, maternal rubella infection, etc. are also emerging as key factors contributing to ASD liability [13].
No generally accepted biomarkers for the diagnosis or diagnosis of the severity of ASD exist to date. Instead, diagnostic evaluation involves a multi-disciplinary team of doctors usually including a pediatrician, psychologist, speech and language pathologist, and occupational therapist. Despite this current state of the art, work in identifying biomarkers that can support the diagnosis process is ongoing. In particular, abnormalities in folate-dependent one-carbon metabolism (FOCM) and transsulfuration (TS) likely contribute to the genetic and environmental predisposition to ASD [15]. FOCM contributes to epigenetic gene expression through DNA methylation and TS is the major contributor to intracellular redox status. An illustration of these pathways overlaid with genetic and environmental contributions to ASD predisposition is presented in Fig 1.
Mutations or altered expression levels of several genes in these pathways have been associated with increased risk of ASD. Adenylosuccinate lyase (ADSL) deficiency leads to a purely genetic form of autism by re-directing a large proportion of FOCM toward purine synthesis to compensate for a reduction in de novo purine synthesis [15, 16]. Methylenetetrahydrofolate reductase (MTHFR) is responsible for generating 5-methyltetrahydrofolate, which in turn is responsible for re-methylating homocysteine to methionine. In particular, the C677T polymorphism has been shown to increase ASD liability, especially in countries where prenatal folate supplementation is low [17]. Limited evidence linking mutations in reduced folate carrier (RFC1) [18, 19], transcobalamin II (TCII) [18], serine hydroxymethyltransferase I (SHMT1) [20], 5-methyltetrahydrofolate-homocysteine methyltransferase reductase (MTRR) [18, 20], and catechol-O-methyltransferase (COMT) [18, 21] to altered prevalence of ASD has also been presented, although these contributions to ASD liability are currently contested [22].
Evidence for the association between environmentally-rooted FOCM/TS dysfunction and ASD predisposition can be seen in prenatal valproate and toxic chemical exposure as well as lack of maternal folate supplementation. Maternal valproate use during pregnancy has been associated with higher incidence rates of ASD [23, 24] and in utero valproate exposure has been used to develop rodent models of autism [25]. Valproate exposure causes DNA hypomethylation [26, 27] in key neurodevelopmental processes that have been mitigated by folate supplementation [28]in vitro. Other chemicals such as heavy metals, ethyl alcohol, pesticides, phthalates, polychlorinated biphenyls, and traffic-related air pollution (TRAP) have also been shown to affect neurodevelopment and increase ASD liability [13, 29]. These organic toxins induce oxidative stress and heavy metals disrupt transsulfuration by binding glutathione, the major contributor to intracellular redox homeostasis [30]. Additionally, glutathione is an important regulator in the intracellular processing of methylcobalamin (vitamin B12), an essential cofactor for methionine synthase and the TS pathway [31]. Air dispersion models coupled with traffic patterns/roadway geometry, meteorological data, and vehicle emission data have been used to find a dose response between ASD prevalence and TRAP exposure [32]. Additionally, common organic pollutants have been associated with increased autism severity in children on the autism spectrum [33]. Two independent studies linked maternal folate supplementation to a reduced risk of having a child with ASD [34, 35]. This protective effect is usually attributed to the involvement of FOCM in early epigenetic regulation of neurodevelopment and neural tube formation [21, 36]. For a more complete description of the evidence for the potential contributions of FOCM/TS dysfunction to the ASD phenotype, see the excellent review by Deth et al. [15].
Although differences between FOCM and TS pathways in children with ASD versus neurotypical controls have been shown previously [18, 37, 38], investigators have struggled with identifying a single, predictive measurement of these pathways that separates individuals with ASD from neurotypical controls or that correlates well with ASD severity. However, in many complex problems one particular measurement may be insufficient and important information can only be extracted by using multivariate statistical analysis. Indeed, incorporating multiple measurements of environmental toxins has been shown to increase the separability of control and ASD participants [39] and better predict autism severity [33, 39].
Latent variable techniques enable the discovery of important multivariate interactions, leading to improved classification and regression performance. Furthermore, latent variable techniques allow assessing the importance of individual variables and are more robust to uninformative variables. One popular latent variable technique for classification problems is Fisher Discriminant Analysis (FDA), which achieves an optimal linear separability using a typically small set of latent variables that are linear combinations of the original variable set. FDA has a long history in biological classification problems and was first used by Rao in 1948 to interpret anthropological data [40]. Extensions of FDA, such as Kernel FDA (KFDA), exist which can take nonlinear relationships into account for classification [41]. Latent variable regression techniques include partial least squares (PLS) and its nonlinear counterpart kernel PLS (KPLS) [42, 43]. Using FDA for classification and KPLS for regression allow multivariate interactions to surface, which are often hidden when only univariate analysis is considered. To guarantee a statistically independent assessment of the multivariate classification and regression models, the presented study utilizes a cross-validatory approach, where the set of samples used for model identification does not contain samples to evaluate the performance of the identified models.
The presented work makes use of these advanced modeling and statistical analysis tools to examine metabolite data of the FOCM/TS pathway in neurotypical participants (NEU) and those on the autism spectrum (ASD) as well as their siblings (SIB). Using FDA, it is possible to clearly distinguish the participants on the spectrum from their neurotypical peers and KPLS unveils a strong correlation between metabolite concentrations of these pathways and adaptive behavior as measured by the Vineland Adaptive Behavior Composite. This work not only analyzes the largest data set of its kind of these pathways in the scientific literature [38], but also results in the strongest evidence to date of the association of FOCM/TS dysfunction with ASD.
Associating dysfunction of FOCM/TS pathways with ASD requires a distinction between or separation of ASD and NEU groups based on FOCM/TS metabolites. Therefore, cross-validatory FDA was performed using measurements of the FOCM/TS metabolites listed in Table 1. A linear classifier based on these FDA scores is then used to classify ASD and NEU participants. FDA scores and estimated probability distribution functions (PDFs) are provided in Fig 2. The cross-validated misclassification rates of only 4.9% and 3.4% for the NEU and ASD samples, respectively, eliminated more complex, nonlinear KFDA analysis from consideration.
The performance of the classifier was then evaluated on the SIB cohort, a more challenging classification problem due to partially shared genetic and environmental effects with the ASD cohort. Using all measurements in Table 1, an FDA model was trained to separate the ASD and NEU cohorts. Then, the trained FDA model was used to evaluate the SIB cohort (which was not used for training). The resulting separation of ASD, NEU, and SIB presented in Fig 3 shows a slight increase in the overlap with the ASD cohort when compared with the performance of the ASD vs. NEU classification. Furthermore, the SIB PDF shows significantly more overlap with the NEU PDF than the ASD PDF. These results support the hypothesis proposed by James et al [38] that the siblings of the participants on the spectrum have FOCM/TS metabolite profiles that are significantly more similar to their neurotypical peers than their siblings, even though genetically they are likely closer to their siblings than participants in the neurotypical control group.
The simultaneous use of multiple measurements promises to increase the separability of the cohorts; however, increasing the number of measurements increases the number of parameters in the projection vector w that maximizes the separability of the two groups (see Materials and methods). Although cross-validation can help mitigate these effects, the increased number of parameters can lead to over-fitting, which would indicate good performance for separation on the existing data set, but poor separation performance when the analysis results are translated to new data. These over-fitting problems can be further mitigated by selecting only the minimum number of variables required to adequately separate the two groups. Therefore, all combinations of up to six variables were evaluated for separability. Select combinations of higher numbers of variables were chosen in a greedy fashion to sequentially add measurements that best improve the separation of the best six variables. Cross-validatory FDA was performed on all variable combinations and probability distribution functions (PDFs) of the FDA scores of the two cohorts were estimated. A receiver-operating-characteristic (ROC) curve was generated based on these PDFs. The C-statistic of the ROC curve provides a measure of the ability of the classifier to separate into ASD and neurotypical groups. A C-statistic of 0.5 represents random classification and a C-statistic of 1.0 represents perfect classification. Fig 4 plots the maximum C-statistic for all combinations of a given number of variables. As the number of variables increases, the C-statistic increases, saturates at 0.997, and then slightly decreases when over-fitting occurs.
From these results, five variables (DNA methylation, 8-OHG, Glu.-Cys., fCystine/fCysteine, % oxidized glutathione) were considered for further analysis; however, it should be noted that select variable combinations distinct from this one provided similar performance for separating ASD and NEU participants. Chlorotyrosine and tGSH/GSSG were added to this set to improve separability of the ASD and SIB groups, increasing the number of metabolites under consideration to seven. The separability of the final minimal classifier based on these seven variables is presented in Fig 5 with Type I and Type II error plots in S1 Fig.
In addition to separation into neurologically distinct cohorts, metabolites in the FOCM/TS pathway were investigated for predictability of adaptive behavior. Due to the inter-dependency of pathway metabolites and possible nonlinear effects on psychological outcomes, nonlinear regression via KPLS was used to evaluate the ability of pathway metabolites to predict adaptive behavior in ASD (as measured by the Vineland Adaptive Behavior Composite score). Just as was done in the FDA analysis, all combinations of a given number of variables were evaluated for predictability. The cross-validatory R2 of the regression was then used to determine the optimal number of variables in the regression analysis. From the results in Fig 6, the R2 begins to decrease when more than five variables are used in the KPLS analysis. The maximum cross-validatory R2 was 0.45, corresponding to the KPLS model with the variable combination GSSG, tGSH/GSSG, Nitrotyrosine, Tyrosine, and fCysteine used as inputs. These regression results are plotted in Fig 6. (It is important to note that a few other variable combinations provided similar results, but only the best regression model is illustrated for clarity.) This strong correlation even after cross-validation indicates the importance of FOCM/TS dysfunction in the pathophysiology of ASD.
The multivariate statistical analysis presented herein provides unprecedented quantitative classification results for separating participants into ASD and NEU cohorts based solely on biochemical data. Existing analyses report differences in mean metabolite levels or provide qualitative illustrations of separating these two groups based on FOCM/TS metabolites [18, 37, 38]. However, these strategies are not designed for classification and thus fail to successfully classify participants. Here, FDA on seven metabolites allows sufficient separation such that a linear classifier can correctly resolve 96.9% of participants. Such low misclassification rates dissuaded the use of more complex, nonlinear methods such as KFDA. Although FOCM/TS dysfunction likely does not completely detail ASD etiology, this biochemical analysis approaches the accuracy needed for a clinical diagnostic tool.
Classification performance on the SIB group fortifies the argument for FOCM/TS involvement in ASD since the large degree of shared genetic and environmental effects with the ASD population only slightly worsens the separation. The sibling recurrence rate for ASD is estimated to be 6.9–18.7% [7, 44, 45] and many siblings perform behaviorally and/or cognitively at intermediate levels between those of ASD and NEU cohorts [45–47] or express traits characteristic of ASD [47–49]. Therefore, the classification performance placing the SIB group between the ASD and NEU groups, albeit much closer to the NEU group, is consistent with the broader scientific literature on psychometric analysis of siblings of people with ASD. Future work would benefit from assessing the SIB and NEU groups on measurements of the Broader Autism Phenotype to validate these hypotheses on mild FOCM/TS dysfunction in the SIB group.
Comparison or meta-analysis of regression analyses across studies is difficult due to differences in metabolites measured, origin of metabolites, available psychometric data, and metrics of model performance. It is emphasized that extreme caution should be used when evaluating fitted versus cross-validated metrics; for example, in [39], the best linear model can achieve a fitting R2 of 0.296, while obtaining a cross-validated R2 of only 0.192. In general, fitting results always surpass cross-validation results; nevertheless, the top-performing KPLS model in this study achieved a cross-validatory R2 of 0.45 due to its ability to reflect nonlinear behaviors/interactions, which surpasses or compares with previous fitting [50, 51] and cross-validated results [39].
Nonlinear regression analysis of FOCM/TS metabolites enables prediction of key FOCM/TS metabolites that are associated with adaptive behavior in ASD. Based upon all variable combinations evaluated in the KPLS regression analysis, top-performing models always incorporated (1) nitrotyrosine, (2) tyrosine, (3) fGSH or tGSH/GSSG, and (4) fCysteine or fCystine/fCysteine. Interestingly, these variables are affected by high quality vitamin supplementation that also decreases ASD severity in at least a subset of cases [51–53]. While this forms an intriguing direction for future studies, it should be noted that these studies should be replicated and empirically tested on a wider scale before more definite conclusions can be drawn. Furthermore, this approach can be extended to include other psychometric instruments (e.g. the Autism Diagnostic Observation Schedule (ADOS) or Childhood Autism Rating Scales (CARS)) that are more appropriate for diagnosis of ASD.
Developmental pediatricians, psychologists and other professionals can effectively use the wealth of information provided by psychometric instruments to diagnose and evaluate patients with ASD. However, these tests can rarely diagnose children under two years old since they are based solely on behavioral assessment. As it is generally acknowledged that an earlier diagnosis can lead to a more favorable outcome in the long run [54], the identification of biomarkers which can be used in conjunction with psychometric measurements would be of significant importance for ASD diagnosis. Furthermore, identification of these biomarkers can facilitate the understanding of these complex disorders, which offers significant potential for developing intervention strategies targeted to normalize these biomarkers in the future. However, it is important to note that these biomarkers may not simply be measurements of certain metabolites but may require nonlinear statistical analysis of the measurements, as is done in this work.
The data used in this study comes from the Arkansas Children’s Hospital Research Institute’s autism IMAGE study [38]. The protocol was approved by the Institutional Review Board at the University of Arkansas for Medical Sciences and all parents signed informed consent. The interested reader is referred to [38] for detailed study design, including demographic information and inclusion/exclusion criteria. Briefly, children between the ages of 3 and 10 years were enrolled to assess levels of oxidative stress. ASD was defined by the Diagnostic and Statistical Manual for Mental Disorders, Fourth Edition, the Autism Diagnostic Observation Schedule (ADOS), and/or the Childhood Autism Rating Scales (CARS; score > 30). FOCM/TS metabolites from 83, 47, and 76 case (ASD), sibling (SIB), and age-matched control (NEU) children, respectively, were used for classification. The metabolites under investigation are tabulated in Table 1 and additional details of these measurements and derivations are presented in [38]. Of the 83 participants on the autism spectrum, 55 also had Vineland II Scores recorded for use in regression analysis (range 46–106). The Vineland Adaptive Behavior Composite evaluates adaptive skills across the domains of communication, socialization, daily living skills, and motor skills through a semi-structured caregiver interview [55]. Data are available in S1 Dataset.
Fisher Discriminant Analysis (FDA) is a dimensionality reduction tool that seeks to maximize differences between multiple classes. Specifically, for n samples of m measurements associated with k different classes, the between cluster variability SB is defined to be
S B = ∑ i = 1 k n i ( x ¯ i - x ¯ ) ( x ¯ i - x ¯ ) T
where x ¯ i represents the mean vector of class i, x ¯ represents the mean vector of all samples, and ni represents the number of samples in class i. The within cluster variation is defined as
S W = ∑ i = 1 k n i ∑ j ∈ i ( x j - x ¯ i ) ( x j - x ¯ i ) T
where xj represents an individual sample. FDA seeks to find at most k − 1 vectors that maximize
J ( w ) = w T S B w w T S W w
In other words, FDA seeks to find linear combinations of variables that project samples in the same group close to each other and project samples in different groups far away from each other. The solution to this optimization problem is the generalized eigenvectors associated with the k − 1 largest generalized eigenvalues of S W - 1 S B.
Kernel density estimation attempts to determine the underlying probability distribution function from a set of reference samples. The main assumption is that additional samples are likely to be found near the reference samples [56–58]. Using a Gaussian kernel, this assumption is formulated into an algorithm by associating a kernel function
K x - x i σ
with each observation xi. Here, x is the additional sample and σ is the kernel parameter that controls the shape of the distribution function. The estimated density function f ^ ( x ) is then given by
f ^ ( x ) = 1 n σ ∑ i = 1 n K x - x i σ
where n is the number of reference samples. The kernel parameter σ is chosen to minimize the mean integrated squared error (MISE) between the unknown density function f(x) and the estimated density function f ^ ( x ):
MISE ( σ ) = ∫ - ∞ ∞ f ( x ) - f ^ ( x ) 2
using a cross-validatory approach [56].
Kernel techniques provide general nonlinear extensions to the popular linear partial least squares (PLS) regression. The KPLS algorithm commences by defining a nonlinear transformation f = ψ(x) on the predictor set x. In this work, ψ(x) is a Gaussian kernel. Rather than regression on x as in linear PLS, y is regressed onto the high dimensional feature space f [42, 43].
To avoid over-fitting and over-stating results, leave-one-out cross validation is employed in both the FDA and KPLS analysis. The approach leaves out a single sample, fits an FDA or KPLS model, and evaluates the prediction of the sample left out. This scheme is repeated for each sample.
|
10.1371/journal.pgen.1004295 | Genotypic and Functional Impact of HIV-1 Adaptation to Its Host Population during the North American Epidemic | HLA-restricted immune escape mutations that persist following HIV transmission could gradually spread through the viral population, thereby compromising host antiviral immunity as the epidemic progresses. To assess the extent and phenotypic impact of this phenomenon in an immunogenetically diverse population, we genotypically and functionally compared linked HLA and HIV (Gag/Nef) sequences from 358 historic (1979–1989) and 382 modern (2000–2011) specimens from four key cities in the North American epidemic (New York, Boston, San Francisco, Vancouver). Inferred HIV phylogenies were star-like, with approximately two-fold greater mean pairwise distances in modern versus historic sequences. The reconstructed epidemic ancestral (founder) HIV sequence was essentially identical to the North American subtype B consensus. Consistent with gradual diversification of a “consensus-like” founder virus, the median “background” frequencies of individual HLA-associated polymorphisms in HIV (in individuals lacking the restricting HLA[s]) were ∼2-fold higher in modern versus historic HIV sequences, though these remained notably low overall (e.g. in Gag, medians were 3.7% in the 2000s versus 2.0% in the 1980s). HIV polymorphisms exhibiting the greatest relative spread were those restricted by protective HLAs. Despite these increases, when HIV sequences were analyzed as a whole, their total average burden of polymorphisms that were “pre-adapted” to the average host HLA profile was only ∼2% greater in modern versus historic eras. Furthermore, HLA-associated polymorphisms identified in historic HIV sequences were consistent with those detectable today, with none identified that could explain the few HIV codons where the inferred epidemic ancestor differed from the modern consensus. Results are therefore consistent with slow HIV adaptation to HLA, but at a rate unlikely to yield imminent negative implications for cellular immunity, at least in North America. Intriguingly, temporal changes in protein activity of patient-derived Nef (though not Gag) sequences were observed, suggesting functional implications of population-level HIV evolution on certain viral proteins.
| Upon HIV transmission, many – though not all – immune escape mutations selected in the previous host will revert to the consensus residue. The persistence of certain escape mutations following transmission has led to concerns that these could gradually accumulate in circulating HIV sequences over time, thereby undermining host antiviral immune potential as the epidemic progresses. As certain immune-driven mutations reduce viral fitness, their spread through the population could also have consequences for the average replication capacity and/or protein function of circulating HIV sequences. Here, we characterized HIV sequences, linked to host immunogenetic information, from patients enrolled in historic (1979–1989) and modern (2000–2011) HIV cohorts from four key cities in the North American epidemic. We reconstructed the epidemic's ancestral (founder) HIV sequence and assessed the subsequent extent to which known HIV immune escape mutations have spread in the population. Our data support the gradual spread of many - though not all - immune escape mutations in HIV sequences over time, but to an extent that is unlikely to have major immediate immunologic consequences for the North American epidemic. Notably, in vitro assessments of ancestral and patient-derived HIV sequences suggested functional implications of ongoing HIV evolution for certain viral proteins.
| Escape from Human Leukocyte Antigen (HLA) class I-restricted CD8+ T-lymphocytes (CTL) in Human Immunodeficiency Virus Type 1 (HIV) occurs along mutational pathways that are broadly reproducible based on the HLA alleles expressed by the host [1]–[4]. The opposite phenomenon (that is, reversion of escape mutations to consensus upon HIV transmission to an individual lacking the restricting HLA) is somewhat more variable. While some escape mutations revert relatively rapidly following transmission [5]–[7], others do so more slowly [8], [9]. Yet others (perhaps because they harbor no fitness costs, or such costs are rescued by the presence of compensatory mutations) revert rarely or not at all [10]–[13]. If escape mutations reverted rapidly and consistently, their prevalence in HLA-mismatched persons would remain stably low (or negligible) over time [9]. However, escape mutations persisting upon transmission could gradually spread throughout the population [10], [12], [14]–[16]. Analogous to the negative impact of transmitted drug resistance mutations on treatment efficacy [17], acquisition of “immune escaped” HIV by persons expressing the relevant HLA allele could undermine the ability of their CTL to control infection. As such, the spread of HIV strains harboring escape mutations throughout the population could gradually undermine host antiviral immune potential, and potentially diminish the protective effects of certain HLA alleles, as the epidemic progresses [11], [12], [18].
The extent to which immune escape mutations are accumulating in HIV sequences over time remains incompletely elucidated – a knowledge gap attributable in part to the scarcity of historic data. Nevertheless, some supportive data exist. It has been suggested that CTL epitopes in European HIV sequences are being “lost” over time through mutational escape, in particular via selection by HLA-B alleles, though this study was limited by the modest number of sequences analyzed [19]. Higher HIV polymorphism frequencies have been reported in modern compared to historic South American HIV subtype B and F sequences, though this study was limited by the lack of host HLA characterization [20]. The high (∼75%) frequency of the B*51-associated HIV Reverse Transcriptase (RT) I135X mutation in Japan, a population where B*51 prevalence approaches 20%, is also consistent with escape mutation accumulation [12], though the possibility that the Japanese epidemic was founded by an HIV sequence containing RT-I135X cannot be ruled out. That certain (though not all) escape mutations are capable of spreading in HIV-infected populations has also been demonstrated via mathematical modeling [9]. However, conclusive assessment of the extent to which escape mutants are accumulating in circulation ideally requires large datasets of linked HLA/HIV genotypes from historic and modern eras, combined with ancestral (founder) sequence reconstruction of the studied epidemics.
The potential pathogenic implications of population-level HIV evolution are also of interest. It has been hypothesized that conflicting selection pressures imposed on HIV by HLA-diverse host populations could lead to (relative) viral attenuation over time, while consistent pressures imposed by populations with limited HLA diversity could increase HIV virulence [21]. However, the complex tradeoffs between immune evasion benefits versus fitness costs of escape, and the context-specific nature of these factors with respect to the host genetic milieu, render this a challenging question to address. A recent meta-analysis of HIV clinical prognostic markers (plasma viral load and CD4+ T-cell counts) in cohorts from North America, Europe and Australia suggested that HIV could be increasing in virulence [22], but other reports have been highly conflicting [23]–[30]. Alternatively, pathogenic implications may be investigated, albeit incompletely and indirectly, via assessment of HIV protein function and/or replication capacity of patient-derived viral sequences – though historic data remain scarce. Reductions in replication capacity of recombinant HIV expressing gag-protease sequences from Japanese patients, a population with relatively constrained HLA diversity [12], [31], have been reported since the 1990s [32], while two earlier studies examining replicative fitness of recombinant viruses expressing HIV RT sequences from historic and modern European isolates yielded opposing results [23], [33].
The goals of the present study are to assess the extent to which HLA-associated polymorphisms are accumulating in HIV sequences over time in a large epidemic region comprising an immunogenetically diverse population (North America), and to investigate whether any genotypic changes have been accompanied by functional implications for the virus. To do this, we genotypically and functionally assessed HIV sequences, linked to host HLA information, from 358 historic (1979–1989) and 382 modern (2000–2011) specimens from four key cities in the epidemic (New York [34], [35], Boston [36], [37], San Francisco [34], [38], [39] and Vancouver, Canada [40]–[42]). We performed ancestral phylogenetic reconstructions to infer North America's most recent common ancestor (MRCA) HIV sequence, and we defined HLA-associated polymorphisms based on independent published sources [43]. We focused on Gag and Nef, as these are immunogenic HIV proteins whose sequence variability is substantially influenced by HLA [43] and whose function is susceptible to immune-mediated attenuation [44]–[46].
Overall, we observed an HIV epidemic that is steadily diversifying (in part due to HLA pressures), where background frequencies of HLA-associated polymorphisms have, on average, increased by a modest extent over the study period. Notably, HIV polymorphisms selected by protective HLA alleles appear to have increased to a greater relative (though not absolute) degree than those restricted by non-protective alleles. Despite these increases, average escape mutation background frequencies remain, in absolute terms, low. As such, we contend that HIV adaptation to host HLA is unlikely to yield imminent negative implications for cellular antiviral immunity, at least in North America. Intriguingly, changes in Nef (though not Gag) activity were observed over the epidemic's course, suggesting functional impacts of ongoing HIV evolution on certain viral proteins.
A total of 358 historic HIV sequences spanning 1979–1989, from observational cohorts of men who have sex with men (MSM) established in four key cities in the North American epidemic (New York [34], [35], Boston [36], [37], San Francisco [34], [38], [39] and Vancouver [40]–[42]), were studied alongside 382 modern North American HIV sequences spanning 2000–2011 from untreated persons belonging to various risk groups. High-resolution HLA class I sequence-based typing, aided where necessary by imputation using a published [47] and extensively validated [43] machine-learning algorithm, was successful for 330 (of 358; 92.2%) historic and 381 (of 382; 99.7%) modern specimens. The lower success rate for historic samples reflects the use of serum or plasma as a genomic DNA source [48]. A limitation of serum-based typing is the potential overrepresentation of homozygous types due to amplification of only one allele of the pair [48]; indeed, this bias was noted (e.g.: HLA-B homozygosity was 9% in the historic compared to 5% in the modern cohort, p = 0.03). Nevertheless, historic and modern cohorts exhibited comparable HLA allele frequencies (Pearson's R = 0.97, p<0.0001, and Figure S1), indicating that our analyses of the spread of HLA-associated HIV polymorphisms are unlikely to be majorly confounded by intercohort differences in the frequencies of their restricting HLA alleles.
Plasma HIV RNA amplification and bulk sequencing of Gag and/or Nef was successful for the above-mentioned 358 historic specimens (of an original total of 497 specimens tested, 72.0% genotyping success rate), yielding 299 Gag and 335 Nef sequences for study. Success rates of historic Gag and/or Nef genotyping, by site, were: New York 73 (of 94; 77.6%), San Francisco 32 (of 75; 42.7%), Boston 242 (of 282; 85.8%) and Vancouver 11 (of 46, 23.9%). Infection stage was unknown for most historic specimens, though these included 67 individuals with known or suspected early infection, all from New York. Gag and/or Nef sequencing was successful for 382 modern specimens in total: 358 (93.7%) for Gag and 337 (88.2%) for Nef, all from individuals with chronic infection. All HIV sequences were subtype B.
Estimated maximum-likelihood HIV Gag and Nef phylogenies exhibited star-like shapes typical of HIV sequences sampled from a population [49] (Figure 1). Despite being a convenience sample, historic sequences exhibited no gross segregation by early (1979–1982; N = 28), mid (1983–1985; N = 122) and later (1986–1989; N = 208) eras. Moreover, unique historic North American sequences in the Los Alamos National Laboratory (LANL) HIV database (totaling 27 Gag and 56 Nef sequences spanning 1982–1989) were interspersed throughout the phylogenies, as were sampled modern LANL sequences spanning 2000-present (Figure 1). Despite some clustering by city and the predominance of historic sequences in two lineages of a combined phylogeny (Figure S2), the historic and modern cohort consensus HIV sequences were consistent with one another as well as the LANL North American and global (worldwide) subtype B consensus sequences (Figure S3), with all differences occurring at highly variable residues. Results thus support our HIV sequences as not grossly unrepresentative of the North American epidemic.
HIV sequence diversity within the modern cohort was substantially greater than that of the historic cohort (Figure 1). Grouped by era, the mean (±standard deviation [SD]) patristic (pairwise) genetic distances in Gag were 0.020±0.004 (1979–1982), 0.027±0.009 (1983–1985), 0.034±0.009 (1986–1989), and 0.074±0.012 (2000+) substitutions per nucleotide site, while those for Nef were 0.043±0.010, (1979–1982), 0.057±0.014 (1983–1985), 0.072±0.015 (1986–1989), and 0.12±0.025 (2000+) substitutions per nucleotide site. Modern HIV cohort sequences (all sampled during chronic infection) exhibited comparable mean pairwise distances to modern acute-phase subtype B sequences not included in the previous analysis (not shown), suggesting that infection stage was not a major confounder of our diversity estimates by era. Taken together, results support a diversifying North American epidemic [50] where average intra-subtype Gag and Nef genetic distances have increased approximately two-fold since the 1980s.
Before claiming that any highly prevalent HIV polymorphism has arisen as a result of its spread through the population over time, it is important to rule out its presence at the epidemic's genesis (i.e. founder effect [51]). We therefore estimated the founder virus sequence of the North American epidemic by reconstructing the most recent common ancestor (MRCA) sequence at the root of the Gag and Nef phylogenies. To this end, we performed ≥50,000 MRCA reconstructions per HIV protein on random subsets of the historic sequence data using BEAST (see methods and [52]), and computed a “grand consensus” MRCA reconstruction per protein (Figure 2). Overall, reconstruction confidence exceeded 80% for all but one codon in Gag (residue 67) and for all but 6 codons in Nef (residues 15, 21, 51, 152, 178 and 205), all of which are highly polymorphic sites (<70% amino acid conservation) (Figure 2). The consensus of Gag sequence reconstructions at the MRCA differed from the LANL North American HIV subtype B consensus at only four residues (A67S, R76K, K91R and E102D), while the consensus of Nef MRCA reconstructions was identical to it (Figures 2 and S3). Note the four ancestor/consensus differences in Gag merit cautious interpretation, as codon 67 was reconstructed with <80% confidence and the remainder are sites with <60% conservation at the amino acid level. MRCA reconstructions undertaken using random subsamples of both historic and modern Gag and Nef sequences were consistent with those computed from historic sequences only (not shown). Finally, the grand mean MRCA date estimate from phylogenetic reconstructions inferred from random subsamples of both historic and modern sequences was 1965 (range 1962–1967). The consistency of this date with published estimates of a 1960s U.S. epidemic origin [53]–[55] provides additional support for our data as representative of the North American epidemic.
A diversifying epidemic will, by definition, feature increasing viral polymorphism frequencies. Thus, to give relevance to our objective of measuring the spread of HLA-driven polymorphisms in HIV sequences over time, it is important to first demonstrate that HIV diversification is driven, at least in part, by HLA pressures. If so, we reasoned that HIV codons known to be under selection by HLA would, on average, have diversified to a greater extent than those not under selection by HLA.
To investigate this, we first needed to independently define a list of HIV sites that are known to be under selection by specific HLA alleles. We defined these based on an independent published study of >1800 treatment-naïve individuals with chronic HIV subtype B infection from cohorts in Canada, the USA and Australia [43], that had no overlap with the historic or modern cohorts studied here. In that study, HLA-associated polymorphisms in HIV were identified using phylogenetically-corrected association testing approaches (see methods and [43]). For the present analysis of HLA selection and HIV diversification, an inclusive definition of “HIV sites under selection by HLA” was warranted; therefore, we defined this as all Gag and Nef codons associated with at least one HLA allele that met a false-discovery rate threshold of <20% (q-value <0.2) in the independent study (see methods and [43]). This totaled 95 (of 500) codons in Gag and 99 (of 206) codons in Nef.
We began with Gag, by aligning historic and modern amino acid sequences to the HIV reference strain HXB2 and computing changes in Shannon Entropy on a per-codon basis (1000 bootstraps). This revealed 69 (of 500; 14%) codons whose entropies were significantly higher (p<0.001, q<0.01) in modern versus historic sequences (Figures 3A, 3B). To minimize circularity of arguments, we next excluded highly (>99%) conserved codons from consideration, as these cannot diversify to any great extent (and as such, are rarely identified as HLA-associated [43]) – leaving 219 “variable” Gag codons for analysis. Stratifying these sites by their HLA status indicated that, of the 95 Gag sites under selection by HLA [43], 45.2% exhibited significantly higher entropy in modern versus historic sequences, compared to 21.0% of the 124 sites not associated with HLA (p = 0.0002, Figure 3C). This indicates that HLA-associated viral sites tend to be those that have diversified the most between historic and modern-era HIV sequences.
While entropy approaches strictly investigate the end products of diversification, dN/dS-based approaches provide a more direct way to investigate elevated substitution rates within the phylogeny. As such, we identified sites under significant pervasive positive (diversifying) selection in a maximum-likelihood phylogeny comprising historic and modern sequences using the fast unconstrained Bayesian approximation for inferring selection algorithm [56]. As expected, after excluding codons that were >99% conserved, sites under pervasive positive selection were more likely to experience a significant increase in entropy (p<1×10−5, not shown) (indicating that positive selection is driving some of this diversification), and were more likely to be HLA-associated (suggesting that HLA represents a major source of this selection pressure) (p = 0.0022, Figure 3D).
We repeated these analyses for Nef, revealing trends consistent with those observed for Gag (Figure S4). Results thus suggest that ongoing HIV diversification is attributable, at least in part, to HLA pressures.
We now turn to our major goal of assessing the spread of HLA-associated polymorphisms in the population over time. If escape mutations in HIV are reproducibly selected in individuals expressing particular host HLA(s), but such mutations consistently and rapidly reverted upon transmission, then we would expect their frequencies to be generally higher among individuals expressing the relevant HLA(s), and generally low among individuals lacking them, at levels that remain stable over time. But, if HLA-associated polymorphisms were to persist upon transmission and gradually spread in the population, we would expect polymorphism frequencies among HLA-matched individuals to remain stably higher, but polymorphism frequencies among individuals lacking the restricting HLA(s) to increase over time. As such, we stratified our HLA-associated polymorphism frequency comparisons between epidemic eras with respect to persons expressing, versus not expressing, the relevant HLA(s).
As before, we defined HLA-associated polymorphisms according to an independent source [43]. Because the present analysis investigated individual viral polymorphisms (rather than just sites) associated with HLA, a more specific definition was warranted. As such, we investigated all HLA-associated “adapted” (escape mutant) forms meeting a false-discovery rate threshold of <5% in the original study (see [43] and methods). This list comprised specific HLA-associated polymorphisms occurring at 71 Gag and 96 Nef codons [43]. HLA-associated polymorphisms in HIV were additionally stratified based on whether they represented consensus or non-consensus viral residues. Though the vast majority of HLA-associated polymorphisms represent non-consensus residues, a minority represent cases where an HLA allele is associated with preservation of the consensus residue at a given site (e.g. HLA-B*07:02 is associated with preservation of consensus G357 in Gag) [43]. We analyzed such cases separately because, under conditions of star-like diversification of a “consensus-like” founder, the null expectation is for polymorphism (i.e. non-consensus) frequencies to increase, and consensus frequencies to decrease, over time. Separating them also allows more intuitive interpretation when polymorphism frequencies are summarized as averages.
We began by investigating the frequencies of 70 non-consensus HLA-associated polymorphisms, occurring at 60 codons in Gag, between HLA-expressing and non-expressing persons in the historic and modern cohorts (Figure 4). As expected, individual polymorphism frequencies varied widely, but they were nevertheless enriched among individuals expressing the relevant HLA(s) (Figure 4A) compared to individuals lacking them (Figure 4B). In accordance with the null expectation, polymorphism frequencies in persons harboring the relevant HLA(s) were consistent across historic (median 18%, Interquartile Range [IQR] 4–54%) and modern (median 23% [IQR 7–45%]) cohorts (p = 0.8; Figure 4A). For example, Gag-242N frequency was ≥70% among persons expressing a B58 supertype allele, regardless of era. In persons lacking the relevant HLA(s), we also observed numerous examples of polymorphisms whose frequencies remained stable between historic and modern eras (e.g. Gag-242N frequency remained <1% in persons lacking a B58 supertype allele) (Figure 4B). Overall though, the average frequencies of these polymorphisms in persons lacking the relevant HLA(s) were modestly, yet statistically significantly, higher in modern (median 3.7% [IQR 2–19%]) compared to historic (median 2.0% [IQR 0.7–10%]) sequences (p = 0.0002; Figure 4B), a result consistent with the spread of many – though not all – HLA-driven polymorphisms in the population. Results remained significant after adjusting for minor inter-cohort differences in HLA frequencies (as these influence rates of polymorphism transmission) (p = 0.001, Wilcoxon one-sample test, not shown).
Under conditions where HLA-associated polymorphisms are, on average, slowly spreading through the population, we would expect the statistical associations between HIV polymorphisms and their restricting HLA(s) to concomitantly weaken. Indeed, this appeared to be the case. The median odds ratios of association between HIV Gag polymorphisms and their restricting HLA(s) were modestly lower in modern (median OR 3.1 [IQR 1.7–7.1]) compared to historic (median OR 3.8 [IQR 1.2–17.5]) cohorts (p = 0.009, Figure 4C).
Similar trends were observed for the 89 non-consensus HLA-associated polymorphisms occurring at 77 codons in Nef. Among persons expressing the relevant HLA(s), Nef polymorphism frequencies remained consistently elevated in historic (median 14% [IQR 3–50%]) and modern (median 15% [IQR 3–41%]) cohorts (p = 0.7; Figure 4D). In persons lacking the relevant HLA(s), examples of polymorphisms whose frequencies remained stable across historic and modern cohorts were noted (e.g. Nef-94E frequency remained ∼1% in persons lacking B*08, while Nef-135F remained ∼10% in persons lacking A*23:01 and A*24) (Figure 4E). Overall though, the average frequencies of these polymorphisms in persons lacking the relevant HLA(s) were modestly higher in modern (median 3.4% [IQR 1–12%]) compared to historic (median 2.0% [IQR 0.6–11%]) sequences, though this did not reach statistical significance (p = 0.054) (Figure 4E). Median odds ratios of association between Nef polymorphisms and their restricting HLA(s) were also slightly lower in modern (median 3.1 [IQR 1.7–7.1]) compared to historic (median 3.8 [IQR 1.2–17.5]) cohorts, though not significantly so (p = 0.065, Figure 4F).
We also investigated HLA-associated polymorphisms occurring at 11 Gag and 19 Nef codons where the association represented the consensus residue [43]. As expected, we observed higher frequencies of these consensus residues in individuals restricting the relevant HLA(s) compared to individuals lacking them (Figure S5). We also observed trends, though not statistically significant, towards lower consensus frequencies at these sites in modern versus historic sequences, regardless of HLA alleles expressed (Figure S5).
Taken together, our results are consistent with a scenario in which, on average, non-consensus HLA-associated polymorphisms have increased in frequency in North American HIV sequences over time. That said, the observed increases for Nef were not statistically significant, and both proteins harbored numerous examples of HLA-driven polymorphisms with stable background prevalence (e.g. Gag-242N, Nef-94E, Nef-135F). Moreover, although results for Gag attained statistical significance, average polymorphism background frequencies remained notably low, regardless of era. Our results thus indicate that not all HLA-driven polymorphisms are accumulating in circulation. Rather, our results suggest a diversity in accumulation rates, with the majority of nonconsensus polymorphisms spreading slowly (and others not at all) – and consensus residues decreasing in frequency overall. These observations confirm slow polymorphism spread predicted by mathematical models [9] and are consistent with an epidemic that is gradually diversifying under selection pressures that include HLA.
Our results suggest that, on average, HLA-associated polymorphisms are spreading in the population, albeit slowly. From an immunological perspective, an increasing burden of escape mutations in circulating HIV strains over time could yield a reduction in the ability of individuals to control the virus via cellular responses as the epidemic progresses. We thus asked: if an individual were to be randomly infected by an HIV sequence from the historic or modern eras, to what extent would the latter contain a higher burden of polymorphisms that are “pre-adapted” to their HLA? To estimate this quantity, we compared each individual's HLA profile against all historic and modern chronic-phase HIV sequences in our dataset, and calculated the percentage of HLA-associated sites in each sequence exhibiting the adapted form specific to each person's total HLA profile. Comparison of the overall per-person averages thus represents the expected extent to which a randomly sampled HIV sequence would be pre-adapted to a given individual, had they been infected by a sequence from that era. Focusing first on non-consensus HLA-associated polymorphisms, our calculations for Gag yielded a median “percentage HIV sites pre-adapted to one's HLA profile” of 14.9% [IQR 10.1–19.5%] for historic versus a median of 17% [IQR 12.7–22.4%] for modern sequences, an average increase of only ∼2% (Figure S6). Inclusion of consensus HLA-associated polymorphisms further minimized this gap (not shown). For Nef, the median “percentage of adapted sites” remained consistent across eras (19.0% in historic versus 18.5% in modern) (Figure S6); moreover, inclusion of consensus polymorphisms resulted in lower overall percentages in modern compared to historic sequences (not shown). Results therefore suggest that, despite HIV diversification, an individual's overall expected risk of acquiring escape mutant viruses specific to their HLA allele profile has increased only minimally for Gag, and not at all for Nef, since the 1980s in North America.
Broadly speaking, at any given point in time, the average background frequencies of HLA-associated polymorphisms in circulating HIV sequences will generally positively correlate with the frequencies of their restricting HLA alleles in the population [12]. This is because higher absolute numbers of persons expressing the HLA will generally translate to higher absolute numbers of polymorphisms selected and thus transmitted (though many factors, including the wide-ranging probabilities of polymorphism selection given their location and restricting HLA, the fact that multiple HLA alleles select the same – or opposing – mutations at a given location, the existence of “consensus” HLA-associations, and the timing of polymorphism selection/reversion, will render this correlation less than perfect). Nevertheless, such a positive trend is observed in both the historic and modern cohorts, as expected (Figure S7). However, we are specifically interested in investigating the extent to which HLA-associated polymorphisms are spreading through the population over time. We thus asked: are polymorphisms restricted by certain HLA alleles increasing to a greater extent than others?
To do this, we analyzed all HLA allele groups for which a minimum of three HLA-associated polymorphisms (regardless of whether they were consensus or non-consensus) were studied (25 alleles total). For each HLA-associated polymorphism, we computed its fold-increase in background frequency over time (for example, a hypothetical polymorphism with a background frequency of 1% in the historic cohort versus 2% in the modern cohort would equate to a two-fold increase). For each HLA allele we then calculated the median fold-increase in frequency of all polymorphisms restricted by it. Overall, we observed no significant correlation between the frequency of a restricting HLA allele and the relative extent to which its polymorphisms spread throughout the population between historic and modern cohorts (Spearman's R = −0.35, p = 0.09) (Figure 5A). Taken together with the results in Figure S7, this indicates that, at any given point in time, polymorphisms restricted by common HLA alleles will generally be found at higher absolute frequencies in a population than those restricted by rarer ones, but such polymorphisms do not appear to be spreading in the population to a greater relative extent (i.e. when expressed in terms of fold-change) over time.
Strong epidemiological links between host carriage of specific HLA class I alleles and HIV disease progression have been demonstrated in natural history studies (e.g.: [57]), with some alleles, notably HLA-B*57 and HLA-B*27, consistently associated with slower progression [57]–[59]. We therefore wished to investigate the relationship between an HLA allele's “protective” status (defined as its published Hazard Ratio for progression to AIDS [57]) and its median fold-increase in polymorphism background frequency between historic and modern eras. Of interest, we observed a significant inverse correlation between these two parameters (Spearman's R = −0.52, p = 0.0076) (Figure 5B), suggesting that polymorphisms restricted by protective HLA alleles have, in relative (fold-change) terms, spread to a greater extent in the population than those restricted by non-protective HLA alleles.
It is nevertheless important to contextualize these results in absolute terms. Of the six HLA-B*57-associated sites studied in Gag, historic sequences harbored a median 0 [IQR 0–1] B*57-associated polymorphisms at these sites, compared to 1 [IQR 0–2] in modern Gag sequences. Of the six B*57-associated sites in Nef (two of which represent “consensus” associations), both historic and modern sequences harbored a median of 2 [IQR 1–3] B*57-associated adapted polymorphisms. It thus remains unclear to what extent these modest absolute increases may compromise the protective effects of certain HLA alleles as the epidemic progresses.
We have thus far defined HLA-associated polymorphisms as those identified in independent modern cohorts by statistical association [43]. To investigate the potential existence of novel historic HLA-associated polymorphisms that are no longer detectable in modern sequences due to their spread throughout the population, we applied association testing approaches to our historic dataset directly. Historic patients with known or suspected early infection were excluded (as these could dilute associations between HLA and HIV polymorphisms due to insufficient within-host evolution), and a false-discovery rate (q-value) cutoff of 0.05 was employed. We were especially interested to see whether HIV codons whose inferred ancestral (founder) amino acid differed from the North American consensus (there were 4 in Gag) or were reconstructed with <80% confidence (1 in Gag and 6 in Nef) could be explained by the existence of historic HLA-associated polymorphisms at these sites. However, no such evidence was observed (Figure 6A, 6B). Instead, analysis revealed 16 HLA-associated polymorphisms occurring at 10 Gag codons and 28 HLA-associated polymorphisms occurring at 13 Nef codons that, with the exception of an association between B*49:01 and the consensus G at Gag codon 62, were wholly consistent with published escape pathways [43] and/or were confirmed in the present modern cohort (not shown). In summary, the strongest HLA-associated polymorphisms in historic sequences are consistent with those identifiable today.
HIV Gag and Nef are highly immunogenic HIV proteins whose sequence variability is substantially influenced by HLA [43] and whose function is susceptible to immune-mediated attenuation [44]–[46]. As such, we investigated whether the gradual spread of immune escape mutations in North American Gag and Nef sequences may be accompanied by overall changes in the average viral replication capacity and/or protein function of patient-derived HIV sequences. We began with Gag, by generating a recombinant HIV strain expressing the epidemic's inferred Gag ancestral sequence, and another expressing the published global subtype B consensus (Figure S3) in an HIV NL4-3 subtype B reference strain backbone. We also generated recombinant HIV NL4-3 strains expressing a single representative clonal Gag sequence from 108 (of 120 originally selected; 90.0% success rate) historic and 58 (of 71 originally selected; 82% success rate) modern specimens (Figure 7A). A clonal (rather than quasispecies [60]) approach was adopted for the patient-derived sequences, as variations in viral stock diversity resulting from differential integrity of historic versus modern specimens could bias replicative measurements. We assayed the in vitro replication capacity of these recombinant viruses using a published reporter T-cell assay [60]–[63]. Replication capacities (RC) were normalized to that of parental NL4-3, such that values >1 and <1 indicate RC greater or less than NL4-3, respectively.
The replication capacities of recombinant viruses encoding the inferred ancestral and global subtype B consensus sequences were comparable to those of parental NL4-3 (Figures 7B and S8). Recombinant viruses expressing historic or modern Gag clonal sequences displayed a broad range of growth phenotypes, with median RCs approaching that of NL4-3 (Figure 7B). Although there appeared to be a trend towards lower RC among Gag recombinant viruses from early historic (1979–1982) patients, this was not statistically significant (Kruskal-Wallis p = 0.6). Furthermore, no correlation was observed between the replication capacity of a given Gag clone and its genetic distance from the Gag NL4-3 sequence (Spearman's R = 0.03, p = 0.6, not shown), arguing against confounding effects attributable to our use of a historic lab-adapted sequence (NL4-3) as a viral backbone.
Similarly, we cloned the inferred ancestral, global subtype B consensus and a single representative Nef sequence from N = 102 historic and N = 86 modern patients into a GFP-expression vector (Figures 8A and S8). As modulation of Nef function over the natural history of infection is supported by some [64], [65] (though not all [66]) studies, and a minority of historic Nef clones were derived from persons with known or suspected early infection, we indirectly assessed infection stage as a potential confounder by including Nef sequences from 52 modern chronic and 34 early infection patients not included in previous analyses (sampled a median of 72 [IQR 48–92] days after infection) in our comparison group. Following transient transfection into an immortalized T-cell line stably expressing CD4 and HLA-A*02, we assessed the ability of these Nef clones to downregulate these molecules from the cell surface by flow cytometry [67], [68] (Figure 8B). The Nef sequence from HIV reference strain SF2 served as a positive control (SF2 is commonly used as a control in Nef functional studies, as it possesses robust CD4 and HLA class I downregulation activities, e.g. [67]); thus, normalized Nef functions of >1 and <1 indicate activity greater or less than SF2, respectively. Nef protein expression was verified by Western blot (Figure S8); 15 poorly functional Nef clones whose expression could not be detected were excluded (since in vitro cloning defects or other artifacts could not be ruled out), leaving 93 historic and 80 modern clones for analysis.
CD4 downregulation activity of ancestral Nef was comparable to that of reference strain SF2 (Figure 8B), while that of global subtype B consensus Nef was ∼3% lower (not shown). Nef clones from historic and modern patients were generally highly functional for CD4 downregulation and exhibited relatively narrow dynamic ranges. Nevertheless, historic patient-derived Nef sequences exhibited significantly lower CD4 downregulation abilities compared to modern sequences (Kruskal-Wallis p<0.0001), with the early (1979–1982) Nef clones exhibiting the lowest function overall (Figure 8B). Nef-mediated CD4 downregulation of modern Nef clones from individuals in early and chronic infection were comparable (p = 0.9, Figure 8B and not shown), arguing against infection stage as a major confounder of this result.
The ability of the ancestral Nef sequence to downregulate HLA-A*02 was ∼3.5% higher than reference strain SF2 (Figure 8C), while that of global subtype B consensus Nef was equivalent to SF2 (not shown). Although Nef clones from both historic and modern patients were in general highly functional, historic Nef sequences exhibited significantly lower HLA downregulation abilities compared to modern Nef sequences (Kruskal-Wallis p<0.0001), with the early (1979–1982) Nef clones displaying the lowest function overall (Figure 8C). HLA downregulation capacities of modern early Nef sequences were on average 1% higher than those from modern chronic Nef sequences (p = 0.14, Figure 8C and not shown), arguing against infection stage as a major confounder. The significantly lower Nef-mediated CD4 and HLA downregulation observed in historic versus modern sequences was robust to inclusion/exclusion of the 15 clones whose Nef expression was not detectable by Western Blot (not shown).
Taken together, the lack of significant functional differences between ancestral, subtype B consensus, and median patient-derived Gag clones from historic and modern eras argues against major replicative consequences of HIV Gag diversification during the North American epidemic. In contrast, our Nef results suggest the introduction of a highly functional founder virus to North America in the 1960s, followed by a subsequent decline in average Nef-mediated CD4 and HLA downregulation functions of patient-derived sequences in the 1980s, that were restored to original (“founder”) levels by the 2000s. The mechanisms and potential role for host pressures in this phenomenon require further investigation.
The present study examined linked host (HLA) and HIV (Gag/Nef) datasets from historic (1979–1989) and modern (2000–2011) eras in North America to estimate the extent to which HLA-driven polymorphisms may be spreading throughout circulating HIV sequences over time on this continent. Phylogenies inferred from historic and modern samples of HIV Gag and Nef sequence variation were star-like in shape, yielding a reconstructed ancestral (epidemic founder) virus sequence that was essentially identical to North American subtype B consensus. Mean pairwise distances between modern HIV Gag and Nef sequences were approximately two-fold greater than those between historic sequences, supporting a diversifying epidemic. Notably, Gag and Nef codons exhibiting the most significant entropy increases over time were enriched for known HLA-associated sites, consistent with a key role of HLA in driving HIV diversification [69], [70].
Also consistent with an approximate two-fold increase in HIV diversity since the mid-1980s in North America, the average “background” frequencies of HLA-associated polymorphisms (i.e. in individuals lacking the restricting HLA) were roughly two-fold higher in modern compared to historic sequences. These differences reached statistical significance for Gag, though not for Nef. As expected, in both historic and modern cohorts, a general positive correlation was observed between the frequency of an HLA allele and the background frequency of its associated polymorphism in the general population. However, the polymorphisms that, over time, appeared to be spreading to the greatest relative extent (in terms of fold-change) were not those restricted by common HLA alleles (Figure 5A) but rather those restricted by protective HLA alleles [57] (Figure 5B). This observation, along with our lack of identification of novel historic HLA-associated polymorphisms restricted by common HLA alleles, indicates that HIV is not simply adapting to the most frequent HLA alleles in a given host population. Instead, our findings are consistent with protective HLA alleles as those imposing the strongest evolutionary pressures on HIV, an observation that is consistent with previous reports that protective HLA alleles are more likely to induce strong selection at key conserved sites [43], [71]–[73].
The spread of HLA-associated polymorphisms in circulation could lead to a reduction in host antiviral immune potential over time [12]. We thus wished to interpret our results in terms of the imminence of this potential outcome. First and notably, the extent of HLA-driven polymorphism accumulation in Nef did not reach statistical significance. Second, though observations for Gag did achieve significance, average polymorphism background frequencies remained low in absolute terms (i.e. 2.0% in the 1980s versus 3.7% in the 2000s) – differences that, when expressed in terms of the average estimated extent to which circulating HIV Gag sequences are “pre-adapted” to an individual's HLA profile, translated into an overall increase of only ∼2% between historic and modern eras. Moreover, we observed numerous HLA-associated polymorphisms whose prevalence remained stable in the population (e.g. B58-supertype-associated Gag-242N, B*08-associated Nef-94E, A*2301/A*24–associated Nef-135F), observations that are consistent with their rapid reversion upon transmission [5], [9], [74] (though estimates of the reversion rate for B*08-Nef-94E are somewhat conflicting [9], [74]). That some - though certainly not all - HLA-driven escape mutations are capable of spreading through the population has been demonstrated via mathematical modeling [9], indicating that the reproducible selection of specific escape mutations in persons harboring the relevant HLA does not always translate into rapid evolution at the population level [9]. That certain HIV sites simultaneously display strong signals for diversifying selection, yet stable polymorphism prevalence, is also consistent with “toggling” between consensus and escape forms [75] as HIV disseminates in a genetically diverse host population.
Although our study did not formally attempt to model the dynamics of HLA-driven polymorphism spread in the North American population, our observations suggest that this is happening slowly. Very gradual polymorphism spread is also consistent with mathematical models projecting that, even in the case where an escape mutation never reverts, it could take centuries for it to reach fixation following its initial appearance in the population [9]. Moreover, it has been projected that any reversion (however slow) would prevent a polymorphism from ever becoming fixed [9]. Also consistent with slow spread is the near-identity of the reconstructed epidemic MRCA (founder) HIV sequence to the North American consensus - which suggests that, between the North American epidemic's genesis and the present day, no polymorphism, HLA-driven or otherwise, has spread to an extent where it now outcompetes that of the original founder residue. Our lack of identification of novel historic HLA-associated polymorphisms at the seven Gag/Nef codons where the inferred ancestor was reconstructed with <80% confidence and the four (highly variable) Gag codons where it differed from the modern consensus also argues against the spread of any historic HIV escape mutation in North America to the point where it now defines consensus. Note however that some caution is merited when interpreting the estimated founder viral sequence, since rapid selective sweeps occurring between the epidemic's foundation [54], [55] and the earliest 1979 sampling date would not have been detected and therefore cannot be ruled out. Acknowledging these caveats, the near-identity between the estimated North American founder virus and modern consensus additionally suggests that statistical associations between particular HLA alleles and the HIV consensus residue at a given site (e.g. B*07:02 with Gag-G357) have not arisen as a result of their selection and subsequent spread in the population to the point where they define the consensus [10]. Rather, these residues were most likely present at the epidemic's foundation - and, if anything, are gradually decreasing in frequency as HIV continues to diversify. We propose that such “consensus HLA associations” represent cases where the founder virus happened to be adapted to certain HLAs (perhaps because the original founder or earlier hosts expressed them), and that these HLAs continue to exert purifying selection on these sites over time.
Despite inferred overall slow rates of accumulation, the observation that polymorphisms restricted by protective alleles appear to be spreading to a greater (relative) extent than others is potentially important. Indeed, the stabilization of certain protective allele-associated escape mutations by secondary (compensatory) substitutions has been documented: the S173A mutation (which allows the B*27-associated Gag-R264K mutation to persist upon transmission in an HIV subtype B context [11], [13]) and the S165N mutation (which stabilizes B*57-associated mutations within the p24Gag KF11 epitope in a subtype C context [8]), are examples.
Despite this, we urge caution in extrapolating that the protective effects of HLA alleles will diminish rapidly in North America. Again, it is important to consider that absolute polymorphism background frequencies remain low: modern Gag and Nef sequences together harbor, on average, only one additional B*57-associated polymorphism compared to historic sequences. Similarly, despite polymorphism spread, a B*27-expressing individual still has a >90% chance of acquiring HIV with the immunologically susceptible consensus R at critical Gag codon 264. Besides, the protective effects of most such alleles (including, to a certain extent, B*27 [76]) are attributable to consistent and strong CTL responses against multiple HIV epitopes [43], [77], [78]. It is also important to consider that protective HLA-restricted CTL retain activity against polymorphic variants in many cases [79], [80], and de novo [81] or cross-reactive [82] CTL responses to in vivo escape variants can, and do, arise. Further integrated evolutionary and molecular studies are therefore required to assess the potential immunologic impact of polymorphism spread on HIV control by protective HLA alleles.
Our study also investigated whether HIV evolution in North America has been accompanied by changes in viral replication capacity or protein function. Consistent with previous in vitro assessments of HIV sequences reconstructed using Center-of-Tree approaches [83], our inferred Gag and Nef ancestral sequences were highly functional. Despite substantial increases in Gag diversity over time, the average replication capacities of recombinant NL4-3 viruses expressing patient-derived clonal Gag sequences from historic and modern eras were comparable to that of NL4-3 expressing the inferred Gag ancestral sequence, arguing against major replicative consequences of HIV Gag diversification during the North American epidemic. These results contrast with reductions in replication capacity of recombinant viruses expressing patient-derived Gag-protease sequences from Japanese patients from the mid-1990s to present [32], a difference possibly due to the greater homogeneity of HLA alleles in Japanese compared to North American populations, that may exert consistent selection pressures driving the selection of fitness-reducing mutations.
In contrast, the average Nef-mediated CD4 and HLA downregulation activities of historic patient-derived sequences were modestly yet significantly lower than modern ones. This is intriguing since the inferred Nef ancestral sequence displayed high function. We therefore speculate that, following the introduction of a functional ancestral Nef sequence into North America, initial HIV adaptation to this new population led to decreases in Nef function that were subsequently rescued upon continued Nef diversification. The higher Nef-mediated HLA class I downregulation function of modern compared to historic sequences, combined with the observation of modest HLA-driven polymorphism spread through the population during this same period, raises the interesting possibility that, compared to viruses circulating in the 1980s, modern North American HIV sequences may exhibit greater immune evasion potential via enhanced HLA class I downregulation [84] function. However, further studies will be required to elucidate the underlying mechanisms and pathogenic implications of these observations.
An anticipated criticism is our definition of HLA-associated polymorphisms by statistical association studies of modern cohorts [43]. This approach could underestimate the average extent of polymorphism spread over time, for two reasons. First, such lists could exclude historic escape mutations that are no longer detectable in modern cohorts due to polymorphism spread. To address this we applied statistical association approaches to identify HLA-associated polymorphisms detectable at the population level in the historic cohort. However, all identified polymorphisms save one were consistent with known HLA-associated escape pathways, indicating that the strongest mutations detectable historically remain readily detectable today. A second limitation is that association testing approaches, even those that incorporate phylogenetic correction (as ours do), systematically favor the identification of HLA-associated mutations that escape and revert rapidly [85], which by definition would not be expected to spread quickly in a population [9]. However, this limitation is somewhat offset by the substantial size (N>1800) of the cohort used to define HLA associations. Mathematical models indicate that at such sample sizes, with phylogenetic correction, significant associations can be detected between HLA alleles and polymorphisms even if these escape and/or revert on a timescale of decades [85]. Moreover we have previously demonstrated that cohorts of this size are powered to detect very rare HLA-associated polymorphisms, as well as those that are nearly universally observed in the population [43].
This study possesses additional limitations, many inherent to convenience sampling and technical challenges of working with historic samples. Although our sequences date back to 1979, the lack of data from the critical period between HIV's introduction into North America and the late 1970s is a major limitation of this and all other studies undertaken to date. Nevertheless, our historic HIV sequence dataset is 10-fold (Gag) and 7-fold (Nef) larger than existing data from this era and region, and includes the oldest North American sequences ever published. Another limitation is that specimens were obtained from only four sites in North America, and all historic specimens were derived from observational studies of individuals from a single risk group (MSM) [34], [36]–[42]. As such, our HIV diversity estimates, particularly for the historic era, may represent underestimates. Nevertheless, the dispersion of published North American HIV sequences throughout all phylogenies, the consistency of historic and modern consensus sequences, and our estimated epidemic founder dates that are compatible with published estimates [53]–[55] suggest that our sequences are not grossly unrepresentative of the North American epidemic. Concerns regarding our ability to faithfully amplify the original quasispecies diversity from historic specimens by PCR led us to adopt a single representative clone (rather than bulk) approach for our functional assessments of Gag and Nef in order to minimize in vitro bias associated with differences in the diversity of viral stocks. The presence of individuals with known or presumed early infection in our historic cohort and the general lack of clinical staging information are also limitations. To reduce confounding, early sequences were excluded from relevant analyses (e.g. identification of HLA-associated polymorphisms in the historic cohort and calculation of Odds Ratios of association between HLA and polymorphisms), while other analyses verified the appropriateness of pooling data by comparing early and chronic sequences directly to rule out differences between them (e.g. Nef functional assessments). The absence of pVL and CD4 information on historic patients also precluded the investigation of trends in disease markers over time.
On the other hand, our development of a sensitive HLA sequence-based typing assay capable of utilizing genomic DNA extracted from plasma/serum [48] allowed us to perform HLA typing of historic specimens, yielding, for the first time, the ability to directly investigate HLA-associated selection pressures over the course of an epidemic. A known limitation of serum-based HLA typing is the overrepresentation of homozygous types due to amplification bias [48], an effect that was noted in our historic dataset. Though this could lead us to overestimate the historic background frequencies of HLA-associated polymorphisms by erroneously including individuals expressing the relevant HLA into our calculations, the low average background frequencies of HLA-associated polymorphisms in modern sequences indicate that any overestimations would not substantially impact our overall conclusions. A notable strength is the lack of overlap between study cohorts and those from which the reference list of HLA-associated polymorphisms was derived [43], thus ensuring independence of source and query data.
In conclusion, HLA-associated polymorphisms are, on average, slowly spreading throughout North American HIV sequences as the epidemic continues to diversify. This slow adaptation to host cellular immune responses parallels the observed drift of HIV towards a more neutralization-resistant phenotype as a result of population-level viral adaptation to humoral immune pressures [86], [87]. However, the absolute frequencies of these polymorphisms in circulation remain on average low on this continent, as do the estimated risks of acquiring HIV “pre-adapted” to one's HLA profile. As such, our results are unlikely to translate into major imminent consequences to CTL-mediated control of HIV, at least in the North American region.
That said, we acknowledge that even modest changes can have biological implications. Indeed, one could contend that modest increases in the frequency of “pre-adapted” HIV strains are not inconsistent with reports suggesting increased HIV virulence over time [22]. Furthermore, it is important to emphasize that the potential rates, and thus immunologic implications, of HLA-associated polymorphism spread may be substantially greater in populations where HLA diversity is far lower and/or HIV prevalence far higher than North America. Rates and implications of polymorphism spread may also be more profound in populations where transmission tends to occur later in infection, thereby increasing the probability of transmitted escape mutations (though mathematical models have suggested that realistic differential transmission rates between acute and chronic infection would impact population escape mutation prevalence only minimally [9]). As such, we recommend that similar analyses of virus-host adaptation be undertaken to assess the rate of accumulation of immune-driven polymorphisms, and its pathogenic implications, in other epidemic regions where historic specimens are available. In conclusion, though our results remain somewhat open to interpretation, we suggest that they be considered in light of the major advances in HIV treatment and prevention [88]–[92] that have occurred during the timecourse of the present study. Combined with current efforts in prevention and cure research [93]–[95], these advances give us firm hope that the end of HIV/AIDS will precede the virus' ability to fully subvert host cellular immunity through population-level adaptation.
Research subjects, all adults, were enrolled under REB-approved protocols and provided written informed consent to participate in the original studies for which specimens were collected. Ethical approval to conduct this study was obtained from the Institutional Review Boards at Providence Health Care/University of British Columbia and Simon Fraser University.
A total of 497 historic plasma/serum specimens from unique patients enrolled in observational studies of men who have sex with men (MSM) at four North American sites between 1979–1989, were obtained for study. Of these, 94 and 75 were from the New York Blood Center (NYBC; 1979–1989) and the San Francisco Department of Public Health (SFDPH; 1979–1984), respectively, and represented participants of hepatitis B observational studies whose archived sera were retrospectively tested for HIV [34], [38], [39]. A further 282 and 46 were obtained from the Fenway Community Health Clinic in Boston (Fenway; 1985–1989) [36], [37] and the Vancouver Lymphadenopathy-AIDS Study in Vancouver, Canada (VLAS; 1984–1987) [40]–[42]. With the exception of 67 NYBC patients whose dates of HIV infection were estimated to be within 6 months prior to specimen collection, all other patients were known or presumed to be in chronic infection. Specimen integrity varied by cohort. Whereas sera from NYBC, SFDPH and Fenway were stored at −70°C since collection, VLAS specimens had been stored at −20°C and bore evidence of freeze-thaw cycles. No clinical information (i.e. plasma viral load, CD4) was available for historic specimens; furthermore, sociodemographic and other identifying information were not sought. Our modern comparison cohort comprised 382 individuals for whom HIV Gag and/or Nef sequences were available: 26 were recruited through the Aaron Diamond AIDS Research Center in New York, 91 from Massachusetts General Hospital in Boston and 265 from various cohort studies based at the BC Centre for Excellence in HIV/AIDS in Vancouver, Canada. The modern cohort comprised MSM, injection drug users and individuals with unknown HIV risk group.
HIV RNA was extracted from plasma or serum using standard methods. Gag and Nef regions were amplified by nested RT-PCR using sequence-specific primers and amplicons were bidirectionally sequenced on a 3130xl and/or 3730xl automated DNA sequencer (Applied Biosystems). Data were analyzed using Sequencher v5.0 (Genecodes) or RECall [96] with nucleotide mixtures called if the height of the secondary peak exceeded 25% of the height of the dominant peak (Sequencher) or 20% of the dominant peak area (RECall). All HIV sequences were confirmed as subtype B using the recombinant identification program (RIP; http://www.hiv.lanl.gov/content/sequence/RIP/RIP.html). HXB2-alignments were performed using an in-house tool based on the HyPhy platform [97]. Phylogenetic trees were constructed using maximum-likelihood approaches [98] and visualized using FigTree (http://tree.bio.ed.ac.uk/software/figtree/). Patristic (pairwise) genetic distances were computed using PATRISTIC [99]. Intercohort comparisons of Shannon entropy scores (featuring 1000 randomizations with replacement) were performed using Entropy-two (http://www.hiv.lanl.gov/content/sequence/ENTROPY/entropy.html). Detection of HIV Gag and Nef codons exhibiting significant evidence of pervasive positive selection (defined as having a posterior probability ≥0.9 that the site-specific nonsynonymous rate exceeds its synonymous rate) in the combined historic/modern datasets was performed using the fast unconstrained Bayesian approximation for inferring selection algorithm [56], implemented in Datamonkey [100], [101].
Consensus sequences were calculated by plurality rule. North American Gag and Nef HIV subtype B consensus sequences were computed from all available Gag and Nef sequences from unique patients annotated with Canada (CA) or United States (US) country labels in the Los Alamos HIV sequence database (N = 1624 and N = 1141 Gag and Nef amino acids sequences, respectively, spanning 1983–2011, accessed June 25, 2013). Historic plasma HIV RNA Gag and Nef sequences, annotated with year and country of collection, have been deposited in GenBank (Accession numbers KF701643–KF701941 for Gag and KF701942–KF702276 for Nef).
HLA class I typing was performed using an in-house sequence-based typing protocol capable of using plasma or serum as a source of genomic DNA [48] and types were assigned using an in-house algorithm. Where necessary, data were imputed to high resolution using a machine learning algorithm trained on a dataset of complete high resolution HLA-A, B and C types from >13,000 individuals with known ethnicity ([47]; http://research.microsoft.com/en-us/projects/bio/mbt.aspx#HLA-Completion) and assigned the highest-probability allele combination. HLA types could not be imputed when data were missing from more than one locus.
Gag and Nef sequences were annotated with sample dates. Putative recombinants were identified using SCUEAL [102] and removed. The most recent common ancestor (MRCA) sequences of Gag and Nef were estimated using Bayesian evolutionary analysis by sampling trees (BEAST) [52] via 6 (Gag) or 5 (Nef) replicate chains, each analyzing a different set of 200 sequences selected at random from the dataset, and yielding 10,000 ancestral reconstructions per chain, as follows. Trees were sampled at random from the posterior distribution of trees given an exponential relaxed molecular clock [103], a Bayesian skyline model of effective population size, and a time-reversible nucleotide substitution model determined by an Akaike information criterion-based model selection procedure in HyPhy [97]. Sampling was run for 2×108 steps, with the first half discarded as burn-in and the remainder thinned to 100 trees sampled at intervals of 106 steps in the chain. Convergence of replicate chains was assessed using the Tracer application in the BEAST software package. For each tree, 100 ancestral sequence reconstructions were sampled at random from the posterior distribution defined at the root under a Muse-Gaut codon substitution model in HyPhy. The inferred ancestral sequence was taken as the consensus of these 60,000 (Gag) and 50,000 (Nef) reconstructions (10,000 each per chain for 6 [Gag] and 5 [Nef] chains). Timing of each ancestral reconstruction (tMRCA) was estimated in BEAST by computing the mean estimate for each replicate chain and then computing a grand mean. The “consensus ancestor” Gag and Nef nucleotide sequences were commercially synthesized (Invitrogen LifeTech) for use in functional analyses.
The reference list of HLA-associated polymorphisms in modern HIV subtype B sequences was defined in an independent multicenter cohort of >1800 chronically subtype-B infected individuals from Canada, the USA and Australia recruited in the 1990s and 2000s, that did not overlap with historic and modern cohorts analyzed herein, using phylogenetically-informed methods [43]. The same methods [43] were used to identify HLA-associated polymorphisms in the historic dataset, as follows. Briefly, maximum likelihood phylogenetic trees were constructed using Gag and Nef sequences, and a model of conditional adaptation was inferred for each observed amino acid at each codon. Here, the amino acid is assumed to evolve independently along the phylogeny, until it reaches the tree tips (representing the present host). In each host, selection via HLA-mediated pressures and HIV amino acid covariation is directly modeled using a weighted logistic regression, in which the individual's HLA repertoire and covarying amino acids are used as predictors and the bias is determined by the possible transmitted sequences as inferred by the phylogeny [104]. To identify which factors (HLA and/or HIV covariation) contribute to the selection pressure, a forward selection procedure is employed where the most significant association is iteratively added to the model, with p-values computed using the likelihood ratio test. Statistical significance is reported using q-values [105], the p-value analogue of the false discovery rate (FDR). Q-values denote the expected proportion of false positives among results deemed significant at a given p-value threshold; for example, at q≤0.05, we expect 5% of identified associations to be false positives. HLA-associated polymorphisms are grouped into two categories: (1) amino acids significantly enriched in the presence of the HLA allele in question (“adapted” forms), and (2) amino acids significantly enriched in the absence of the HLA allele in question (“nonadapted” forms).
Second round Gag amplicons were selected from 120 historic and 71 modern patients with known or presumed chronic infection and cloned into the pCR2.1-TOPO TA vector (Life Technologies, Burlington, ON, Canada). A single representative clone harboring an intact Gag reading frame that closely resembled the patient's bulk plasma HIV RNA was selected for virus generation [60], [61]. Gag was amplified by PCR from each clone using 100 bp-long primers matching the NL4-3 sequence upstream and downstream of Gag, designed to facilitate homologous recombination of the amplicon with the pNL4-3Δgag backbone. The plasmid pNL4-3Δgag was developed by inserting unique BstEII restriction sites at the 5′ and 3′ ends of gag using the QuikChange XL kit (Stratagene), followed by deletion of the intervening region via BstEII digestion (New England Biolabs), gel-purification, and re-ligation (T4 DNA ligase; New England Biolabs). PNL4-3Δgag was maintained in Stbl3 E. coli cells (Invitrogen). To generate recombinant viruses, 10 µg of BstEII-linearized pNL4-3Δgag plus 50 µl of 2nd round Gag amplicon (∼5 µg) were mixed with 2.5×106 cells of a GFP-reporter T-cell line (CEM-derived GXR25 cells [106]) in 125 µl of Mega-Cell medium (Sigma), and transfected by electroporation in 96-well plates (exponential protocol: 250 Volts, 2000 µF; 25 millisecond pulse duration; BioRad MxCell_Pro). Following transfection, cells were rested for 15 min at room temperature, transferred to 25 cm2 flasks containing 1 million GXR cells resuspended in 5 mL of R20+ medium (RPMI 1640 containing 20% FCS, 2 mM L-glutamine, 100 units/mL penicillin, and 100 µg/mL streptomycin), and fed with 5 mL R20+ medium on day 5 and with replacement thereafter. Tat-driven GFP expression, indicating productive HIV infection of GXR cells, was monitored by flow cytometry (Guava 8HT, Millipore) starting on day 12 [60], [61]. Once GFP+ expression exceeded 15% among viable cells, supernatants containing recombinant viruses were harvested and aliquots stored at −80°C. Patient origin of all recombinant viruses was confirmed via sequencing of the Gag region.
Viral titers and replication capacity (RC) assays were performed using GXR25 GFP-reporter T-cells, as described [60], [61]. RC assays were initiated at MOI = 0.003, and included one negative control (uninfected cells only) and one positive control (NL4-3 Gag re-introduced into the NL4-3Δgag backbone using identical methods) per 24-well plate. For each virus, the natural log slope of the percentage (%) of GFP+ cells was calculated during the exponential phase of viral spread (days 3–6). This value was divided by the mean rate of spread of all NL4-3 controls such that RC values <1.0 or >1.0 indicate rates of spread that were slower than or faster than NL4-3, respectively. Each virus was assayed in a minimum of two independent experiments and average RC values are reported.
First-round Nef amplicons from 102 historic and 86 modern patients were originally selected and amplified using second round primers featuring EcoRI (forward) and SacII (reverse) restriction sites. Amplicons were PCR-purified (GeneJET PCR Purfication Kit, Thermo Scientific) and cloned into the pIRES2-EGFP expression vector (Clontech) as described in [67], [68]. For each patient, a single representative clone harboring an intact Nef reading frame that closely resembled the original bulk plasma HIV RNA sequence by phylogenetic analysis was selected for functional assessment.
CD4 and HLA class I downregulation activities for each Nef clone were measured using a CEM-SS derived T cell line that expresses high levels of surface CD4 and HLA-A*02 (CEM-A*02), constructed as described in [107]. To assess Nef-mediated CD4 and HLA downregulation, 3×105 CEM-A*02 cells were transfected with 5 µg plasmid DNA encoding Nef protein and GFP by electroporation (BioRad GenePulser MX). Twenty hours later, cells were stained with APC-labeled anti-CD4 and PE-labeled anti-HLA-A*02 antibodies (BD Biosciences) and cell surface expression was measured in transfected (GFP-positive) cells by flow cytometry (Guava easyCyte 8HT, Millipore). For patient-derived Nef clones, the median fluorescence intensity (MFI) of CD4 or HLA-A*02 expression in GFP-positive cells was normalized to the MFI of CD4 or HLA-A*02 expression for the negative control (empty pIRES2-EGFP plasmid) and positive control (nef reference sequence SF2, cloned into pIRES2-EGFP) to determine the relative CD4 or HLA-A*02 downregulation capacity. As such, a normalized value of 0.0 indicates no downregulation activity and a value of 1.0 indicates downregulation capacity equivalent to that of the positive control NefSF2. All assays were performed in triplicate and results are presented as the mean of these measurements.
Steady state Nef protein levels were measured by Western blot for the minority of Nef clones that displayed poor (<50%) function for either CD4 or HLA-A*02 downregulation activity, alongside 20 randomly-selected clones with activities above this threshold. A total of 5×106 CEM-A*02 cells were transfected by electroporation with 10 µg of plasmid DNA, and cell pellets were collected 20 hours later for preparation of total cell lysates, using a protocol modified from [107]. Lysates were subjected to SDS-PAGE in duplicate and electro-blotted onto PVDF membrane. To ensure detection of patient-derived Nef, duplicate blots were probed using anti-Nef polyclonal antisera developed from rabbit (NIH AIDS Research and Reference Reagent Program Catalog #2949, [108]) or sheep (ARP 444; NIBSC Center for AIDS Reagents, UK). Actin expression was simultaneously assessed. Band intensities were quantified on an ImageQuant LAS 4000 (GE Healthcare Life Sciences). A total of 15 poorly functional Nef clones whose expression could not be detected by Western Blot were excluded from analysis, as in vitro cloning or other defects cannot be ruled out. This left 93 historic and 80 modern Nef clones for analysis.
|
10.1371/journal.pgen.1006078 | The Contribution of GWAS Loci in Familial Dyslipidemias | Familial combined hyperlipidemia (FCH) is a complex and common familial dyslipidemia characterized by elevated total cholesterol and/or triglyceride levels with over five-fold risk of coronary heart disease. The genetic architecture and contribution of rare Mendelian and common variants to FCH susceptibility is unknown. In 53 Finnish FCH families, we genotyped and imputed nine million variants in 715 family members with DNA available. We studied the enrichment of variants previously implicated with monogenic dyslipidemias and/or lipid levels in the general population by comparing allele frequencies between the FCH families and population samples. We also constructed weighted polygenic scores using 212 lipid-associated SNPs and estimated the relative contributions of Mendelian variants and polygenic scores to the risk of FCH in the families. We identified, across the whole allele frequency spectrum, an enrichment of variants known to elevate, and a deficiency of variants known to lower LDL-C and/or TG levels among both probands and affected FCH individuals. The score based on TG associated SNPs was particularly high among affected individuals compared to non-affected family members. Out of 234 affected FCH individuals across the families, seven (3%) carried Mendelian variants and 83 (35%) showed high accumulation of either known LDL-C or TG elevating variants by having either polygenic score over the 90th percentile in the population. The positive predictive value of high score was much higher for affected FCH individuals than for similar sporadic cases in the population. FCH is highly polygenic, supporting the hypothesis that variants across the whole allele frequency spectrum contribute to this complex familial trait. Polygenic SNP panels improve identification of individuals affected with FCH, but their clinical utility remains to be defined.
| Familial combined hyperlipidemia (FCH) is a familial dyslipidemia and the most common familial risk factor for premature coronary heart disease. Its genetic architecture is poorly understood. Rare high-impact variants have been identified in some patients, but have not explained a substantial portion of the trait. FCH has previously been speculated to be a polygenic disorder, but genetic data supporting this hypothesis have so far been incomplete. We provide experimental evidence for the polygenicity and heterogeneity of FCH in a large set of affected families using comprehensive genome-wide variant data. Approximately a third of the affected FCH individuals in our sample had high polygenic burden, and only a minority carried high-impact variants identifiable by genotyping. We show that the polygenic burden of affected FCH family members is comparable to that observed in individuals with similar lipid phenotypes in the general population. Genetic variants identified in large-scale population studies can also underlie the typical phenotypes observed in complex familial diseases such as FCH. Advances in genetic diagnosis based on population samples may thus also benefit FCH families. Families without high polygenic burden are good candidates for sequencing studies to identify rare variants not observable with genotyping.
| Familial combined hyperlipidemia (FCH), classically defined by elevations in serum total cholesterol (TC), triglycerides (TG), or both, in two or more first degree relatives, displays a prevalence of greater than 1% in Western populations [1, 2]. It is the most common familial risk factor for premature coronary heart disease (CHD), occurring in 11–14% of individuals with this condition, and raising by up to five-fold the CHD risk in first- and second-degree relatives of affected individuals [3, 4]. In clinical practice, FCH is characterized by elevations of low-density lipoprotein cholesterol (LDL-C), TG, or both [5]. The phenotype within a family shows high inter- and intraindividual variability of lipid values (TG, LDL-C, high-density lipoprotein cholesterol [HDL-C], and apolipoprotein B) and therefore the diagnosis is commonly missed. Despite attempts to identify rare high-impact variants underlying FCH, no such variant has explained a substantial proportion of the trait [6]. Also, so far the genetics of FCH has not been addressed using high-density genotyping panels [7].
Here, we present a comprehensive evaluation of the genetic background of FCH using a dense-marker genotyping panel. We hypothesized that FCH risk could derive in part from a combination of common variants associated, in population cohorts, with LDL-C and/or TG as well as uncommon variants in genes that have been implicated in Mendelian lipid syndromes. To test our hypothesis we evaluated Finnish FCH families to determine if they demonstrated an enrichment of known lipid-associated variants, compared to Finnish general population samples, and assessed whether such variants could account for differences in lipid levels and degree of aggregation of disease observed between the FCH families. Additionally, following the logic of recent studies that have demonstrated the combined impact of multiple small effect variants (polygenic scores) on risk for familial hypercholesterolemia (FH) and other lipid disorders [8–11], we measured, in these families, the relative contributions of LDL-C and TG associated polygenic scores to the lipid traits and the FCH phenotype.
We evaluated the contribution of lipid-level associated genetic variation to FCH in 715 genotyped individuals (234 of whom were considered affected by FCH; Table 1, S1 Table, and S1 Fig) from 53 FCH families (average family size 13 individuals). Of the 234 FCH-affected individuals, 78 (33%) were identified because of elevated (≥ 90th age- and sex-specific population percentile) TC without elevated TG, 76 (32%) had elevated TG without elevated TC, and 80 (34%) had both lipids elevated. FCH probands (n = 48, because five probands had DNA unavailable) and other FCH affected individuals showed higher mean levels of both LDL-C (4.45 and 4.29 mmol/l, respectively) and TG (4.06 and 2.51 mmol/l, respectively) compared to all FCH family members (LDL-C and TG 3.64 and 1.60 mmol/l, respectively) and to the general Finnish population (LDL-C and TG 3.43 and 1.43 mmol/l, respectively) (Fig 1). The median LDL-C and TG values of the genotyped family members ranged from 1.08 to 4.69 mmol/l for LDL-C and from 0.70 to 4.44 mmol/l for TG demonstrating considerable variation among families (S2 Fig).
Compared to 18,715 random Finnish population samples, alleles that elevated LDL-C and TG levels in the population were overall enriched and alleles that lowered LDL-C and TG levels were depleted in 234 affected FCH individuals (sign test p = 0.0025 for LDL-C and p = 0.0016 for TG). We included in this analysis all SNPs with at least one affected carrier (194 of the 212 SNPs previously associated with either LDL-C or TG in population genome-wide association (GWA) studies or with monogenic dyslipidemias). In total, 60 out of 95 LDL-C elevating SNPs were more common in EUFAM than in population samples, and 57 out of 99 of LDL-C lowering SNPs were less common. Similarly, 61 out of 96 TG elevating SNPs were enriched and 57 out of 98 TG lowering SNPs were depleted (Fig 2, S3 Table). S3 Fig shows the enrichment ratios for probands and all family members compared to the population samples.
We then examined if high-impact variants implicated in monogenic dyslipidemias would be enriched and contribute to FCH. APOA5 rs3135506, which predisposes to hypertriglyceridemia in homozygous form [12], was 1.8–fold more frequent in the affected individuals of the FCH families than in the general population (minor allele frequency [MAF] in affected FCH individuals = 0.11, MAF in population = 0.062). In all FCH family members, there were 105 heterozygotes (43 of them affected) and six homozygotes (five of them affected). Homozygotes had a mean TG level (mmol/l) of 2.61, heterozygotes 1.97 and wild type carriers 1.55 (S4 Fig). APOE variations have a well-established role in dyslipidemias yet incompletely defined contribution to FCH. By genotyping rs7412 and imputing rs429358 we were able to determine the phenotyped apolipoprotein E isoform with a minimum accuracy of 95% (Table 2). The APOE ɛ2ɛ2 haplotype that predisposes to type III hyperlipoproteinemia [13], was observed in three FCH family members (two of them affected), all with elevated cholesterol and TG concentrations in very low-density lipoprotein and intermediate-density lipoprotein fractions (Table 2). Its frequency was 0.0023 in the Finnish population, 0.0045 in all FCH families and 0.0085 among FCH affected individuals. LIPC rs28933094 predisposes to hepatic lipase deficiency, the cardiovascular effects of which are unclear but may be dependent on the underlying lipid phenotype [14, 15]. It is 4.8–fold more frequent in Finns compared to other Europeans and additionally 2.6–fold more frequent in affected FCH individuals compared to the Finnish population (MAF in affected FCH individuals = 0.041, MAF in Finnish population = 0.016, MAF in Non-Finnish Europeans = 0.0033). Finally, it is of special interest whether the FCH families carry any of the classical FH variants. In the Finnish population there are five major extensively documented, FH-associated LDLR variants [16]. As the genotyping array does not capture these, we successfully imputed one out of the five variants (FH-Pogosta), but observed no carriers. In summary, variants that are strongly linked to known forms of dyslipidemia (in APOA5 and APOE), could at most explain only 7 (3.0%) of all 234 affected FCH individuals, thus being of minor importance in our FCH family sample.
We then estimated what proportion of the affected FCH individuals had a high cumulative burden of enriched LDL-C and TG elevating variants or carried high-impact variants. Out of the 234 affected FCH individuals, 83 individuals (35%) showed high accumulation of known LDL-C or TG elevating variants, having either one or both polygenic lipid scores over the 90th percentile in the population (Table 3). Three additional FCH affected individuals carried Mendelian variants and did not have elevated polygenic scores. In 14 out of the 53 (25%) families, over half of the affected individuals had high polygenic scores or carried known Mendelian variants (Fig 3). In six out of the 53 (11%) families, all affected members had high polygenic scores or carried known Mendelian variants. This did not appear to be driven by differences in genetic correlation between affected individuals in the families (S5 Fig).
Finally, we evaluated how well the high polygenic scores predicted FCH affectedness in the families or among hyperlipidemic individuals in the population (Table 4). Having a high polygenic score (≥ the 90th population percentile) for either LDL-C or TG had a positive predictive value (PPV) of 0.45 in the FCH families and 0.23 in the general population. Negative predictive values (NPV) were 0.71 and 0.89, respectively.
In summary, we observed enrichment of both common and uncommon lipid-elevating variants in the FCH families. Only a minority (7 out of 234, 3%) of affected individuals were carriers of high-impact variants of Mendelian dyslipidemias. The polygenic lipid scores contribute to FCH in over one third of 234 FCH subjects in these families, with considerable heterogeneity between families. In more than half of the families we did not observe either an increased load of common variants or carriers of high-impact variants.
Our results demonstrate an enrichment of many known lipid-level elevating variants in FCH families. This enrichment was observed for both uncommon and common variants known to affect either LDL-C or TG levels in populations. When the known variants were combined into polygenic lipid scores for each individual, 17% and 25% of the affected individuals had polygenic scores that were higher than the 90th percentile of the population for LDL-C and TG, respectively. In 13 families three or more affected individuals had high polygenic score and in 22 families none of the affected individuals had polygenic score above the 90th percentile of the population.
Our results allow us to draw several conclusions about the genetic background of FCH. First, variants across the whole frequency spectrum were enriched and contribute to the high LDL-C and/or TG levels in the affected FCH individuals. This emphasizes the polygenic nature of FCH and is in line with previous results with familial hypercholesterolemia demonstrating that this less prevalent syndrome has a polygenic component [9]. Genes in which we observed enriched variants include APOE, LIPC and APOA5, whose role in dyslipidemias is already established, but also include several genes whose function in lipid metabolism is not yet clear (e.g. UBR1, MTHFD2L and the PIGV-NR0B2 region).
Second, the majority of the enriched variants were originally identified in random population samples. This finding confirms that variants identified in population screens play a considerable role also in the complex familial disease FCH, and highlights the potential of using variants identified in population-based GWA studies to characterize familial dyslipidemia cases genetically.
Third, the observed enrichment was stronger for TG loci than for LDL-C loci. Many of the enriched TG SNPs were located in genes known to contribute to hypertriglyceridemia, such as the APOA1-C3-A4-A5 cluster, showing the central role of genetically driven TG in FCH [7, 17].
Fourth, over a third of the affected FCH individuals had a high load (polygenic lipid scores over the 90th percentile of the population) of either known LDL-C or TG associated variants. This proportion with high polygenic score was comparable to the population samples with similar levels of high LDL-C or TG in the population. This observation suggests that the role of the known variants is similar in familial hyperlipidemias and randomly ascertained high LDL-C and/or TG individuals. However, the accumulation of LDL-C or TG-elevating alleles in families was highlighted by the prediction analysis results. High load of known LDL-C or TG associated variants predicted affectedness in FCH families almost two times better than comparable hyperlipidemia in the population.
There were large differences among families in the proportion of affected individuals with high polygenic scores. While in nine families more than two thirds of the affected individuals had high polygenic score, in over a third of the families, none of the cases had high score and the underlying genetic architecture remains unexplained in the majority of families. As genotyping and imputation allow for the evaluation of only a portion of catalogued high-impact variants, future sequencing studies might pinpoint rare causal variants in these families.
Our study supports the hypothesis that FCH is a genetically heterogeneous disease, reflected by its heterogeneous lipid phenotype [18–20]. large number of LDL-C associated variants are mainly responsible for the elevation of LDL-C in the FCH lipid profile. In contrast, a handful of low- to moderate-impact TG variants drive the elevation of triglycerides. Our study provides support for the polygenic rather than monogenic nature of FCH, and highlights the central role of genetically driven TG in FCH.
Written informed consent was obtained from all study participants. All samples were collected in accordance with the Helsinki declaration and study protocols were approved by the ethics committees of the participating centers (The Hospital District of Helsinki and Uusimaa Coordinating Ethics Committee, approval number 184/13/03/00/12).
As part of the European Multicenter Study on Familial Dyslipidemias in Patients with Premature Coronary Heart Disease (EUFAM), the Finnish FCH families were identified from patients admitted to university hospitals with a diagnosis of premature CHD who demonstrated levels of TC, TG, or both that were ≥ 90th Finnish age- and sex-specific population percentile (S4 Table) [21, 22]. Families that contained at least one other first-degree relative affected with hyperlipidemia (according to the same lipid-level criteria as used for the probands) were included in the study. Additionally, in the included families, at least one of the affected individuals had high TG. All family members with hyperlipidemia were then considered affected by FCH. Probands with a diagnosis of FH (screened with a functional LDL receptor test), and any subjects with diabetes or other chronic diseases were considered unaffected and did not contribute to establishing the family’s FCH status (S1 Text).
For analyses of continuous lipid traits, individuals using lipid-lowering or estrogen medication at the time of sampling were excluded. Samples from the Finnish National FINRISK study were used as a Finnish population-specific comparison group, and individuals with known diabetes or cancer were excluded from the analyses (S1 Text). For the FCH families, venous blood samples were obtained after an overnight fast. FINRISK participants were advised to fast for four hours before the examination and avoid heavy meals earlier during the day. For both the EUFAM and FINRISK samples, circulating biochemical markers were measured from the venous blood samples using standard methods, as described in the S1 Text.
All FCH individuals with DNA available (715 out of 1161 from all 53 FCH families) and the FINRISK samples (n = 20,626) were genotyped with the HumanCoreExome BeadChip (Illumina Inc., San Diego, CA, USA) using standard methods (S1 Text). Additionally, over nine million variants were imputed using a combined reference panel of 1000 Genomes Phase I integrated haplotypes and 1943 Finnish genomes (S1 Text).
To construct a polygenic lipid score, we catalogued and used 212 SNPs, representing either the lead SNPs for genome-wide significant associations to LDL-C or TG from GWA studies of these traits (171 SNPs), or SNPs catalogued in the Online Mendelian Inheritance in Man (OMIM) database as located in genes implicated in primary and secondary monogenic dyslipidemic syndromes (44 SNPs, four of which overlap with the GWA lead-SNPs, S2 Table) [23–26]. The SNPs were assigned to LDL-C and TG scores based on their previously reported associations (S1 Text). The scores were calculated as the sum of the risk alleles weighted by their effect estimates drawn from a multiple linear model estimating all SNP effects on the trait (LDL-C or TG) at the same time in the FINRISK samples (S1 Text). The FINRISK samples used to estimate the weights were independent from the FCH samples.
We estimated allele frequencies for all 212 SNPs in the 234 affected FCH individuals and in 18,715 FINRISK samples after excluding individuals with known diabetes or cancer similarly to the EUFAM exclusion criteria. The minor allele in FINRISK was designated as the effect allele. We then calculated enrichment ratios by dividing the effect allele frequencies in affected FCH individuals with the effect allele frequencies in the FINRISK samples. Under the null hypothesis of no enrichment, we estimated the null distribution for enrichment testing by calculating the enrichment statistic (enrichment ratio) for all variants with MAF > 0.001% (12,234,754 SNPs) across the genome excluding regions within 50 000 base pairs from the lipid-associated SNPs. We estimated the null distribution in MAF bins across the genome while keeping the family structure fixed (FINRISK minor allele frequency intervals [0.1%,0.5%); [0.5%,1%); [1%,1.5%); [1.5%,2%); [2%,2.5%), [2.5%,5%); [5%,10%); [10%,15%); [15%,20%); [20%,25%); [25%,30%); [30%,35%); [35%,40%); [40%,45%); and [45%,50%)). The observed lipid SNP allele frequency enrichment ratios were compared to the null distribution to provide one-sided p-values for each SNP. We used sign test for all SNPs together to estimate whether the direction of enrichment (more or less common in the affected FCH individuals than in the population) was associated with the SNPs’ effects on the lipid in question (elevating or lowering).
We applied linear mixed models to test for differences in metabolic and clinical characteristics between FCH affected and unaffected individuals (S1 Text). An empirical genetic correlation matrix between individuals was included as the covariance structure of a random effect. Linear mixed models were applied with MMM (version 1.01) [27], and the other statistical analyses were performed using R (version 3.2.1) [28].
|
10.1371/journal.pgen.1003516 | Ikbkap/Elp1 Deficiency Causes Male Infertility by Disrupting Meiotic Progression | Mouse Ikbkap gene encodes IKAP—one of the core subunits of Elongator—and is thought to be involved in transcription. However, the biological function of IKAP, particularly within the context of an animal model, remains poorly characterized. We used a loss-of-function approach in mice to demonstrate that Ikbkap is essential for meiosis during spermatogenesis. Absence of Ikbkap results in defects in synapsis and meiotic recombination, both of which result in increased apoptosis and complete arrest of gametogenesis. In Ikbkap-mutant testes, a few meiotic genes are down-regulated, suggesting IKAP's role in transcriptional regulation. In addition, Ikbkap-mutant testes exhibit defects in wobble uridine tRNA modification, supporting a conserved tRNA modification function from yeast to mammals. Thus, our study not only reveals a novel function of IKAP in meiosis, but also suggests that IKAP contributes to this process partly by exerting its effect on transcription and tRNA modification.
| The process of meiosis is responsible for gamete formation and ensures that offspring will inherit a complete set of chromosomes from each parent. Errors arising during this process generally result in spontaneous abortions, birth defects, or infertility. Many genes that are essential in regulating meiosis have also been implicated in DNA repair. Importantly, defects in DNA repair are common causes of cancers. Therefore, identification of genes important for normal meiosis contributes not only to the field of reproduction but also to the field of cancer biology. We studied the effects of deleting mouse Ikbkap, a gene that encodes one of the subunit of the Elongator complex initially described as an RNA polymerase II–associated transcription elongation factor. We demonstrate that Ikbkap mutant mice exhibit infertility and defects in meiotic progression. Specifically, homologous and sex chromosomes fail to synapse (become associated), DNA double-strand breaks are inefficiently repaired, and DNA crossovers are significantly decreased in Ikbkap males. We also demonstrate that the requirement for Elongator in tRNA modification, which has been shown in lower eukaryotes, is conserved in mammals. Our findings suggest novel roles for Ikbkap in meiosis progression and tRNA modification, which have not been reported previously.
| Meiosis is a fundamental and highly regulated process that takes place during gamete generation. Faithful execution of this process is essential for maintaining genome integrity. Errors and various types of disruption during meiosis can cause aneuploidy and result in developmental defects, including mental retardation in Trisomy 21, infertility, to name two [1].
During the prophase I stage of the first meiotic division, homologous chromosomes undergo pairing and synapsis. Synapsis is mediated by a protein complex namely the synaptonemal complex (SC), and is accompanied by chromosome recombination [2]. Unlike homologous autosomes, the X and Y chromosome synapsis occurs only at a very small region of homology, a pseudoautosomal region (PAR) [3]. Formation of the fully synapsed autosomal SCs as well as the partially synapsed sex chromosome are essential for DNA repair, recombination and subsequent desynapsis [4]. Consequently, DNA damage response (DDR) is initiated upon the recognition of the DNA lesion made by SPO11, which is a type II-like topoisomerase that induces double-stranded breaks (DSBs) [5]. At the DSB sites, the DNA repair machinery generates DNA recombination between homologous chromosomes to ensure proper disjunction at metaphase I. The genetic studies in yeast and mouse helped identify many factors important for meiosis [6], [7], [8], such as: the master regulators meiosis-inducing protein 1 (Ime1) in yeast, and A-MYB (MYBL1) in mouse [9], [10]. Despite great progress in understanding the transcriptional regulation of the meiotic process [2], very little is known about the role of translational control during this process. Our data presents evidence that the evolutionarily conserved factor Ikbkap/Elp1 governs meiotic progression at the level of both transcription and translation.
Elp1, also referred to as IKAP (Inhibitor of kappaB kinase -associated protein), functions as a scaffold protein that assembles the Elongator and is encoded by Ikbkap gene (we will use the MGI nomenclature, IKAP for the protein, and Ikbkap for the gene, hereafter). Elongator is a protein complex comprised of two copies of the core complex, Elp1–3, and a sub-complex, Elp4–6 [11]. The protein complex “Elongator” was first purified in budding yeast through its association with the elongating RNA polymerase II (RNAP II) [12]. Similar protein complex was subsequently purified from human cells [13], [14], [15]. Interestingly, the components of the protein complex are highly conserved in different species that include yeast and human. The Elongator complex has important biological functions as deletion or mutation of any of its subunits results in severe phenotypes in yeast. Among the Elongator components, Elp3 likely serves as a catalytic subunit, because it not only harbors motifs characteristic of the GCN5 family of histone acetyltransferases (HATs), but also has been shown to directly acetylate H3 lysine 14 (H3K14) and possibly H4K8 in vitro [16]. These findings, combined with the studies demonstrating the association of Elongator with RNAP II holoenzyme, its ability to bind to nascent pre-mRNA, and to facilitate RNAPII transcribes through chromatin in an acetyl-CoA-dependent manner, support its role in transcription regulation [12], [14], [17].
Accumulating evidence suggest that Elongator, in addition to participating in transcriptional regulation, also plays pivotal role in the regulation of translation. The first evidence implicating the involvement of the Elongator in translation came from a genetic screen, which demonstrated that all genes encoding the yeast Elongator subunits are required for the formation of 5-carbamoylmethyl (ncm5), and 5-methoxycarbonylmethyl (mcm5) side chains on uridines at the wobble position of certain tRNAs [18]. These modified nucleosides are important for efficient decoding of A- and G- ending codons through stabilizing codon-anticodon interactions during translation [19], [20], [21]. Studies have shown that all the six subunits of the Elongator are required for the early step of mcm and ncm side chain formation [18]. Although it is currently unclear how Elongator contributes to the generation of the modified tRNAs, this function is conserved in S. cerevisiae, S. pombe, C. elegans, and A. thaliana [18], [22], [23], [24]. Whether such function is conserved in mammals remains to be determined.
Using a loss-of-function approach, we demonstrate that IKAP plays an important role in male meiosis. First, we show that IKAP is highly expressed in male germ cells. Targeted deletion of Ikbkap in mice resulted in increased apoptosis in male germ cells and male infertility. Interestingly, autosomal and sex chromosome synapsis defects are observed in Ikbkap mutant spermatocytes. In addition, sustained RAD51 foci are observed on the autosomes of mutant spermatocytes, suggesting a homologous recombination repair defect. Detailed molecular studies revealed that the expression of a few meiotic genes is down-regulated in mutant testes. Furthermore, the levels of the Elongator-dependent tRNA modifications are reduced in the mutant testes. Our study thus reveals a critical function of Ikbkap in male meiosis, and demonstrates a conservation tRNA modification function in mammalian cells.
To explore a possible role of IKAP in gametogenesis, we analyzed the expression pattern of IKAP during spermatogenesis by immunofluorescence staining. This analysis revealed that IKAP is expressed in Tra98-positive gonocytes as early as postnatal day 0 (P0) with a predominant cytoplasmic localization (Figure 1A). This expression and localization pattern is maintained at P8, as prospermatogonia developed into PLZF-positive undifferentiated spermatogonia (Figure 1A). At P21, IKAP expression remains in SYCP3-expressing meiocytes (Figure 1A). At late stage of spermatogenesis, IKAP was detected in RNA polymerase II-positive round spermatids (Figure 1A, arrows), but not in transition protein 1 (TNP1)-positive elongated spermatids at P35 (Figure 1A). In contrast to specific expression in germ cells, IKAP is undetectable in the GATA1-positive Sertoli cells (Figure 1B, arrow) or 3β-HSD positive Leydig cells (Figure 1B). We also used the conditional knockout testes (as below) as negative controls for the purpose of antibody validation (Figure S1). Collectively, immunofluorescence staining revealed that IKAP is expressed in all stages of male germ cells except the elongated spermatids, but it is almost undetectable in somatic cells of testis.
The germ cell-specific expression pattern of IKAP revealed above suggests that IKAP might have a role in spermatogenesis. Previous studies have demonstrated that Ikbkap null mutant mice die of cardiovascular and neuronal developmental defects at embryonic day E10 [25], [26]. To bypass the embryonic requirement for Ikbkap, we used mice harboring a conditional knockout allele for Ikbkap with exon 4 flanked by two loxP sites (Figure S2A). Mice homozygous for the Ikbkapflox conditional allele were viable and were born at Mendelian ratio (data not shown). To explore the function of Ikbkap in spermatogenesis, we inactivated Ikbkap in the male germ line by crossing with the Vasa-Cre mice (also known as Ddx4-Cre). Vasa-Cre induces recombination in germ cells starting from E15.5, and is expressed in all spermatogenic cells postnatally [27]. The germ linage conditional Ikbkap mutant mice (genotyped as Vasa-Cre; Ikbkapflox/−, referred to as CKO hereafter) were obtained by crossing Ikbkapflox/flox females with Vasa-Cre; Ikbkapflox/+ males. The genotypes of control mice were either Vasa-Cre; Ikbkapflox/+, or Ikbkapflox/−, or Ikbkapflox/+. RT-qPCR analysis using P16 mouse testes demonstrates that the deletion efficiency is more than 80% (Figure S2B). Western blot analysis and immunostaining using two commercial antibodies revealed marked reduction of IKAP protein in the CKO testes (Figure S2C, S2D and data not shown).
To test for a possible function of IKAP in spermatogenesis, CKO and control male mice were mated with wild-type females and the breeding capacity was monitored for 3 months. While the control mice gave birth at an average litter size 6.7±1.5, no pups were obtained from wild-type females mated with CKO males, even though copulatory plugs were frequently observed in the females. These results suggest that loss of function of Ikbkap in male mice causes infertility.
To determine the potential cause of infertility, we examined the size of male gonads and the presence of spermatozoa in the epididymis. We found that the size of the testes is significantly decreased in the CKO mice, and the testicular weight to body weight ratio was reduced by 25% at the age of 14 month (Figure 2A). No spermatozoa were found in the epididymis of 2-month old CKO mice (Figure 2B). Histological analyses indicated stage IV (mid-pachytene) arrest of the seminiferous epithelium (Figure 2B), which is typical of diverse mouse meiotic mutants [28]. Indeed, in contrast to control seminiferous tubules, CKO testes lacked postmeiotic spermatids. To determine the stages at which Ikbkap deficiency causes germ cells perturbation, we examined the first round of spermatogenesis using juvenile testes. Immunostaining with the meiocyte maker SYCP3 revealed no obvious histological change in CKO testes at P14 (Figure 2C), suggesting that CKO germ cells enter meiosis and progress to the prophase I normally as in testes of control mice. However, at P21, while most of the tubules of control testes have postmeiotic round spermatids, CKO testes have almost no postmeiotic cells with SYCP3-expressing meiocytes disorganized and scattered through seminiferous tubules (Figure 2C). At P35, while control testes showed a full spectrum of spermatogenic cells, including primary spermatocytes, rounds spermatids, and spermatozoa, no post-meiotic cells, such as round and elongated spermatids, were found in the CKO testes (Figure 2C and Figure S3A). These results suggest that Ikbkap deletion causes spermatogenesis arrest at meiotic prophase (Figure 2C). Consistent with the lack of post-meiotic germ cells in the mutant testes, Terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) assays revealed an increase in the number of apoptotic cells in the CKO testes compared to that of the controls (Figure 2D; p<0.001), indicating that apoptosis at least partly explains the lack of round and elongated sperms in the CKO testes. Consistent with the absence of IKAP in somatic cells, no morphological change in Sertoli cells or Leydig cells was observed in CKO testes, and the density of GATA1-positive Sertoli cells and 3β-HSD-positive Leydig cells was also not altered in response to Ikbkap deletion (Figure S3B). Taken together, the above results demonstrate that Ikbkap deletion impedes spermatogenesis during meiotic stage, resulting in increased apoptosis.
The lack of post-meiotic germ cells in the CKO testes prompted us to examine the impact of Ikbkap deletion on meiotic process. The meiotic prophase is divided into five stages, including leptotene, zygotene, pachytene, diplotene, and diakinesis [2]. We first examined the stage distribution of spermatocytes in control and CKO testes based on chromosomal morphology and sex body status. In the control P21 testes, pachytene spermatocytes are the most abundant, accounting for half of the total cell population, followed by diplotene, zygotene and leptotene stage cells (Figure 3A). In contrast, the CKO testes exhibited an accumulation of zygotene spermatocytes, and a significant decrease of pachytene spermatocytes (Figure 3A, p<0.01), suggesting that the impairment of meiotic progression takes place between zygotene to pachytene stage.
To characterize the observed meiotic defect in detail, we performed co-immunostaining of meiotic chromosome spreads using antibodies against axial/lateral (SYCP3) and central (SYCP1) elements of the synaptonemal complex. In normal meiosis, the axial element of synaptonemal complex starts to form at the leptotene stage (Figure S4). Synapsis is initiated at the zygotene stage, as determined by the appearance of SYCP1- a marker of fully synapsed chromosome segments (Figure 3B). Synapsis is completed at the pachytene stage as chromosome cores contained 20 fully synapsed bivalents, including XY pair synapsed at the pseudoautosomal region (PAR). Despite normal development of axial elements, CKO spermatocytes exhibited an increase in the unpaired SC at zygotene stage. Zygotene-like nuclei, with approximately 40 or more short fragmented stretches of SYCP3 and no SYCP1 staining were observed in 61.3% (n = 60) of CKO zygotene spermatocytes (Figure 3B), suggesting a synapsis defect. To further analyze the synaptic defects, we investigated the centromere distribution by immunostaining with centromere marker CREST and SYCP3. Prior to synapsis, 40 centromeres are usually observed in the control leptotene spermatocytes. As the synapsis progresses, the number of visible centromeres reduces and becomes 21 centromeric foci (19 from synapsed autosomes and 2 from the XY bivalent) at the pachytene stage (Figure 3C). In the CKO zygotene-like spermatocytes, we observed greater than 20 centromeric foci, most containing 40 CREST foci (Figure 3C), indicating CKO spermatocytes failed to complete homologous chromosome pairing. Although 27.7% of CKO spermatocytes proceed to pachytene stage and exhibit 19 fully synapsed autosomal bivalent chromosomes (Figure 3A and 3B), XY asynapsis, in which X and Y axes were not associated, were frequently observed in CKO spermatocytes (67.5%, n = 225), as judged by the absence of SYCP1 (Figure 3B and 3D). Taken together, Ikbkap deficiency in mouse spermatocytes leads to the disruption of synapsis.
To further characterize the XY synapsis defect, we stained spermatocytes for γH2AX (a phosphorylated form of histone H2AX), a marker of DSBs, which are abundant in the silenced sex body. At early stage of prophase I, phosphorylation of H2AX is induced by SPO11-catalyzed DSBs in meiotic DNA [5]. γH2AX exhibits a diffuse staining pattern during the leptotene and zygotene stages, and becomes exclusively localized on the sex chromosomes (so called sex body) within pachytene and diplotene spermatocytes [29], [30]. We observed slightly decrease of γH2AX staining in CKO leptotene and zygotene spermatocytes as compared to the control, indicating inefficient initiation of DSB in the CKO testes (Figure S4). At pachytene stage, γH2AX localization was only restricted to the sex body, but not autosomes in the control spermatocytes (Figure 4A). Although sex bodies are formed in the CKO pachytene spermatocytes, the γH2AX signals were more concentrated in chromatin surrounding the XY axes rather than in more distant region of chromatin (Figure 4A). Moreover, not only weak γH2AX foci were persistent abnormally in the pachytene stage of CKO chromosome (Figure 4A, arrows), but more than one localized γH2AX signals were frequently observed in CKO spermatocytes (Figure S5). We further confirmed that the axes of the autosomes in CKO spermatocytes were indeed covered by γH2AX cloud using confocal microscopy (Figure S5). These results suggest an accumulation of unrepaired DSBs. Next, we analyzed sex chromosome-specific synaptic defects in details, and found that 77% of mutant spermatocytes exhibited distinct dissociation of X and Y axis (Type I in Figure 4A), whereas 10% represented illegitimate association of an end of X axis to autosomes (Type II in Figure 4A), and 13% displayed persistent γH2AX signals along the autosomes multiple chromosome within sex body (Type III in Figure 4A) (distribution of the three types is presented in Figure 4B). Taken together, our results indicate that Ikbkap is essential for both autosomal and XY synapsis as well as DSB repair.
The persistence of γH2AX foci in CKO spermatocytes prompted us to investigate a possible DSB repair defect. In normal meiotic recombination, SPO11-induced DSBs are repaired by the eukaryotic RecA homologs RAD51 and DMC1 (meiosis specific), which catalyze the invasion and strand exchange reaction between non-sister chromatids on homologous chromosomes [31]. To this end, we characterized the distribution of the RAD51 and DMC1 recombinases by immunostaining with an anti-RAD51 antibody that recognizes both proteins. In normal meiosis, RAD51 formed numerous foci at leptotene and zygotene stage, but these foci disappeared from autosomal axes and remained only at unsynapsed region like X and Y axis during pachytene stage [32], [33]. Consistent with a decreased γH2AX staining pattern in the CKO spermatocytes, the number of RAD51/DMC1 foci was significantly decreased in the CKO leptotene and zygotene spermatocytes as compared to the control (Figure S6). These results indicate a defect in early meiotic recombination. Moreover, while RAD51/DMC1 foci were detected in sex chromosome of the control pachytene spermatocytes (Figure 4D), they remained not only on the X axis, but also in autosomes in CKO pachytene spermatocytes. These results suggest an impairment in DSB repair in Ikbkap mutant spermatocytes. We next asked whether meiotic process can past the mid-pachytene stage in CKO spermatocytes by staining for Histone 1t (H1t), a mid-pachytene marker. Both control and the remaining CKO spermatocytes at mid/late pachytene stage expressed H1t (Figure 4C), suggesting that the remaining CKO spermatocytes progressed to mid-pachytene stage. To examine whether the defect in early recombination events led to the development of reciprocal exchanges (crossovers) between homologous chromosomes, we examined the distribution of mismatch repair protein MLH1, which marks the locations of crossovers [34]. Mid-pachytene control spermatocytes had 1–2 MLH1 foci on each synapsed chromosome and ∼23 per nucleus (Figure 4C). In contrast, a decreased number of MLH1 foci per nucleus were observed in CKO spermatocytes (Figure 4C), suggesting a defect in crossover formation in CKO spermatocytes.
We next sought to investigate the underlying mechanism by which Ikbkap deficiency causes spermatogenic arrest in pachytene stage. Depletion of IKAP in human cells has been previously liked to transcriptional and cell migration defects [35]. To identify potential Ikbkap-regulated genes in meiosis, we performed microarray analysis using RNAs purified from control and CKO P15 testis. We chose to use P15 testis because the first wave of spermatogenesis progresses to the pachytene stage around this time. We observed that the levels of 1810 transcripts were significantly altered (paired t-test, P<0.05). However, the changes are all less than 2-fold, as indicated by scatter plot analysis (Figure 5A). Among the altered transcripts, 1103 were down-regulated, as illustrated by the heat map (Figure 5B). Gene ontology (GO) analysis revealed that the affected genes that are most enriched are involved in cell cycle and M phase processes (P value = 10−14∼10−15). Other terms with a significant P value (<10−4) include meiosis, DNA repair, spermatogenesis and male gamete formation (Figure 5C). By comparing Ikbkap-affected genes with a list of genes that were previously demonstrated to be required for synapsis, we identified Spo11, a type II like topoisomerase (including α and β isoforms), Rad18 (ubiquitin ligase), and subunits of cohesion, including Smc1β, Rec8 and Stag3. RT-qPCR analysis confirmed their down-regulation in CKO testes (Figure 5D). In addition, we verified the down-regulation of several spermatogenesis relevant genes including the boule-like (BOLL), and Tudor domain containing 1(Tdrd1) (Figure 5D). Interestingly, Spo11, Smc1β, Rec8 and Rad18 are known to play a role in meiotic DSBs repair. Given the phenotypical similarity between Ikbkap, Spo11 [36], [37], Smc1β [38], Rec8 [39], [40] and Rad18 mutants [41], we believe that down-regulation of Spo11, Smc1β, Rec8 and Rad18 at least partly contribute to the Ikbkap CKO phenotype.
In mammals, heterologous unsynapsed chromatins (sex chromosomes) are transcriptionally silenced during meiosis, a phenomenon called “meiotic sex chromosome inactivation” (MSCI) [42]. Failure in MSCI leads to apoptosis of pachytene spermatocytes, which has been proposed to be the reason for the elimination of the spermatocytes of asynaptic mutants, such as Spo11 or Dmc1 [43], [44]. Given the sex chromosome synapsis defect exhibited in the Ikbkap mutant, we asked whether it affects MSCI. At P15, when MSCI is established in normal meiosis, genes on the sex chromosomes were not repressed and were significantly up-regulated in CKO testes as compared with control testes (Kolmogorov-Smirnov test, P<0.05) (Figure 6A). We furthered confirmed the sex chromosome specific up-regulation in CKO testes as compared to autosomes (Figure 6A). To further validate the results, we analyzed the gene expression level of selected X-, Y-, and autosome-linked genes by RT-qPCR. Among the three Y-linked genes (Zfy1, Zfy2, Ube1y1) analyzed, Zfy2 was significantly up-regulated in CKO testes (Figure 6B). We also analyzed four X-linked genes that are expressed in meiotic and postmeiotic cells (Ccnb3, Nxt2), or in premeiotic cells but repressed in meiotic and postmeiotic cells (Tex16, Hprt) [45]. In addition to few autosomal genes that we examined in Figure 5D, we also examined additional five autosomal genes that are expressed in meiotic cells (Syce1, Syce2, Sycp1, Sycp3 and Tex12). While the expression of the autosomal genes was either not altered or down-regulated, the X-linked genes Ccnb3, Tex16, and Hprt showed significant increase in CKO testes (Figure 6B). Taken together, only X- and Y-linked genes, which are normally repressed during prophase I in male germline, showed increased expression in CKO testes, suggesting that Ikbkap is important for MSCI.
Elongator has been shown to directly interact with tRNA in vitro [11], [18] and is required for wobble uridine tRNA modification in yeast, plant and worm [18], [22], [23], [24].To determine whether this function is conserved in mouse, we asked whether wobble uridine tRNA modification is affected by Ikbkap deletion. To test this possibility, total tRNA was extracted from P15 testis and subjected to nucleotide digestion before LC-MS-MS analysis. We used synthetic nucleoside standards to determine the retention time and nucleoside-to-base ion transition. Similar to S. cerevisiae, S. pombe and C. elegans tRNAs, and in accordance with previous studies [46], we found that mouse tRNA contain mcm5U, ncm5U, and mcm5s2U (Figure 7A and 7B). Importantly, the levels of mcm5U, ncm5U, and mcm5s2U in tRNA are significantly reduced in the CKO testes (Figure 7B). Quantification indicates that mcm5U, mcm5s2U, and ncm5U levels in the tRNAs of the CKO testes are only about 33%, 37%, and 47% that of the control levels, respectively (Figure 7C). Similar results were obtained from tRNAs isolated from P21 or 2 month-old testes (data not shown). tRNA from Elongator mutants of budding yeasts showed accumulation of 2-thio uridine (s2U), which is absent in wild-type tRNA, probably reflecting the thiolation of unmodified wobble uridine [18]. Indeed, we found correspondingly that, while s2U was readily detectable in the tRNAs of the CKO testes, it was not detectable in control testes (Figure 7B). This result suggests that IKAP deficiency causes accumulation of unmodified wobble uridine, some of which is thiolated into s2U. Taken together, our result suggests that, similar to yeast Elp1p and C. elegans ELPC-1, mouse IKAP is responsible for early steps of mcm5s2U, and ncm5U modification of tRNAs. The incomplete elimination of the formation of wobble uridine modification in CKO testes could be due to the incomplete deletion of Ikbkap and/or possible compensation by other pathways.
Elongator has been shown to play important roles in both transcription and translation, and is regarded as an essential regulator of normal development. Indeed, germ-line deletion of Ikbkap in mouse causes embryonic lethality at E10 likely due to impaired cardiovascular development and/or function [25], [26]. Loss of function of any of the six subunits of the Elongator impairs development in a wide variety of organisms, including yeast, fly and worm. In this study, we demonstrate that loss of function of IKAP results in multiple meiotic defects including defects in synapsis, DSB repair, and meiotic progression. In addition, we show loss of function of IKAP results in down-regulation of meiotic genes. Furthermore, we show that IKAP is required for the formation of mcm5 and ncm5 at U34 in tRNA. Taken together, our results support the notion that Elongator is required for tRNA modification and this function is conserved in all eukaryotes.
Key events in meiotic prophase I include: (1) introduction of SPO11-dependent double-stranded breaks (DSBs), (2) synapsing of the homologous chromosomes, (3) meiotic sex chromosome inactivation (MSCI), and (4) repair of DSBs by homologous recombination [2]. Ikbkap- deficient spermatocytes arrest at pachytene stage and show various meiotic phenotypes, including aberrant homologous and sex chromosomal synapsis, accumulation of unrepaired DSBs, lack of crossing over, as well as defective MSCI. Defects in synapsis and DSB repair are observed in CKO spermatocytes, suggesting that Ikbkap plays a role in these processes. One of the possible causes of meiotic arrest in the CKO spermatocytes could be activation of a pachytene checkpoint. Spermatocytes with defects in chromosome synapsis and/or recombination commonly trigger pachytene checkpoint control that can delay or arrest meiosis at the pachytene stage of prophase I [47]. However, the accumulation of mid-pachytene marker H1t suggests that the remaining CKO spermatocytes transits past mid-pachytene stage, which is later than pachytene checkpoint (Figure 4C). Another plausible explanation for spermatocyte elimination could be defective MSCI. MSCI is a quality control system unique to spermatocytes, and malfunction of MSCI is sufficient to trigger apoptosis of the pachytene spermatocytes [48], [49]. Indeed, we observed up-regulation of transcripts from the sex chromosomes in the CKO spermatocytes, suggesting a deficiency in MSCI. In particular, Zfy2 expression was significantly up-regulated (Figure 6B). Zfy1/2 paralogs are thought to be stage IV killer genes as ectopic expression of Zfy1/2 in XY males is sufficient to phenocopy the pachytene arrest phenotype [43]. Such spermatocytes undergo apoptosis and are eliminated at stage IV of the testicular epithelial cycle. Taken together, our results suggest that loss of Ikbkap in germ cells likely triggers a pachytene checkpoint, which together with defective MSCI leads to spermatocyte arrest and apoptosis.
Another question raised in our study was how Elongator contributes to meiotic defects in male germ cells. Our study showed that the expression of major meiotic genes involved in synapsis, including Spo11 (inclusive of α and β isoforms), Rad18, Smc1β, Rec8 and Stag3 are down-regulated in P15 juvenile testes. Among them, Smc1β, Rec8 and Stag3 belongs to the cohesin complex which provides sister chromatid cohesion and ensures chromosome segregation in mitosis and meiosis [50]. In mammalian germ cells, meiotic-specific cohesin complex contains four evolutionarily conserved protein subunits: two SMC (structural maintenance of chromosomes) proteins, SMC1β and SMC3, which heterodimerize, and two non-SMC subunits, REC8 and STAG3 [50]. They form a ring-shaped structure which embraces sister chromatids [51]. Knockout mouse models for SMC1β [38] and REC8 [39], [40] have been developed. While male meiosis of Smc1β-deficient mice is blocked in pachytene stage, Rec8-deficient spermatocytes could not proceed to pachytene. Interestingly, both Smc1β and Rec8-deficient mice show severe defects in synapsis, recombination, as well as crossover [38], [39], [40], which phenocopy Ikbkap meiotic phenotypes. In addition to cohesion complex, SPO11, which introduces DSBs during meiotic prophase, was also down-regulated in CKO spermatocytes. Given that Spo11 deficiency results in failure in the initiation of meiotic recombination [36], [37], inefficient generation of DSBs in CKO leptotene/zygotene spermatocytes might result from the down-regulation of Spo11. Furthermore, Spo11α, one of major Spo11 isoforms, and Rad18, are important for XY pairing. In contrast to high expression of Spo11β in the early prophase, Spo11α, a smaller isoform of Spo11, is highly expressed in mid- to late prophase [52]. Importantly, the XY pairing takes place later in meiotic prophase than autosomal pairing. In fact, mice that lack Spo11α exhibit abnormal synapsis in sex chromosomes while autosomal homologous pairing and synapsis are normal, suggesting that Spo11α plays a role in XY synapsis [53]. RAD18, an E3 ubiquitin ligase, has an essential function in the repair of meiotic DSBs and loss of function of RAD18 also results in XY asynapsis [41]. Therefore, it is likely that down-regulation of Smc1β, Rec8, Stag3, Spo11, and Rec18 is at least partly responsible for the CKO phenotype. Whether Ikbkap directly regulates these genes in the male germ cells by facilitating their transcriptional elongation remains to be determined. Despite a predominant cytoplasmic location, a few studies have reported that in certain organisms, some subunits of the Elongator can localize to the nucleus [14], [15], [54], [55], [56], [57]. Moreover, studies in human cells have demonstrated that Elongator is preferentially recruited to the open reading frames of a number of genes [35], supporting a role for Ikbkap in transcription. Our results also suggest that Ikbkap positively regulates critical genes involved in synapsis and autosomal DSB repair. Due to the lack of chromatin immunoprecipitation (ChIP)-grade IKAP antibodies, we were unable to address whether IKAP directly contributes to transcription regulation by performing ChIP assays. Thus, we cannot exclude the possibility that IKAP contributes to transcription indirectly.
Accumulating evidence indicate that Elongator has an important role in tRNA modification, which has been well-documented in several model systems, including yeasts, nematode, and plants [18], [22], [23], [24]. However, it remains unknown whether this function of the Elongator is conserved to mammals. In this study, we present the first evidence demonstrating that Elongator complex is required for the formation of mcm5 and ncm5 side chains at wobble uridines of tRNA in mammalian cells, supporting its conserved function in all eukaryotes. The conservation of this function raises the question of whether the Elongator complex itself is directly involved in the wobble uridine tRNA modification. One previous study has shown that the S-adenosyl-methionion (SAM) binding domain present in Elp3 is able to transfer methyl groups to RNAs [58]. In addition, mutations in the conserved residues of the histone acetyl transferase (HAT) domains of Elp3 also affect tRNA-modifying activity [18]. Therefore, Elp3 appears to harbor at least two enzymatic activities. Interestingly, recent studies on the crystal structure of the Elp4–6 sub-complex have revealed its potential role in substrate recognition and tRNA modification [11], [59]. Elp4, Elp5, and Elp6 all share the same RecA-like protein fold, and Elp4/5/6 forms a hetero-hexameric conformation resembling hexameric RecA-like ATPase [11]. The ring-like structure of the sub-complex together with the hydrolysis of ATP are essential for its binding to the anti-codon stem-loops of tRNA as mutations in the homologous nucleic acid binding loop (L2) of Elp6 resulted in the loss of tRNA binding capacity [11]. Thus, it is possible that removal of IKAP may affect the integrity of the complex and thereby affecting its function in tRNA modification. This possibility is supported by studies in yeast and human cells demonstrating that deletion of Ikbkap leads to the loss of Elp3 as well as the integrity of the Elongator [35], [60]. Further structural and enzymological analyses of the Elp1–3 sub-complex and the Elongator holo-complex will help clarify the mechanism by which Elongator contributes to tRNA modification.
Interestingly, phenotypes observed in Elongator mutants, including those in RNAPII transcription and exocytosis, could be suppressed by overexpression of two tRNAs (tRNALysUUU and tRNAGlnUUG) in budding yeast [21], suggesting that the phenotypes are caused by lack of mcm5s2U modification on certain tRNAs. Therefore, it is likely that the primary effect of Ikbkap deficiency in germ cells is caused by tRNA modification defect, rather than dysregulation on transcription.
Mutations in the human Ikbkap genes have been shown to cause familial dysautonomia (FD) (also known as Riley-Day syndrome). FD is an autosomal recessive disease characterized by defects in the development and maintenance of autonomic and sensory nervous system [61], [62]. FD has been mainly associated with a single nucleotide substitution in the splice site of intron 20 of the Ikbkap gene, which ultimately leads to decreased expression of IKAP in a tissue-specific manner. Dietrich et al. has generated mice harboring exon 20 deletion allele (IkbkapΔ20) which phenocopy Ikbkap null mutations [26]. The mutants display severe cardiovascular phenotypes and die at E10 [26]. To circumvent the embryonic lethality of Ikbkap mutants, they further generated Ikbkap flox/flox mice with exon 20 floxed (referred to as Ikbkap floxE20/floxE20 hereafter). In contrast to our Ikbkap flox/flox mice, which are viable and normal, Ikbkap floxE20/floxE20 mice display low body weight, and skeletal and neuronal abnormalities. Biochemical analyses showed that Ikbkap floxedE20 allele results in severe reduction in expression of full-length IKAP protein [63]. IKAP expression in Ikbkap floxE20/floxE20 and Ikbkap Δ20/floxE20 brains is 10% and 5% that of wild-type mice, respectively. Interestingly, both models recapitulate the major phenotypic and neuropathological features, including optic neuropathy, seizures, ataxia, impaired development and maintenance of sensory and autonomic systems, reduced number of fungiform papillae on the tongue, gastrointestinal dysfunction as well as skeletal abnormalities [63]. Our study suggests that IKAP-mediated tRNA modification may play a role in the pathogenesis of FD. Characterization of a brain-specific Ikbkap knockout model may reveal how IKAP contributes to FD.
A mouse line harboring a FRT-flanked βGeo cassette upstream of loxP-flanked exon 4 of Ikbkap gene was obtained from Knockout Mouse Project (KOMP) Repository. Ikbkap flox/flox mice were generated by crossing mice carrying the Ikbkapβ-geo-flox allele to Rosa26R-FLP mice. Vasa-Cre transgenic mice were provided by Dr. Diego H. Castrillon [27]. All mouse strains were maintained in a mixed genetic background (129/Sv×C57BL/6) and received standard rodent chow. The primer sequences used for genotyping are listed in Table S1. Experimental animals and studies were approved by the Institutional Animal Care and User Committee (IACUC) of University of North Carolina at Chapel Hill.
Testis tissues were fixed with 4% paraformaldehyde (PFA), dehydrated, and embedded in paraffin. For histological analysis, sections (7 µm) were stained with hematoxylin and eosin (H&E) or periodic acid schiff's (PAS). For immunohistochemistry, deparaffinized sections after antigen retrieval were blocked with 5% donkey serum and a biotin-blocking system (Dako, http://www.dako.com/). The following antibodies were used: anti-IKAP (LSBio, https://www.lsbio.com/), anti-SCP1, anti-SCP3, anti-DMC1 (Abcam, http://www.abcam.com/), anti-PLZF, anti-GATA1, anti-3β-HSD (Santa Cruz Biotech, http://www.scbt.com/), anti-c-Kit (Cell Signaling Technology, http://www.cellsignal.com/), anti-γH2AX, anti-RAD51(this polyclonal antibody recognize both RAD51 and DMC1 [9]) (Millipore, http://www.millipore.com/), anti-MLH1, anti-Ki67 (BD Biosciences, http://www.bdbiosciences.com/), anti-Tra98 (Bio Academia, http://www.bioacademia.co.jp/en/), and anti-H1t (a gift from M.A. Handel, The Jackson Laboratory). Sections were washed with 0.1% Triton X-100/phosphate-buffered saline (PBST) buffer and incubated with biotinylated secondary antibodies (Jackson ImmunoResearch, http://www.jacksonimmuno.com/). Signal detection was carried out with the Avidin-Biotin Complex kit (Vector Laboratories, http://www.vectorlabs.com/) or Tyramide Signal Amplification system (TSA, Invitrogen, http://www.invitrogen.com/). Peroxidase activity was visualized with 3,3′-diaminobenzidine (DAB, Vector Laboratories). Nuclear staining was carried out with 4,6-diamidino-2-phenylindole (DAPI; Sigma-Aldrich, http://www.sigmaaldrich.com/), and sections were mounted with fluorescence mounting medium (Dako) prior to imaging. Images were captured with a Zeiss Axiophoto fluorescence microscope or a Zeiss laser-scanning confocal microscope with a spinning disk (CSU-10, Yokogawa).
TUNEL assays were performed on paraffin-embedded tissue sections using In Situ Cell Death Detection Kit (Roche, http://www.rocheusa.com/), following the manufacturer's instruction.
Testes were removed, decapsulated and shredded by needles in PBS, and the cell suspension was filtered with 100 mm mesh to remove the debris. The suspension was incubated with equal volume of a 2× hypotonic extraction buffer (30 mM Tris-HCl pH 8; 5 mM EDTA; 1.7% sucrose; 0.5% trisodium citrate) for 7 minutes. After centrifugation, cells were suspended with 100 mM sucrose solution to release hypotonized nuclei. A drop of nuclear suspension was spread onto slides that have been dipped in fixation solution (1% paraformaldehyde; 0.15% Triton X-100; 3 mM dithiothreitol (Sigma-Aldrich)). The slides were dried slowly in a humidified chamber for overnight, washed in 0.4% Photo-Flo solution (Kodak, http://www.kodak.com/), and dry again for storage. For immunofluorescent staining, the slides were permeabilized with 0.4% Triton X100/PBS for 20 minutes, rinsed with 0.1% tween 20/PBS, blocked with 5% donkey serum for 1 hour at room temperature (RT), and then incubated with primary antibodies at an optimized concentration overnight at 4°C. After wash, the slides were incubated with Alexa fluorophore conjugated secondary antibodies (Invitrogen) for 1 hour at RT. The standard protocol was followed as described above.
The stage of prophase I for each spermatocyte was determined by chromosomal morphology and sex body status. We use SYCP3 and γH2AX staining to visualize the chromosomal changes and XY body, respectively.
P15 juvenile testes from control or CKO mice (n = 3 for each genotype) were collected and total RNAs were extracted from mouse testes using Trizol (Invitrogen) and were cleanup using RNeasy kit (Qiagen). Samples were submitted to the UNC Functional Genomics Core Facility for RNA labeling, amplification, hybridization, and scanning. Samples were applied on Affymetrix Gene 1.0 ST assays (Affymetrix), and the procedures were followed according to the manufacturer's instructions. Data were analyzed and the expression patterns were presented as a scatter plot using GeneSpring GX software (Agilent Technologies).
Total RNAs were extracted from mouse testes using Trizol (Invitrogen) and were cleanup using RNeasy kit (Qiagen). The RNAs were treated with DNase I and first-strand cDNA were synthesized by SuperScriptIII reverse transcriptase using random hexanucleotide primers according to the manufacturer's instructions (Invitrogen). Quantitative RT-PCR analyses were carried out using the ViiA7 Real-Time PCR System (Applied Biosystems) and FastStart Universal SYBR Green Master (Roche Applied Science). All expression data were normalized to Gapdh. The primer sequences for RT-qPCR are listed in Table S1.
Mass spectrometric analysis of nucleosides was performed essentially as previously described [46], [64]. For sample preparation, 1 µg of total tRNA were heat-denatured, hydrolyzed with 90 U of Nuclease S1 (Sigma) in Buffer 1 (0.5 mM ZnSO4, 14 mM sodium acetate, pH 5.2) at 37°C for 1 hour (total volume is 44.5 µL), followed by the addition of 5 µL 10× Buffer 2 (560 mM Tris-Cl, 30 mM NaCl, 10 mM MgCl2, pH 8.3), 0.5 µg of phosphodiesterase I (Worthington) and 2 U of Calf Intestinal Alkaline Phosphatase (New England Biolabs) for an additional 1 hour (final volume 50 µL). The digested DNA was then filtered with Nanosep3K (Pall Corporation), and 10 µL of filtered samples were subjected to LC-MS/MS analysis using an UPLC (Waters) coupled to a TSQ-Quantum Ultra triple-quadrupole mass analyzer (ThermoFinnigan) using heat assisted electrospray ionization (HESI) in positive mode (spray voltage of 3000 V, API temperature of 250°C, sheath gas flow rate 35 arb, AUX gas flow rate 25 arb, capillary temperature of 285°C). Liquid chromatography (LC) was performed with a 2.1×100 mm HSS T3 1.8 µm column (Waters) with gradient elution at flow rate of 200 µl/min using 0.02% acetic acid in water as mobile phase A and methanol as mobile phase B. The gradient was 0→3.5 min, 3% B, 3.5→12.5 min, 3%→16.2%B, 12.5 →13 min, 16.2%B→30%B, 13→15 min, 30%B, 15→16 min, 30%→3%B, 16→20 min, 3%B. The eluent was directed to the mass spectrometer that was running in multiple reaction monitoring (MRM) mode, monitoring the transition of m/z 317.0 to 153.0 (mcm5U), m/z 302.0 to 153.0 (ncm5U), m/z 333.0 to 169.0 (mcm5s2U), m/z 261.0 to 129.0 (2-thio-U) and m/z 245.0 to 113.0 (U) for RNA samples.
We investigated reproductive capacities of VasaCre; Ikbkap flox/flox male mice by mating one male with two wild-type females for 3 months. Female mice were checked for vaginal plugs each morning, and the litter sizes were recorded.
Results are presented as mean ± SEM. Statistical analysis was carried out by Student's t test. The statistical analysis of boxplot was carried out by Kolmogorov-Smirnov test. P values less than 0.05 were considered statistically significant.
|
10.1371/journal.pgen.1005783 | Glycine and Folate Ameliorate Models of Congenital Sideroblastic Anemia | Sideroblastic anemias are acquired or inherited anemias that result in a decreased ability to synthesize hemoglobin in red blood cells and result in the presence of iron deposits in the mitochondria of red blood cell precursors. A common subtype of congenital sideroblastic anemia is due to autosomal recessive mutations in the SLC25A38 gene. The current treatment for SLC25A38 congenital sideroblastic anemia is chronic blood transfusion coupled with iron chelation. The function of SLC25A38 is not known. Here we report that the SLC25A38 protein, and its yeast homolog Hem25, are mitochondrial glycine transporters required for the initiation of heme synthesis. To do so, we took advantage of the fact that mitochondrial glycine has several roles beyond the synthesis of heme, including the synthesis of folate derivatives through the glycine cleavage system. The data were consistent with Hem25 not being the sole mitochondrial glycine importer, and we identify a second SLC25 family member Ymc1, as a potential secondary mitochondrial glycine importer. Based on these findings, we observed that high levels of exogenous glycine, or 5-aminolevulinic acid (5-Ala) a metabolite downstream of Hem25 in heme biosynthetic pathway, were able to restore heme levels to normal in yeast cells lacking Hem25 function. While neither glycine nor 5-Ala could ameliorate SLC25A38 congenital sideroblastic anemia in a zebrafish model, we determined that the addition of folate with glycine was able to restore hemoglobin levels. This difference is likely due to the fact that yeast can synthesize folate, whereas in zebrafish folate is an essential vitamin that must be obtained exogenously. Given the tolerability of glycine and folate in humans, this study points to a potential novel treatment for SLC25A38 congenital sideroblastic anemia.
| Mutations in the SLC25A38 gene cause an inherited anemia. In this study we determine that the function of SLC25A38, and its yeast homolgue Hem25, is to act as mitochondrial glycine importers providing a molecular explanation for why patients with SLC25A38 mutations have low hemoglobin levels and become anemic. Using this new knowledge, we go on to determine that supplementation with glycine and folate restore hemoglobin levels in a zebrafish model of the disease pointing to a potentially new, safe, and cost effective treatment for SLC25A38 congenital sideroblastic anemia.
| Sideroblastic anemias are a group of disorders principally defined by a decreased level of hemoglobin in erythrocytes (red blood cells) and the presence of pathological iron deposits in perinuclear mitochondria of erythroblasts (red blood cell precursors found in bone marrow) [1–5]. Sideroblastic anemias can be congenital or acquired with both primarily being due to a defect in heme/hemoglobin synthesis. One of the main reasons for acquired sideroblastic anemia is a nutritional deficiency in vitamin B6 (pyridoxine) as several of the enzymes required to synthesize heme and heme precursors require pyridoxal 5’-phosphate (PLP) as a cofactor. Alcohol abuse, copper deficiency, lead poisoning, some antimicrobial drugs, and myelodysplastic syndrome can also result in acquired sideroblastic anemia. Mutations in several genes cause congenital sideroblastic anemia (CSA) including ALAS2, SLC25A38, ABCB7, GLRX5, SLC19A2, PUS1, and YAR2.
The two most common types of CSA are an X-linked form due to mutations in ALAS2 and the more recently identified autosomal recessive form due to mutations in SLC25A38 [6–9]. ALAS2 and SLC25A38 are primarily expressed in erythroid precursor and red blood cells. ALAS2 is a PLP-dependent enzyme that catalyzes the first enzymatic step of the heme/hemoglobin biosynthesis pathway utilizing glycine and succinyl-CoA to synthesize 5-aminolevulinic acid (5-Ala). A subset of ALAS2 CSA patients, those with mutations that decrease PLP binding, can be treated with high levels of pyridoxine. ALAS2 CSA patients with mutations outside of the PLP binding region, and all SLC25A28 CSA patients, are refractory to pyridoxine treatment.
Pyridoxine refractory CSA patients suffer severe clinical consequences including a microcytic transfusion-dependent anemia that usually appears in infancy resulting in sequelae typical of chronic transfusion therapy, and can suffer significant long term morbidity and mortality related to iron overload [10]. Recently, the adoption of effective and tolerable oral iron chelation therapies predict an increase in life expectancy comparable to that found for transfusion-dependent adequately chelated patients with hemoglobinopathies [11]. Despite these advances, oral iron chelators carry their own risks [12] and lifetime transfusion is associated with high financial quality of life burdens, with additional medical complications including alloimmunization and acquired infectious agent transmission including hepatitis B and C [13]. There is a clear need to decrease transfusion dependence for CSA patients. Here, we employ yeast and zebrafish as complementary preclinical models to determine the function of SLC25A38 and go on to propose a potential therapy for SLC25A38 CSA patients.
The function of SLC25A38 is not known (Fig 1A). To determine how SLC25A38 mutations cause CSA we sought to determine its function using a yeast model. The Saccharomyces cerevisiae SLC25A38 homologue YDL119c, which we name HEM25 (Heme synthesis by SLC25 family member), was inactivated in the yeast genome and the level of heme was determined. The hem25Δ yeast cells exhibited a 50% decrease in the level of heme (Fig 1B). Human SLC25A38 was expressed in hem25Δ cells in order to determine if the human protein could complement the absence of the yeast protein. Heme level was restored to normal upon expression of the human SLC25A38 protein in yeast cells with an inactivated HEM25 gene indicating conservation of function between the yeast and human proteins. SLC25A38 is a member of the mitochondrial SLC25 family which are subdivided into keto acid, adenine nucleotide, and amino acid carriers [14]. Phylogenetic analysis of human SLC25A38 grouped SLC25A38 with the amino acid carriers (S1 Fig). This is consistent with previous predictions that SLC25A38 could be a mitochondrial glycine or serine transporter required for the synthesis of heme [6].
We first sought to determine if Hem25 is a mitochondrial glycine importer. To do so, we used two separate scenarios where the efficient uptake of glycine by mitochondria was required for yeast cell growth. In the first scenario, we took advantage of the fact that glycine can be used as the sole nitrogen source by yeast cells, but only if glycine is efficiently imported into mitochondria for conversion to NH3 by the glycine cleavage system (GCV) (Fig 2A). The ability of hem25Δ cells to grow with glycine as the sole source of nitrogen was determined. Inactivation of the HEM25 gene impaired the ability of cells to grow when glycine was the sole nitrogen source, although not to the extent observed for inactivation of an enzyme of the GCV, dihydrolipoamide dehydrogenase encoded by LPD1 (Fig 2B). The results are consistent with Hem25 being an important importer of glycine into mitochondria.
In the second scenario, we used the observation that in yeast cells with an inactivated SER1 gene [15] glycine supports the growth of ser1Δ cells but only if glycine can be efficiently imported into the mitochrondria. In this context, glycine becomes the major source of one carbon units via its catabolism by the mitocondrial glycine cleavage system (GCV) to produce CH2-THF, and of serine through mitochondrial serine hydroxymethyltransferase which itself requires CH2-THF (Fig 2C) [16–19]. To determine if Hem25 is required for effective glycine import into mitchondria, the HEM25 gene was inactivated in a ser1Δ background and the ability of ser1Δ and ser1Δ hem25Δ strains to grow in the presence of serine or glycine was determined. As expected, serine supported the growth of all strains (Fig 2D). Congruent with the dependence on GCV for the generation of one carbon units and subsequent serine synthesis, ser1Δcells grew poorly on non-supplemented medium. Importantly, the absence of Hem25 in ser1Δ cells exacerbated the growth defect. Glycine supported growth of ser1Δ cells, however, ser1Δ hem25Δ cells grew poorly in medium containing low concentrations of glycine (Fig 2D), consistent with Hem25 being required to efficiently import glycine into mitochondria. Interestingly, as the levels of glycine were titrated upward, there was an increase in growth of the ser1Δ hem25Δ cells, implying other mitochondrial glycine transporter(s) (of lower glycine affinity) may exist.
To determine if other members of the yeast SLC25 family member could be secondary glycine transporters, each gene encoding a putative amino acid carrier that was a member of the SLC25 family was inactivated simultaneously with the HEM25 gene. The ability of each single and double mutant to synthesize heme in the presence of glycine or 5-Ala was determined. If any of these SLC25 family members encode a secondary glycine transporter, then 5-Ala, but not high levels of glycine should restore heme levels to hem25Δ cells also lacking a specific SLC25 family member. This was observed for hem25Δ ymc1Δ cells. The double hem25Δ ymc1Δ mutants had a significant reduction in heme content compared with the single ymc1Δ or hem25Δ mutants (Fig 3), consistent with both genes being on parallel pathways contributing to the same downstream product, in this case heme. Double mutant cells grown on media supplemented with 5 mM glycine did not show an increase in heme content whereas double mutant cells supplemented with 5-Ala showed an increase in heme content, although not to wild type levels.
The data strongly support the notion that SLC25A38/Hem25 is a glycine transporter required to synthesize heme. However, it is also clear that in yeast Hem25 is not the sole mitochondrial glycine transporter and secondary glycine transporter(s) exist. We identify Ymc1 as a putative secondary mitochondrial glycine transporter that contributes to heme synthesis when glycine concentrations are high.
We next sought to determine the de novo source(s) of glycine used for heme synthesis. The genes encoding enzymes that synthesize glycine under glucose grown conditions are GLY1 encoding a cytoplasmic threonine aldolase, SHM2 encoding a cytoplasmic serine hydroxymethyltransferase, and SHM1 encoding a mitochondrial serine hydroxymethyltransferase (Fig 4A). The GLY1, SHM1 and SHM2 genes were inactivated and the levels of glycine and heme determined in gly1Δ, shm2Δ, and shm1Δ single mutant cells. Inactivation of the GLY1 gene, encoding cytoplasmic threonine aldolase, reduced cellular heme by 75% and glycine mass by 90% (Fig 4B). Upon inactivation of either the cytoplasmic or mitochondrial serine hydroxymethyltransferases, a 50% decrease in glycine was observed, however, the level of heme level was not significantly different from that of wild type cells. As Gly1 produces glycine in the cytoplasm, this is consistent with a cytoplasmic source of glycine being the main contributor to heme synthesis through its import by the mitochondrial SLC25A38/Hem25 glycine transporters for subsequent use as substrate in the first enzymatic step in the synthesis of heme catalyzed by ALAS2/Hem1.
A second hypothesis for the function of SLC25A38/Hem25 was as a serine transporter. This was based on the metabolic pathway whereby serine, once imported into the mitochondria, can be converted to glycine by mitochondrial serine hydroxymethyltransferase (Shm1) for subsequent use by ALAS2 (Hem1 in yeast) to initiate heme synthesis (Figs 1A and 4A). If this was the case, then loss of function of Shm1 should lead to a significant decrease in the level of heme, however, in shm1Δ cells heme level was similar to wild type (Fig 4B), as opposed to the profound effect on heme level observed in hem25Δ cells (Fig 1B).
To further investigate if Hem25 could be a serine transporter, we took advantage of the fact that it has been reported that shm1Δ cells grow with rates similar to wild type cells whereas shm2Δ cells exhibit a growth impairment associated with a shortage of one-carbon units that can be alleviated by formate supplementation (formate is transported between the mitochondria and cytoplasm as part of the folate cycle (Fig 2C) [15]. Consistent with this observation, there are two studies that showed that shm2Δ shm1Δ cells have impaired growth compared to shm2Δ cells [15,20] with the addition of exogenous formate restoring growth to wild type level. If Hem25 were a serine transporter, the growth phenotype of the shm2Δ hem25Δ cells should be worse than shm2Δ cells as the absence of Hem25 would prevent the entry of serine into the mitochondria and restrict the generation of one-carbon units by Shm1. However, there was no difference in growth of shm2Δ versus shm2Δ hem25Δ cells (Fig 4C). As the absence of Hem25 does not aggravate the growth phenotype of shm2Δ cells this suggests that Hem25 is not a serine transporter.
The third experiment we performed to determine if Hem25 was a serine importer was based on the observation of McNeil [15] et al that reported that shm2Δ gcv1Δ cells had a slight growth defect, which was dramatically worsened by inactivation of SHM1. Both strains had their growth rates restored to wild type by formate addition. If Hem25 was involved in serine import, shm2Δ gcv1Δ hem25Δ cells should grow slower compared to shm2Δ gcv1Δ cells as the absence of Hem25 would deprive Shm1 of its substrate serine, limiting the generation of one-carbon units. We determined if there was a difference in growth for shm2Δ gcv1Δ isolates and shm2Δ gcv1Δ hem25Δ strains and observed that the double and triple mutants strains grew at similar rates (S2 Fig), again indicating that Hem25 does significantly not contribute to the import of serine into the mitochondria for its subsequent conversion to glycine by Shm1.
We hypothesized three scenarios to ameliorate the defect in heme synthesis in hem25Δ cells (i) supplementation with exogenous glycine to increase substrate availability for the first step in heme synthesis, (ii) addition of excess serine to drive endogenous glycine synthesis in mitochondria, or (iii) addition of the downstream metabolite, 5-Ala, within the heme biosynthesis pathway. Each compound was added to wild type, hem25Δ, and hem1Δ cells. Serine had no effect on the growth of any of these strains. The addition of 5-Ala restored heme to normal levels in hem1Δ and hem25Δ cells, while only the addition of exogenous glycine restored heme level to normal in hem25Δ cells (Fig 5A). These latter two results are consistent with Hem25 lying upstream of Hem1 in the synthesis of heme. We also determined the level of heme in the hem25Δ gly1Δ cells and observed that it was not substantially diminished beyond that observed in gly1Δ cells alone (Fig 5B), consistent with two proteins on a linear pathway synthesizing the same downstream product.
Interestingly, we noted a growth defect for gly1Δ and hem25Δ gly1Δ cells that could be differentially restored by the addition of 5-Ala (Fig 5C). For the single gly1Δ single mutant, 5-Ala did not increase growth whereas glycine restored growth to wild type levels. Thus, although the level of heme is depleted in gly1Δ cells its level does not limit cell growth, as 5-Ala does not restore growth. Some other glycine dependent process must be limiting gly1Δ cell growth when heme levels are normal. In the absence of supplements, the growth of the hem25Δ gly1Δ double mutant was slower than either single mutant alone. The addition of glycine or 5-Ala to the hem25Δ gly1Δ double mutant restored heme to normal levels (Fig 5D). Similar to gly1Δ cells, 5-Ala supplementation partially restored cell growth while glycine completely restored cell growth, indicating that growth impairment of the double mutant is only partly due to a decrease in heme level.
In the absence of exogenous glycine, cytoplasmic Gly1 is the major supplier of glycine for the synthesis of heme, consistent with the requirement for a high affinity mitochondrial glycine transporter being required for heme synthesis. When Gly1 function is inactivated glycine becomes limiting for cellular functions resulting in impaired growth, with simultaneous loss of Hem25 function resulting in a more severe growth defect. The growth defect of gly1Δ and hem25Δ gly1Δ cells cannot be restored through restoration of heme synthesis alone through the addition of 5-Ala, but can be restored by the addition of glycine, implying other glycine dependent functions are what impair cell growth.
The zebrafish homologue of SLC25A38 is duplicated in the zebrafish genome, like many zebrafish genes. To further understand the loss-of-function phenotypes of both slc25a38a and slc25a38b, we studied their expression by in situ hybridization at 24 and 34 hrs post-fertilization (hpf), time points at which each of the two waves of definitive hematopoiesis occur in zebrafish [21]. At 24 hpf, slc25a38b was expressed predominantly in the posterior blood island, posterior cardinal vein and circulating blood. By contrast, slc25a38a was also expressed in somites, brain and retina at 24 hpf (Fig 6A and S3 Fig). At 34 hpf expression patterns of both genes became more strongly restricted to the posterior blood island and circulating blood cells.
Morpholino knockdown of both slc25a38a and slc25a38b was required to produce an anemia in zebrafish embryos as determined using o-dianisidine staining to assess hemoglobin content (Fig 6B and [6]). Specificity of the morpholino oligos used was controlled by using five mismatch versions, as well as a control morpholino, whose injection resulted in normal phenotypes of the resulting morphants (Fig 6B). We observed a 50% decrease in hemoglobin level in slc25a38a+b morphants and larger more immature appearing cells with less compact nuclei. This was despite the absence of pathological sideroblasts in erythroid cells (S4 Fig). This finding is in keeping with other zebrafish models of CSA, such as the mutant sauternes (sau), which results in loss-of-function of the zebrafish orthologue of ALAS2 [22]. The absence of ringed siderocytes in the zebrafish may be a result of the early time point in embryogenesis in which the cells are analyzed without sufficient time for either endogenous iron overload, or excess iron that accumulates from transfusion therapy in patients.
We next determined if the addition of glycine or 5-Ala could rescue the capacity of slc25a38a+b, or alas2 knockdown fish, to synthesize heme. Optimization of glycine and 5-Ala dosing was performed in wild type zebrafish embryos by toxicity studies evaluating the frequency of death and developmental abnormalities relative to the dose of drug, a dose of ~50% of the maximum tolerated dose was chosen for treatments (S1 Table). Glycine or 5-Ala was added to the egg water of slc25a38a+b or alas2 morphants, and o-dianisidine staining was used to assess hemoglobin content. Interestingly, neither glycine nor 5-Ala was able to restore hemoglobin level in alas2 or slc25a38a+b morphants (Fig 6B and 6C and S5 Fig).
We know from our work with yeast lacking Hem25 function that decreasing the de novo capacity to synthesize glycine impaired cell growth implying glycine dependent processes beyond an inability to synthesize heme can impact cell fitness. One of these downstream glycine dependent processes is folate synthesis. Interestingly, yeast that can synthesize folate de novo while higher eukaryotes including zebrafish and humans have a dietary requirement for folate. With this in mind, we sought to determine if the addition of exogenous glycine plus folate together could ameliorate the zebrafish model of CSA. When glycine and folate were added in combination to slc25a38a+b morphants, the level of hemoglobin was increased to 80% that observed in normal zebrafish (Fig 6B and 6C).
We determined that Hem25, the yeast homologue of human SLC25A38, is a mitochondrial glycine transporter that plays an important role in providing glycine for the first enzymatic step in heme synthesis, which is catalyzed in the mitochondria by Hem1 in yeast and ALAS2 in red blood cells and their precurors in humans. The function of Hem25 was determined through assessing phenotypes in yeast cells lacking Hem25 function where the import of glycine into mitochondria was required for yeast cells to grow and included the use of glycine as a sole nitogen source and glycine as a source of one carbon units. We also determined that Hem25 is not a mitochondrial serine importer using three lines of evidence. First, inactivation of gene encoding the mitochondrial enzyme that converts serine to glycine, SHM1, did not decrease heme levels; second, inactivation of the SHM1 gene in cells lacking Shm2 function results in a growth defect due to a decreased capacity to synthesize one carbon units, while inactivation of the HEM25 gene in cells lacking Shm2 function did not decrease growth; and third, similarly to cells lacking Shm1, cells lacking the GCV also grow slowly in cells lacking Shm2 function due to an inability to synthesize one carbon units. Inactivation of HEM25 in cells lacking both Shm2 and the GCV did not further decrease growth. Expression of human SLC25A38 in yeast cells where the HEM25 was gene deleted was able to restore heme to a normal level indicating conservation of function. We conclude that Hem25 and SLC25A38 are mitochondrial glycine importers. Future directions will be biochemical glycine transport experiments to determine substrate affinity, and whether SLC25A38 is a facilitative transporter or if there is a secondary active metabolite.
The conclusion that Hem25 and SLC25A38 are mitochondrial glycine importers was boltered by our work that determined the de novo source of glycine for heme synthesis in yeast. We found that Gly1, a cytosolic threonine aldolase, was the main source of glycine for heme synthesis, while the mitochondrial and cytosolic serine hydroxymethyltransferase, Shm1 and Shm2 respectively, did not significantly contribute to heme synthesis. A cytosolic source of de novo glycine as the main source of glycine for heme synthesis is consistent with the requirement for a high affinity glycine importer for subsequent synthesis of heme. As expected from this conclusion, loss of function of Gly1 along with Hem25 did not further decrease heme level as both are on a linear pathway for the synthesis of heme. Interestingly, loss of function of Gly1 along with Hem25 did result in a growth defect. This growth defect was partially restored by the addition of 5-Ala and completely restored by glycine supplementation. This indicates that the growth impairment of the double mutant is partly due to a decrease in heme level and partly due to second important role for glycine in the mitochondria such as the synthesis of one carbon units and/or the synthesis of serine. In line with this conclusion was our observation that there was no growth impairment in cells lacking only Hem25 function, and both glycine and 5-Ala restored heme level to normal in cells lacking Hem25.
We subsequently determined if glycine or 5-Ala could increase the level of hemoglobin in a more complex vertebrate model of CSA. We employed the zebrafish model as a substantial number of human blood disorders including inherited anemias have been accurately recapitulated in zebrafish [23,24]. Like many genes, SLC25A38 is duplicated in the zebrafish genome thus translation-blocking morpholinos were designed to target slc25a38a+slc25a38b morphant embryos, which demonstrated reduced hemoglobin content. Surprisingly, the addition of glycine or 5-Ala at half of the maximum tolerated dose did not restore hemoglobin levels. In the case of 5-Ala, a higher dose may be required as 2 mM 5-Ala was found to rescue morphants where the expression of the protein required for PLP binding to alas2 (clpxa) was decreased [25], whereas we used a dose of 0.3 mM as we observed some developmental abnormalities occuring with doses of 5-Ala above 1 mM. Another explanation is that clpxa may not provide as complete a block in 5-aminolevulinic acid synthase activity as compared to reducing the expression of alas2 itself. In addition, 5-Ala was unable to restore growth to some yeast strain backgrounds lacking genes beyond HEM25, implying that glycine metabolism requirements for cells beyond heme synthesis can impact the capacity of 5-Ala to restore cell fitness. Surprisingly, the addition of glycine was also unable to increase the level of hemoglobin in slc25a38a+b morphants, especially as glycine restored heme and growth to yeast lacking HEM25 and yeast strains lack HEM25 in combination with impaired glycine metabolism.
The inability of glycine to increase hemoglobin levels in zebrafish prompted us to consider differences between yeast and vertebrate organisms that might impact glycine metabolism. A major difference is the ability of yeast to synthesize their own folate, whereas in zebrafish and higher eukaryotes folate is a vitamin that must be consumed in the diet. Mitochondria contain up to 40% of total cellular folate [26,27], and coupled with the fact that folate and its metabolites are normally found near the level required for use in folate dependent catalysis, a change in folate availability can have significant consequences [18]. It is also known that an inability to import folate into the mitochondria of Chinese hamster ovary cells results in a requirement for glycine for cell survival [28,29]. Finally, the CH2-THF dependent synthesis of glycine from serine by serine hydroxymethyltransferases has an essential role in one-carbon metabolism in mitochondria in mammalian cells [30–32]. We found that the addition of glycine plus folate substantially increased hemoglobin level in the zebrafish SLC25A38 congenital sideroblastic anemia model. In the absence of slc25a38a+b function the ability to supply mitochondrial glycine becomes limiting. As de novo mitochondrial glycine synthesis requires folate derivatives, the addition of folate would increase mitochondrial glycine, providing glycine for the production of heme/hemoglobin (S6 Fig).
Human red blood cells rely primarily on the CH2-THF dependent serine hydroxymethyltransferases enzymes SHMT1 and SHMT2 for endogenous glycine synthesis [33,34], as the human threonine aldolase (GLY1) homologue is a non-functional pseudogene [35], and an alternate source of glycine found in many eukaryotes threonine dehydrogenase (including mice and zebrafish, but not yeast) is also a non-functional pseudogene in humans [36]. This implies an increased reliability on exogenous glycine and folate in humans. The substitution of medically, socially and financilly burdensome chronic transfusion therapy with commercially established and available glycine and folate is currently being explored as a potential innovative therapy for SLC25A38 CSA patients on a clinical protocol that has been approved by our institutional research ethics board.
All zebrafish studies were approved by the Dalhousie University Committee on Laboratory Animals (Protocol #11–130) and conducting in accordance with the Canadian Council on Animal Care guidelines.
Yeast cells were maintained in YEPD (1% bacto-yeast extract, 2% bacto-peptone, 2% dextrose) medium or SD (0.67% bacto-yeast nitrogen base without amino acids, 2% dextrose) medium supplemented as required for plasmid maintenance and nutrient auxotrophies. Yeast cells were grown at 30°C.
The yeast strains used in this study are shown in S2 Table. Single deletion mutants in the BY4741 background were obtained from Euroscarf. The genetic KanMX6 marker was replaced with the nourseothricin acetyltransferase (Nat) gene or hygromycin resistance gene cassette using plasmid pAG25 or pAG32, respectively. Gene deletions were transfered from the BY4741 strains into the W303 background by PCR amplification of the corresponding selectable marker disrupted gene flanked at both ends by ~0.2 Kb of yeast genomic DNA, followed by transformation of W303 yeast cells and selection and characterization of transformants for the gene deletion by genomic PCR. Genomic PCR was used to verify all strain constructions. To screen for a putative second glycine importer among the members of the SLC25 family double gene deletion strains were constructed in the BY4741 background by standard genetic crosses of hem25Δ::NAT with the thirty one single gene deletion strains carrying mutations for each member of the yeast SLC25 family, followed by sporulation and haploid cell selection. Routinely, cells from the BY4741 background express a wild type allele of HAP1 from a plasmid. The human SLC25A38 open reading frame was PCR amplified from a B lymphocyte cDNA library and subcloned into the yeast expression vector p416-GPD. The DNA sequence of the SLC25A38 open reading frame was confirmed by Sanger sequencing.
Logarithmically growing cells were harvested, washed in ice-cold water, and resuspended in 10 mM Tris-HCl, pH 8, 150 mM NaCl. Cells were lysed by glass bead beating for two periods of 1 min intercalated with 1 min on ice. Cell debris was removed by centrifugation at 500 × g for 3 min, and heme and protein content were assayed on the supernatants. Heme determination was carried out with the Hemin Assay Kit from BioVision using a Thermo Labsystems Multiskan Ascent Microplate Reader. Protein was determined by a modified Lowry method.
Free glycine content in yeast cells was estimated by the ninhydrin method. Briefly, yeast cells grown into log phase in SD media were harvested, washed with ice-cold water and resuspended at a density of 4 x108 cells per ml of 10% sulphosalicylic acid containing 0.5 mM norleucine. Cells were broken by glass bead beating for 4 minutes and whole cell extracts were clarified by centrifugation at 18,000 x g for 30 min. Cell extracts were subjected to amino acid quantification using the Biochrom 30 Amino Acid Analyzer.
Casper zebrafish (Danio rerio) were raised and mated using routine procedures. Antisense morpholinos (MO) were ordered from Gene Tools. Antisense MOs targeted to the region immediately at the translational start site, with sequences 5′-CCGGATGAGCCACAGAGAACTCCAT-3′ for slc25a38a and 5′-CAGGATGAGCCAGGGCAACTTCCAT-3′ for slc25a38b. The 5-mismatch negative control morpholinos for slc25a38a were CCcGATGAcCCACAcAGAAaTCaAT MO and for slc25a38b were CAGcATGAcCCAGcGCAAaTTCaAT, the MO 5'-AACGAACGAACGAACGAACGAACGC-3’ was also used as an additional negative control. An antisense MO blocking the zebrafish alas2, 5′-CAGTGATGCAGAAAAGCAGACATGA-3′, was used as a positive control for phenotypes assocaited with decreased heme/hemoglobin synthesis. For each, 1.5 nl of MO was microinjected into casper embryos at the one-cell stage at a concentration of 0.5 mM, the maximum dose of MO that resulted in the least overt developmental abnormalities during the first 48 h of embryonic development. Attempts to detect slc25a38a or slc25a38b encoded protein using several commercially available antibodies were unsuccessful. In addition, we oberved that injection of mRNA coding for zebrafish slc25a38a+b, or human SLC25A38, at increasing doses resulted in either no phenotype or embryo death making rescue experiments impossible.
Optimization of glycine and 5-Ala dosing was performed in wild type zebrafish embryos by toxicity studies evaluating the frequency of death and developmental abnormalities relative to the dose of drug, with 50% of the maximum tolerated dose chosen for treatments which were 100 mM for glycine and 1 nM for 5-Ala. At 24 hours postinjection (hpi) with MO, glycine or 5-Ala was added to egg water of the slc25a38a+b, alas2, and control morphants. Folic acid was converted to its sodium salt and added to fish water to give a final concentration of 1 mM at 6 hpf and then exchanged once at 24 hpf. The folate level used was that previously used to fully ameliorate selenite toxicity in zebrafish embryos [37]. Embryos were subsequently stained with o-dianisidine solution.
At 48 hpi, embryos were decorionated and anesthetized with 0.02% Tricaine. The dechorionated embryos were stained in the dark for 15 min in o-dianisidine solution (0.6 mg/ml) containing 0.65% hydrogen peroxide, 40% ethanol, 10 mM sodium acetate, pH 4.5) for 15 min, fixed in 4% parafomaldehyde in phosphate buffered saline overnight at 4°C, and then embedded into 2.0% (w/v) low-melting point agarose for imaging.
RNA in situ hybridization for zebrafish slc25a38a and slc25a38b genes was performed as previously described[38]. In vitro transcription was performed using the Roche DIG RNA labeling kit. Probes were generated from PCR products amplified directly from an oligo-dT primed 24 hpf zebrafish cDNA preparation using primers specific for slc2538a and b
Transgenic gata1::EGFP embryos were dissociated to single cells as described previously[39] and then sorted using the FACS Aria I at the IWK Health Centre FACS Core Facility. Positive and negative sorted cells were put on ice immediately after collection, centrifuged and immediately processed for RNA extraction. The purity of sorted cell populations was verified by a repeated FACS analysis of sorted cells.
RNA for qPCR experiments was extracted from 100,000 to 500,000 FACS-sorted cells, genomic DNA was removed, and cDNAs for qPCR assays were generated using 9-mer random primers and M-MuLV reverse in the presence of Protector RNAse Inhibitor. For SYBR Green I-based expression quantification we used the QuantiFast SYBR Green RT-PCR Kit (Qiagen). qPCR gene expression assays for hbbe1, hbbe3, slc25a38a and slc25a38b and 18S ribosomal RNA were performed.
|
10.1371/journal.pntd.0004341 | Allopurinol Resistance in Leishmania infantum from Dogs with Disease Relapse | Visceral leishmaniasis caused by the protozoan Leishmania infantum is a zoonotic, life threatening parasitic disease. Domestic dogs are the main peridomestic reservoir, and allopurinol is the most frequently used drug for the control of infection, alone or in combination with other drugs. Resistance of Leishmania strains from dogs to allopurinol has not been described before in clinical studies.
Following our observation of clinical disease relapse in dogs under allopurinol treatment, we tested susceptibility to allopurinol of L. infantum isolated from groups of dogs pre-treatment, treated in remission, and with disease relapse during treatment. Promastigote isolates obtained from four treated relapsed dogs (TR group) showed an average half maximal inhibitory concentration (IC50) of 996 μg/mL. A significantly lower IC50 (P = 0.01) was found for isolates from ten dogs before treatment (NT group, 200 μg/mL), as well as for five isolates obtained from treated dogs in remission (TA group, 268 μg/mL). Axenic amastigotes produced from isolates of the TR group also showed significantly higher (P = 0.002) IC50 compared to the NT group (1678 and 671 μg/mL, respectively). The lower sensitivity of intracellular amastigotes from the TR group relative to those from the NT group (P = 0.002) was confirmed using an infected macrophage model (6.3% and 20% growth inhibition, respectively at 300 μg/mL allopurinol).
This is the first study to demonstrate allopurinol resistance in L. infantum and to associate it with disease relapse in the canine host. These findings are of concern as allopurinol is the main drug used for long term control of the disease in dogs, and resistant L. infantum strains may enhance uncontrolled transmission to humans and to other dogs.
| Visceral leishmaniasis caused by Leishmania infantum is a widespread, devastating zoonotic disease. Domestic dogs are considered the major reservoir for human L. infantum infection in an area stretching from Portugal to China, and across South America. Dogs not only carry infection and transfer it to humans via sand flies, but frequently suffer from a severe and fatal disease. Allopurinol is the main drug used for treatment of dogs with leishmaniasis, by itself or in conjunction with other drugs. Following treatment, most dogs experience clinical remission, but this is often not followed by parasite elimination. We found that dogs suffer from disease relapse during treatment with allopurinol, and showed that L. infantum promastigotes isolated from treated relapsed dogs were roughly four times less susceptible to allopurinol as compared to isolates from dogs before treatment or from dogs that resolved their clinical disease during treatment. We also confirmed that resistance to allopurinol was present in axenic and intracellular amastigotes. This study describes resistance of L. infantum clinical isolates associated with disease relapse. Allopurinol resistance may increase L. infantum transmission to humans and to other dogs, as canines with uncontrolled infection are highly infectious to sandflies.
| Visceral leishmaniasis caused by L. infantum is a potentially fatal disease, which is a serious public health concern in Europe, Asia, North Africa and Latin America. The dog is the main reservoir for this zoonotic infection and it has been estimated that in Southwestern Europe alone there are about 2.5 million infected dogs [1,2].
Allopurinol is a purine analog used mainly as a xanthine oxidase inhibitor to reduce serum urate concentration and is prescribed for the management of gout in humans [3]. Allopurinol’s anti-leishmanial activity was first described in 1974 by Pfaller and Marr [4], and is attributed to the inhibition of the leishmanial enzyme hypoxanthine-guanine phosphoribosyl transferase (HGPRT). HGPRT takes part in the parasite's purine salvage pathway, converting dephosphorylated purines to nucleoside monophosphates. Phosphorylated allopurinol is most likely incorporated into nucleic acids, leading to disrupted protein translation and selective parasite death [5,6]. Allopurinol has limited use as an agent for treatment of human visceral and cutaneous leishmaniasis, and it was rarely administered in conjunction with amphotericin B or pentavalent antimonials [1,7]. In veterinary medicine however, it is considered the major first line drug for long term treatment of canine leishmaniasis (CanL), often in combination with pentavalent antimonials or miltefosine for the first month and then continued alone [8,9]. Among its merits are wide availability, low cost, good safety profile, lack of known resistance, and the fact that it is rarely used for the treatment of human leishmaniasis. For the latter reason, allopurinol is the only drug recommended by the World Health Organization for the treatment of CanL [1]. Since antimonials and miltefosine are not commercially available for treatment of CanL in Israel, allopurinol is currently the only drug used against this disease in dogs.
Long term allopurinol treatment, whether combined with an initial course of meglumine antimoniate or miltefosine [8,10,11], or administered alone [12,13], results in most cases in improvement of the dog's clinical disease signs within a few weeks and a reduction in the dog's infective potential. Despite that, cases of disease relapse following secession of combined or exclusive allopurinol treatment [14,15,16] or even during treatment [17,18] have been recorded. Despite a growing concern over the development of resistance to other antileishmanial drugs in humans and increased efforts to uncover the genetic and biochemical basis of resistance mechanisms [19,20,21,22,23,24], very little information is available on resistance to anti-leishmanial drugs in dogs and no direct association has been documented between disease relapse and drug resistance [25,26,27]. In this study, following our observation of clinical disease relapse in dogs under allopurinol treatment, we tested susceptibility to allopurinol of L. infantum strains isolated from relapsed and non-relapsed dogs aiming to explore the possibility of drug resistance.
Nineteen privately owned dogs with natural L. infantum infection were included in the study. The inclusion criteria were: diagnosis of leishmaniasis as detailed under “diagnosis of leishmaniasis”, documented medical history from diagnosis, no concurrent non-Leishmania related medical conditions and owners consent to participate in the study. Dogs were allocated to one of three groups, based on a clinical score [28] established for each dog on its enrollment to the study: (1) ten newly diagnosed dogs with moderate to severe clinical signs (clinical score >2) due to L. infantum infection, prior to initiation of allopurinol treatment (non-treated; NT); (2) five previously diagnosed dogs with clinical disease undergoing allopurinol treatment and currently in remission for 3 months or longer (clinical score >2) (treated-asymptomatic; TA); and (3) four previously diagnosed dogs undergoing allopurinol treatment, with recurrence of clinical signs (clinical score >2) after being disease free for 3 months or longer (treated-relapsed; TR). Dogs belonging to the TA and TR groups were treated with allopurinol at 10 mg/kg orally every 12 hours from diagnosis onward, with no intermissions. None of the dogs experienced any previous episodes of leishmaniasis or was treated with allopurinol for any indication before the described diagnosis, nor did they receive any other anti-leishmanial drug at any stage. For dogs in the TR and TA groups, clinical score was established for time of disease diagnosis based on information in their medical records. Complete blood count, albumin, total protein, creatinine and urea were measured for each dog on enrollment to the study.
Diagnosis of leishmaniasis was made in all dogs based on the presence of compatible clinical signs, positive ELISA serology, and parasite isolation in culture followed by molecular characterization. Amplification of L. infantum DNA from blood by PCR was attempted for each patient. Serology was tested on crude L. infantum antigen ELISA as previously described [29] with a cutoff value set at 0.6 optical density (OD). Real time PCR was used to detect and quantify Leishmania DNA in blood. DNA was extracted (illustra, GE Healthcare, UK) from 200μL blood and a 120 bases long leishmanial kDNA fragment was amplified by PCR using primers JW11 (5-CCTATTTTACACCAACCCCCAGT-3) and JW12 (5-GGGTAGGGGCGTTC-TGCGAAA-3) [30]. Reactions were done in a total volume of 20μL, including; 10 μL Fast SYBR Green Master Mix (x2) (Applied Biosystems, Foster City, CA), 1μL DNA, ultra-pure water and a final concentration of 500nM of the forward and reverse primers. Reactions were carried out using the StepOnePlus real-time PCR thermal cycler (Applied Biosystems, Foster City, CA) with the following thermal profile; initial denaturation for 20 seconds at 95°C, followed by denaturation for 3 seconds at 95°C and annealing for 30 seconds at 59°C for 40 cycles. Amplicons were subsequently subjected to a melt step with the temperature raised to 95°C for 15 seconds and then lowered to 60°C for 1 min. The temperature was then raised to 95°C at a rate of 0.3°C per second. Amplification, melt profiles and quantity based on analysis of standard curve values, were analyzed using the StepOne V2.2.2 software (Applied Biosystems, Foster City, CA). Samples were tested in duplicates including negative (Leishmania-free dog DNA) and positive L. infantum (MCAN/IL/2002/Skoshi) controls. Only samples with CT value below 35 and a melt curve matching that of the positive control were considered positive. Quantitation of leishmanial DNA in the samples was made by comparing samples to a calibration curve. Leishmania DNA standards representing 100−106 parasites DNA per μL were produced by mixing 200μL of Leishmania free dog blood with 2*102−2*108 cultured L. infantum promastigotes, followed by DNA extraction as described above.
Parasites were isolated from popliteal or prescapular lymph nodes of all dogs by needle aspiration prior to initiation of treatment (NT group) or during treatment (TA and TR groups). Samples were aspirated aseptically and seeded in semi-solid medium [31] for 5–10 days at 27°C. Viable promastigotes were transferred to complete medium 199 (M-199) adjusted to pH 6.8 [32]. Promastigotes were tested for allopurinol susceptibility after 5–10 passages.
Axenic amastigotes were generated from promastigote isolates as described by Sereno and Lemesre [33], with the following modifications; M-199 was complemented as described above [32] only with 20% FCS, pH was adjusted to 5.8 with sodium phosphate buffer and 5x106 logarithmic phase promastigotes per mL from each tested isolate were used for induction of differentiation. Axenic amastigotes were kept at 37°C and passaged thrice to ensure viability before testing.
DNA from promastigote isolates was extracted (illustra, GE Healthcare, UK) and a 300 bases long leishmanial ITS1 fragment was amplified by PCR using the primers LITSR (5'-CTGGATCATTTTCCGATG-3') and L5.8S (5'-TGATACCACTTATCG-CACTT-3') as previously described [34]. Amplicons were sent for DNA sequencing and the resulting sequences compared to L. infantum sequences deposited in GenBank using the BLAST program (www.ncbi.nlm.nih.gov/BLAST).
Growth curves were made for three isolates from the NT and TR groups each, in order to determine whether exposure to allopurinol affected the isolates’ growth kinetics. Samples containing 1x106 per mL stationary phase promastigotes were transferred into complete medium. Promastigote concentrations were determined for each isolate every 24 hours for six consecutive days in triplicates using a Neubauer counting chamber. Average daily values were plotted, and the experiment was repeated twice for each isolate.
Half-maximal inhibitory concentration (IC50) values were determined by a viability-based assay as described [32] with modifications. For each isolate, 1x106 log phase promastigotes or axenic amastigotes were aliquoted in triplicates (100μL/well) into 96-well flat-bottom plates (Nunc, Denmark). Allopurinol (Sigma-Aldrich, St. Louis, MO) from a stock solution of 50mg/mL in 1N NaOH was 1:2 serially diluted in complete medium and added (100μL/well) to final concentrations of 0–1600 μg/mL. Three wells per plate containing complete medium were used as blanks. Plates were incubated for 66 hours, followed by addition of 20μL resazurin (alamarBlue, AbD Serotec, UK) per well. Plates were incubated for a further 6 hours, and fluorescence (λex = 540nm, λem = 590nm) was read using a fluorescent microplate reader (Synergy2, BioTek Instruments, Winsooki, VT). Data was analyzed using the Prism 5 software (GraphPad Software, San Diego, CA). Experiments were repeated twice for each isolate tested, 7–14 days apart, and the average IC50 value for each isolate and stage was used in further analysis.
The cytotoxicity of allopurinol for five different macrophage cell lines was first determined in order to identify the best host cell line for testing drug activity on intracellular amastigotes. Cytotoxicity IC50 was measured, essentially as described above for the parasites, with some modifications. Human THP-1 and U-937 (provided by Prof. Charles Jaffe’s laboratory, Hebrew University), murine RAW 264.7 (kindly provided by Professor Nachum Shpigel’s laboratory, Hebrew University) and J774.1, and canine DH-82 cell lines (the latter 2 cell lines kindly provided by Prof. Shimon Harrus’s laboratory, Hebrew University) were cultured in the following media: RPMI for THP-1 and U-937, DMEM for RAW 264.7 and J774.1, and MEM for DH-82; all supplemented with 10% FCS, 2mM L-glutamine and antibiotics (100IU penicillin G and 100μg/mL streptomycin). For each cell line, 1x104 cells were aliquoted in triplicates into 96-well plates. Allopurinol was added to final concentrations of 0, 12.5, 25, 50, 100, 200, and 400μg/mL. Cells were incubated for 66 hours at 37°C, 20μL resazurin were added and cultures were further incubated for an additional 4 hours. Two separate experiments were performed for each cell line and the average IC50 value was used for analyses.
Allopurinol susceptibility of intracellular amastigotes was evaluated using isolates from the NT and TR groups. DH-82 cells in complete MEM medium, chosen based on the cytotoxicity study above, were plated (1x104 cells/0.5mL/well in duplicates) on round glass slides in 24 well plates (Nunc, Denmark) and left for 6 hours to adhere. Promastigotes from a stationary culture of the test isolate were then added for 16 hours at 37°C using an infection ratio of 10:1 parasites:macrophage. Cells were washed twice with MEM without FCS to remove extracellular parasites, and 1mL complete MEM containing allopurinol added (final concentration being either 0 or 300μg/mL). Cells were incubated for 72 hours, washed twice with PBS, followed by fixation in methanol and Giemsa staining of the glass slides. Intracellular parasites for each isolate were counted in 100 host cells, and percent amastigote growth inhibition was calculated compared to the non-treated controls. Each experiment was repeated twice and the average percent growth inhibition was calculated.
Significance of differences in IC50’s, as well as other quantitative variables such as clinical score, serology OD or blood and biochemistry indices among the three study groups was determined using the non-parametric Kruskal-Wallis test. The Mann-Whitney non-parametric test was used to compare IC50’s between two independent groups and for multiple pairwise comparisons (as a Post Hoc test, with the Bonferroni correction of the significance level). The Fisher's exact test was applied to examine the association between two categorical variables. Association between IC50 and the serology OD or treatment duration was estimated by calculation of the Pearson and the non-parametric Spearman rank correlation coefficients. Assessing the difference in growth rates between promastigotes from two study groups was done by applying the repeated measures ANOVA model, with the Greenhouse-Geiser test for effects within each group. Comparison of clinical scores at two time point within a group was done for the TR and TA groups using the Wilcoxon Signed Ranks Test. All statistical tests applied were two-tailed, and a P-value of 5 percent or less was considered statistically significant. Average values are given with the respective standard deviation. Analysis was made using the PASW 18 software (SPSS Inc., Chicago, IL).
Procedures to identify Leishmania infection in the dogs and medical treatment were carried out as a part of their routine management with informed consent of the dog’s owners, and in compliance with the Hebrew University's animal use ethical guidelines. The animal care protocol was approved by the Hebrew University’s Institutional Animal Care and Use Committee (IACUC) which adheres to the NIH guidelines; approval no. MD-08-11476-2.
The 19 parasite strains included in this study were derived from 10 non-treated dogs (NT), and from 9 dogs undergoing allopurinol treatment including five dogs that have gone into remission following treatment (TA) and 4 relapsed (TR). Complete information of each dog and respective parasite strain can be found in Tables 1 and S2.
The ages of infected dogs varied between 1.5 and 11 years, with mean of 6.4±3.7, 6.2±2.9 and 6.8±3.0 years for the NT, TA and TR groups, respectively. No significant difference was found between the average ages of the three groups (Kruskal-Wallis test, P = 0.92). Eleven females and eight males were included in the study groups. No significant difference in gender distribution was found between the study groups (Fisher's Exact test, P = 0.154), even though females were over-represented in the TR group.
All dogs presented clinical signs compatible with CanL at diagnosis, and in group TR also at enrollment (for group NT diagnosis and enrollment were at the same time point) [8]. The average clinical score at disease diagnosis was 4.5±1.0, 4.8±1.1 and 5±1.5 for the TR, TA and NT groups, respectively, with no significant difference between the groups (Kruskal-Wallis Test, P = 0.79). The clinical score for the TA group was significantly lower on enrollment, both compared to group’s average score at diagnosis (1±0 versus 4.8±1.1, Wilcoxon Signed Ranks Test, P = 0.042) and the score at enrollment for the NT and TR groups (Mann-Whitney Test, P = 0.001 and 0.016, respectively). No significant differences were found in blood and biochemistry indices between groups, except for a higher total protein for the NT group compared to the TA group (Mann-Whitney Test, P = 0.028).
The average duration of treatment was 10.3±5.6 and 25±15.1 months for the TA and TR groups, respectively, with no significant difference between groups (Mann-Whitney test, P = 0.19). Nine out of ten dogs in the NT group were positive for L. infantum by blood PCR, compared to one out of eight of tested treated dogs (groups TA and TR). The average ELISA serology OD was 1.42±0.3, 1.11±0.2 and 1.28±0.6 for the NT, TA and TR groups, respectively, with no significant difference between groups (Kruskal-Wallis test, P = 0.256).
All of the isolates obtained from dogs were identified as L. infantum with 99–100% similarity by BLAST to L. infantum ITS1 sequences GI339730635, GI332330750, GI306422161 and deposited in GenBank (S1 Table; GenBank accessions KM677128- KM677146). Comparison of promastigote growth curves of 3 isolates from the NT and TR groups each demonstrated that all tested strains entered the logarithmic growth phase after 2 days in culture (Fig 1). Stationary phase was achieved after an average of 4.67±0.6 or 4.33±0.6 days for the NT and TR strains, respectively. No significant difference was found between promastigote counts for strains from the two groups on any of the days (Repeated measures ANOVA with Greenhouse-Geisser tests, P = 0.784). Based on this data we could conclude that there was no inherent difference in growth kinetics between naïve isolates and those previously exposed to allopurinol, thus we proceeded to test the IC50 using similar incubation times for all isolates.
Allopurinol toxicity for 5 different cell lines, expressed as allopurinol IC50 value for each cell line, was determined (Table 2). The canine macrophage cell line, DH-82, was found to have the highest IC50, 377±3.25μg/mL. Because the other cell lines showed IC50’s 1.6–5.3 folds lower, the DH-82 cell line was chosen for evaluating the susceptibility of intracellular amastigotes to allopurinol. The allopurinol concentration chosen for this assay was 300μg/mL, which caused approximately 20% growth inhibition of the DH-82 cells in the absence of parasites.
All isolates (n = 19) were tested as promastigotes. The average IC50 found for promastigotes from the TR group was approximately fourfold higher (996±372 μg/mL), and significantly different from that of the two other groups (Kruskal-Wallis test, P = 0.01) (Fig 2 and S2 Table). No significant difference was noted between the average allopurinol IC50 for promastigotes in the NT group (200±145μg/mL) compared to the TA group (268±172μg/mL) (Mann-Whitney test, P = 0.594). Sensitivity of axenic and intracellular amastigotes to allopurinol was determined for parasite strains of the NT and TR groups (n = 10 and 4, respectively). The IC50 observed for the axenic amastigote stage (Fig 2 and S2 Table) was consistently higher, by 1.5 to 10.2 fold compared to the corresponding promastigote stage of each strain, regardless of the treatment history of the dog. As seen for promastigotes, average allopurinol IC50 value for axenic amastigotes from group TR (1678±396 μg/mL) was also more than two folds higher than this found for isolates from group NT (671±129μg/mL, Mann-Whitney test, P = 0.002). Finally, allopurinol at 300μg/mL resulted in an average 20±4.9 percent growth inhibition for NT group intracellular amastigotes in DH-82 infected macrophages, which was significantly higher compared to 6.3±4.8 percent growth inhibition for the TR group isolates (Mann-Whitney test, P = 0.002) (Tables 3 and S2). We were unable to use higher allopurinol concentrations in order to determine the exact IC50 value for intracellular amastigotes, since concentrations over 300μg/mL resulted significant cytotoxicity and growth inhibition of the host DH-82 cells (>20%).
No significant correlation was found between the promastigote IC50 and serology results (Spearman’s rho correlation, r < 0.17), or treatment length (r < 0.58) of dogs from which the respective parasites were isolated.
This is the first detailed report of resistance to allopurinol in L. infantum parasites isolated from dogs and associated with clinical relapse. All dogs included in this study were seropositive at diagnosis, had typical clinical manifestations of CanL at the time of their initial diagnosis or when they relapsed, and were positive by lymph node culture. The clinical score was able to reliably reflect the clinical status of each group, while serology, complete blood count results, as well as the biochemical indices tested, did not show significant differences between groups that could be beneficial for their clinical characterization. A previous study has demonstrated that an increase in parasite DNA loads in lymph nodes, but not blood, was correlated with the reappearance of clinical signs in relapsing dogs [17]. Our results shows that parasitemia seems to be greatly reduced or absent following treatment, which may be attributed to the effect of allopurinol, causing a total reduction in parasite load, as reported previously [35]. Thus, these results do not indicate that any ancillary test employed in the study is superior to the clinical score in the diagnosis of relapse in leishmaniotic dogs. Although it is widely accepted that dogs treated for visceral leishmaniasis using the currently available drugs frequently remain infected and often relapse [8,13], only a small number of relapse cases have been described so far [14,15,16,17,18], and no characterization of parasite strains from relapsed dogs was made. As opposed to the large volume of evidence on drug resistance in human leishmaniasis, little is known regarding the development of drug resistance in CanL. One study demonstrated that L. infantum isolates obtained from naturally infected dogs after a therapeutic course of meglumine antimoniate were less susceptible to the drug compared to isolates from the same dogs before treatment [25]. Similar results were also demonstrated in a canine experimental L. infantum infection model followed by meglumine antimoniate treatment [26]. Differences in susceptibility to meglumine antimoniate among L. infantum isolates belonging to two different zymodemes from an area lacking drug pressure have also been shown and were hypothesized to represent an inherent resistance to this drug [27]. As part of a larger study on drug susceptibility in humans and dogs, allopurinol susceptibility was evaluated in two L. infantum isolates from dogs [36]. The IC50 value of an isolate from a non-treated dog was found to be lower compared to an isolate from a dog undergoing treatment. However, the clinical status of the latter dog was not discussed, and no association was made between clinical relapse and decreased drug susceptibility of parasites. We have found a significant 3 to 4 fold increase in allopurinol IC50 for promastigote strains isolated from relapsed dogs (TR) undergoing treatment with this drug compared to non-treated dogs. This increase was evident when testing axenic amastigotes and intracellular amastigotes. Although we did not find a significant correlation between treatment duration and drug IC50, the fact that increase in drug resistance was found in parasites isolated from treated dogs suggests that resistance may at least in some cases be caused by selection under drug pressure. Resistance development over time under the pressure of drug treatment has been described for L. donovani and antimonials [21]. In addition, inherent increased resistance to drugs must also be considered, in view of the considerable variation in IC50 and the % inhibition values demonstrated within each group in all parasite life stages tested. The dissemination of drug resistant parasites may provide another possible explanation for this variation. Cases such as the TR4 isolate, which presented increased drug resistance while being under drug pressure for only short time, may be attributed to either of the above options. Although relapse was associated in this study with parasite drug resistance, a future study with a larger number of relapsed dogs may identify animals with lower IC50’s that relapsed due to other potential causes such as concurrent immune-suppressing conditions and neoplasia. By widening the database of dogs and parasites we may better understand the origin of these resistant parasites and take measures to minimize their prevalence. The association found in this study between clinical relapse and resistant parasites may affect current drug selection and treatment regimens. Since no significant difference was found in serology values between the groups and no parasite DNA was detected in the blood of most relapsed dogs, these routine tests are ineffective in discriminating between infection with sensitive and resistant parasites. Relapse in leishmaniotic dogs presents a clinical challenge, since possible effects of deteriorating leishmanial infection such as kidney disease, as well as nonrelated conditions such as other diseases, should be assessed before parasite drug resistance is defined as the cause of relapse. Despite that, our results suggest that allopurinol resistance should be considered, and that an affordable methodology to test the presence of allopurinol resistant L. infantum in dogs should be available, once clinical relapse is suspected.
In conclusion, this study is the first report of resistance to allopurinol in L. infantum isolates from dogs with clinical disease relapse. Resistance was verified in three forms of the parasite. Clinical relapse associated with naturally developing resistance to allopurinol is a serious cause for concern considering the zoonotic nature of the disease, and allopurinol's major role in long term treatment of CanL. As canines with uncontrollable L. infantum infection are highly infectious to sand flies [37], dogs infected with allopurinol resistant strains may pose great risk for infection of naive dogs as well as humans.
|
10.1371/journal.pcbi.1006846 | Identification of avian flapping motion from non-volant winged dinosaurs based on modal effective mass analysis | The origin of avian flight is one of the most controversial debates in Paleontology. This paper investigates the wing performance of Caudipteryx, the most basal non-volant dinosaur with pennaceous feathered forelimbs by using modal effective mass theory. From a mechanical standpoint, the forced vibrations excited by hindlimb locomotion stimulate the movement of wings, creating a flapping-like motion in response. This shows that the origin of the avian flight stroke should lie in a completely natural process of active locomotion on the ground. In this regard, flapping in the history of evolution of avian flight should have already occurred when the dinosaurs were equipped with pennaceous remiges and rectrices. The forced vibrations provided the initial training for flapping the feathered wings of theropods similar to Caudipteryx.
| The origin of avian flight in the perspective of mechanics has been investigated for the first time. We reported the first evidence for flapping hypothesis based on principle of physical modeling. This is significant because using modal effective mass method and reconstructed Caudipteryx, the most basal non-volant winged dinosaur, we captured significant and negligible modes and realized that resonance oscillation of Caudipteryx wings could occur as the running speed approached to the primary frequencies. Such forced vibrations induced by legs' motions during running trained the Caudipteryx and the other feathered dinosaurs to flap their wings.
| The origin of avian flight has been debated for over 150 years, ever since the discovery of the first fossil of Archaeopteryx in 1861 [1–35]. Being widely considered as the oldest and most basal-known avian taxon, Archaeopteryx is characterized by a long boney tail, three clawed digits forming the manus, teeth throughout the upper and lower jaws, a furcula, a non-ossified sternum, and perhaps most importantly, forelimbs with elongate asymmetrical feathers forming large wings. It is widely accepted that birds are nestled within the derived lineage of theropod dinosaurs, the Maniraptora. However, it is still subject to heavy debate how flight evolved within the Dinosauria, and multiple origins of flight appear increasingly probable [1–12, 24, 31]. Many researchers consider that avian flight evolved through a number of stages from a ground-dwelling quadrupedal reptile [14–18], cursorial bipedal ground-dweller [13–17, 19], and arboreal life [14, 20] including parachuting [14, 21, 22], gliding [14, 16, 20, 22, 23], and eventually achieving active powered flapping flight [14, 16, 20–23]. However, there is increasing support from studies of juvenile birds for a ground up hypothesis in which flight evolved in a terrestrial animal and the flight stroke evolved directly without an intervening gliding phase [36–43]. Among non-avian dinosaurs [20], Caudipteryx represents the most basal taxon with almost completely preserved feathered forelimbs that could be considered ‘proto-wings’ making this taxon important to the study [21] of flight origins [15, 22, 26]. Some other non-volant theropods from the Cretaceous period have been reported with feathered forelimbs [27, 28]. Caudipteryx is a basal member of the Pennaraptora [1], a derived group of maniraptoran dinosaurs, sometimes closely allied with birds and the most primitive group with pennaceous feathers. Caudipteryx has short forelimbs with distally located symmetrical pennaceous feathers and long hindlimbs. The feathering of both fore-and hindlimbs indicates that Caudipteryx was not a volant theropod [26]. Caudipteryx further differs from modern birds which have abbreviated tails and forward centered mass locating near the wings [29]. However, the most primitive winged dinosaur, Caudipteryx, is clearly terrestrial, investigating the aerodynamic properties of the proto-wings of Caudipteryx has the potential to shed light on the origin of avian flight [30].
We estimated the maximum running speed of Caudipteryx to be about 8 m/s. This value was based on the skeletal hindlimb proportions of BPM 00001 and on adopting the assumptions [44, 45] with respect to the limb posture of small theropods and the range of Froude numbers (up to 17) they might have utilized in running (see S1 Speed for detailed calculation about speed) [44, 45]. We also focused our analysis on a literally generic Caudipteryx with a body mass of 5 kg, a realistic value given that an empirical equation for estimating theropod body masses on the basis of femoral length [46] produces results ranging from 4.74 kg to 5.18 kg (mean value = 4.96 kg) for a total of five described specimens (see S1 Mass for detailed calculation about mass) [24, 47, 48].
Any part, mechanism or system has its particular natural frequencies and corresponding mode shapes [46–51]. Mathematically we can compute which natural frequency and related mode shape is significant and effective to take them into account [49, 52]. The theory of modal effective mass is based on natural frequency, modal analysis and effective masses associated with different directions [49]. The modal effective mass is a measure to classify the importance of a mode shape when a structure is excited by the enforced acceleration from base. A high effective mass in a certain direction will lead to a high reaction force at the base and will be easily excited. Resonance phenomenon occurs on the Caudipteryx when the frequency of the forced vibrations excited by running legs is matched with any natural frequency of Caudipteryx. Hence, by detecting effective natural frequencies of the whole body and analysis of corresponding mode shapes, the velocities of the Caudipteryx that stimulate the wings to flap can be obtained (50 cm is measured for the step length of Caudipteryx).
To this end, a simplified mathematical model, a Finite Element Model, a reconstructed physical model of Caudipteryx, and experiment on young ostrich have been utilized. The simplified mathematics model helps us to understand how to face with the kinematics of Caudipteryx. Finite Element (FE) model gives a precise and acceptable result to compare with the reconstructed model on the test rig and experiment on running juvenile ostrich proves the mathematical analyses and simulations.
All experiments using juvenile ostriches, data collection and data analysis procedures in this research were carried out in full accordance with ethical rules for animal welfare and according to the requirements of the Ethics Committee of Tsinghua University.
Effective mass categorizes the significance of a mode shape while a structure is excited by forced vibrations from the base. A higher effective mass will certainly lead to a higher reaction force from the basis, while mode shapes with lower related modal effective masses are hardly excited by base vibration and will provide lower reaction forces at the basis [49–55] (see S1 Text for detailed explanations about modal effective mass method). The analyses using the theory of modal effective masses represent that at which velocities, Caudipteryx could most obviously sense flapping on its wings and shoulder joints. This phenomenon is purely governed by the natural biophysics.
Finite Element Model of Caudipteryx provided a precise analysis as the number of elements were sufficient enough and non-structural masses to cover the whole body mass to reach to 5 Kg were also taken into account. Also, except those elements which have boundary conditions, all elements have full DOF in any direction (Fig 2 and S1 Fig).
We reconstructed the real-sized robot of Caudipteryx on the test rig in accordance with the existing fossils (BPM0001) (S2 Fig). The robot is composed of body, tail, neck, wings and legs, and the skeleton is fabricated from ABS plastic. There is no definite evidence for tertiary feathers of Caudipteryx. We therefore only reconstructed the primary and secondary remiges from the feathers of extant birds. Using metal pins, we attached them to the antebrachial to generate artificial articulated wings.
In the reconstructed wings, we imbedded force sensors (S3A Fig) to collect the data of lift and thrust/drag (S3C, S3D, S3E and S3F Fig).
In order to verify experiment by Caudipteryx robot and the forced vibrations phenomenon induced by legs, we also implemented the experiment on a half-adulted ostrich whose mass is 6.7 kilograms (Fig 3A) as a similar living bird to Caudipteryx. This process was performed through observations on a running juvenile ostrich (S2 Video) and experiments on running ostriches (S3 Video). A device was fixed on the ostrich’s back (Fig 3B) to measure the velocity, acceleration (S3 Fig), rolling angles of body, and wings (S4 Fig). To investigate the responses of the body and the wings in running and the advantage of aerodynamic effects of flapping wings of feathered dinosaurs, we fabricated four different sizes of feathered forearms with the simplest plate wings (Fig 3C) and executed experiments on the ostrich. Therefore, lift and thrust/drag forces produced by artificial wings during running were also measured by the force sensors (S3A and S3B Fig). The connections of the shoulder joints were particularly designed in order to avoid the effect of frictions and inertial forces during locomotion.
Mathematical model shows the first mode of the forced vibrations to flap the wings when Caudipteryx ran on the ground at the speed of about 2 m/s, the mode shape of which is expressed with a vector of n1 = (0.393 0.62 0.62 0.158 0.158 0.432 0.45)T. The FE model analysis results of the modal effective mass of the Caudipteryx (Table E in S1 Table) indicate that the effective natural modes occur only in vertical direction (Y-axis) and they are almost zero in lateral motions (X and Z axes). It expresses that the first natural frequency of about 1.99 Hz is not effective, but the second one of about 2.58 Hz and the third mode of about 5.79 Hz considering the maximum speed of Caudipteryx (the forecasted velocity is about 8 m/s for Caudipteryx) are effective and important. In other words, the oscillation about the torso axis is the first mode (S1 Fig). Therefore, the Caudipteryx should roll its whole body about the torso direction when they ran at a low speed (around 2 m/s) near the first primary frequency. The second primary mode (the most effective mode) occurred as the running speed approached to 2.5 m/s. It means flapping modes were easily excited at low frequency while Caudipteryx ran on the ground at the velocity from around 2.5 m/s to a little faster than 5.8 m/s (S1 Fig).
We fabricated four simplest plate wings with different sizes and did experiments on the ostrich to compare the lift forces obtained from the flapping wings passively applied by forced vibrations during running. At the same running speed, the wings with filament feathers (1st wing) provided the smallest lift, the largest value of which is less than 0.13 N, while the ones with longer feathers could provide larger lift (2nd and 3rd wings), and the longest feather (4th wing) could provide the largest lift which exceeds 0.42 N (S3 Fig).
In the simplified rigid body system of seven degrees of freedom of Caudipteryx, the whole system can be excited by the displacements of feet, x4 and x5 during running. After this excitation, the whole body masses move along their individual vertical directions in this model (Fig 1). It illustrates the kinematics of Caudipteryx mathematically. In order to obtain the precise results using computer simulation, Finite Element Method reveals the phenomenon that the maximum effective mass occurs in the second mode which is a flapping mode. Only in the most effective mode, could the wings of Caudipteryx be excited to flap evidently and then sense lift. Therefore, the results of the FEM model (second model) through Finite Element Method have been considered because of having the highest accuracy. On the other hand, in the FE model simulated by FEM, computer calculations represent that the first natural frequency which had been roughly calculated in the first mathematical model (first model) is almost equal to that of the FEM model; and the other natural frequencies (from the second to the seventh) in comparison with the FEM model (second model) have some deviations but still acceptable. Also in the first model the modal effective masses of each natural mode might not be equal to the accurate FEM model, but the summation of which in simplified seven-degree-of-freedom model must be 5 kilograms. The reason is the limitation on the number of elements/masses (solely seven masses) and having only one DOF in the vertical direction. The effective mass analysis discovers that the first mode has never been effective (Table E in S1 Table). As the speed approached to the second primary frequency, the Caudipteryx output the second oscillation mode. It is the flapping of the wings up and down with the same amplitudes and same directions. The simulation has been extended by either increasing or decreasing the mass of each part of the Caudipteryx (Table D in S1 Table) and assumed eight excessive masses except the actual one (S5 Fig) (by measuring) from 2 kg to 10 kg in a similar geometrical model. Hence, the frequencies and corresponding effective masses in Y-axis have been studied (Table F in S1 Table). The analyses reveal that the performance of effective modes of any model (models A, B, …, I) are identical but at different frequencies. It means that in all mass distribution models, effective mode mainly depends on the creature’s velocity. When the forced vibration frequency is near the second natural frequency, the flapping mode will be occurred. The natural frequency decreases from 4.0 Hz in mass model A to 1.8 Hz in mass model I (S5 Fig) in the second mode. Therefore, as the weight of the creature increases, the velocity necessary to reach flapping mode might be decreased.
With the observation of the experiments, we realized that when the speed of the reconstructed Caudipteryx robot on the test rig (S2 Fig) reached 2.31 m/s (near the value of what has been simulated by FEM model), the robot’s wings started to output most obvious flapping motions which is the resonance of forced vibrations in physics (Fig 4). Using theory of modal effective mass and reconstruction of Caudipteryx zoui (BPM0001) (S6 Fig and Table A in S1 Table), we infer that flapping flight could be developed earlier than gliding in the evolution of avian flight. When the running speed was near the second primary speed of about 2.5 m/s, both wings of the Caudipteryx generated oscillations similar to flapping wings.
Step length in running animals varies with speed and gait and animals do not just have one step length. Any given velocity in this research such as 2, 2.5 and 5.79 m/s dedicated to first, second and third modes was obtained by measuring and assuming some parameters from the fossil such as step length, stiffness and mass (see S1 Text for detailed explanations about Caudipteryx velocity and step length). To eliminate these uncertain values, we used interval analysis which is a powerful mathematical tool in engineering (see S1 Text for detailed explanations about Interval Analysis method). Modal effective mass and Interval Analysis represent that flapping motion occurred at lower velocity. It means, if step length was between 30 cm to 70 cm and if mass was between 3 to 7 kg, Caudipteryx had flapping motion and it occurred at lower velocities (there must be a value that will render the second mode although we do not know the exact number which is in a certain Interval). Hence, the velocities of 2.5 m/s and 5.86 m/s are only two cases among all possibilities. Therefore, the conclusion that the second and the third modes must occur at a certain value is an objective conclusion. Further, the physical phenomenon of flapping motion (induced by forced harmonious vibrations) always be generated in running, but we cannot obtain the precise value of running speed since it might be expressed with an interval of velocity. Hence, the role of body oscillation during a run should be taken into account in order to understand the origin and evolution of avian flapping wings.
Experiment results on ostrich indicated that the vibrations of the feathered wings were easily induced when ostrich ran on the ground. Under the assumption of the same length of forearms for the feathered dinosaurs, the wing with the shortest feathers generated the flapping motions with the largest amplitude while the ones with longer feathers produced the flapping motions with smaller amplitudes (S4 Fig). This is interpreted by the air resistance. The larger the wing area, the larger the resistance, and the smaller the amplitude for the passive vibrations. This experiment suggests that the flapping motion might be developed by the forced vibrations during terrestrial locomotion when the winged dinosaur appeared on the earth. However, the lift obtained from the running-foot forced vibrations shows that the longer and larger the wing was, the larger the lift would be (S3 Fig). Therefore, forced vibrations may represent the earliest stages in the evolution of forelimb flapping in winged theropods. This suggests that flapping behavior evolved in non-volant theropods long time ago before they could actively fly.
Experiments on the Caudipteryx robot based on the fossil (Caudipteryx sp. IVPP V12430) and the experiments on artificial wings placed on the back of a juvenile ostrich indicated that the forced vibrations of plumage forearms during walking and running taught the winged theropods to flap their wings. These analyses suggest that the impetus of the evolution of powered flight in the theropod lineage that lead to Aves may have been an entirely natural phenomenon produced by bipedal motion in the presence of feathered forelimbs.
|
10.1371/journal.pntd.0003129 | Fleas of Small Mammals on Reunion Island: Diversity, Distribution and Epidemiological Consequences | The diversity and geographical distribution of fleas parasitizing small mammals have been poorly investigated on Indian Ocean islands with the exception of Madagascar where endemic plague has stimulated extensive research on these arthropod vectors. In the context of an emerging flea-borne murine typhus outbreak that occurred recently in Reunion Island, we explored fleas' diversity, distribution and host specificity on Reunion Island. Small mammal hosts belonging to five introduced species were trapped from November 2012 to November 2013 along two altitudinal transects, one on the windward eastern and one on the leeward western sides of the island. A total of 960 animals were trapped, and 286 fleas were morphologically and molecularly identified. Four species were reported: (i) two cosmopolitan Xenopsylla species which appeared by far as the prominent species, X. cheopis and X. brasiliensis; (ii) fewer fleas belonging to Echidnophaga gallinacea and Leptopsylla segnis. Rattus rattus was found to be the most abundant host species in our sample, and also the most parasitized host, predominantly by X. cheopis. A marked decrease in flea abundance was observed during the cool-dry season, which indicates seasonal fluctuation in infestation. Importantly, our data reveal that flea abundance was strongly biased on the island, with 81% of all collected fleas coming from the western dry side and no Xenopsylla flea collected on almost four hundred rodents trapped along the windward humid eastern side. The possible consequences of this sharp spatio-temporal pattern are discussed in terms of flea-borne disease risks in Reunion Island, particularly with regard to plague and the currently emerging murine typhus outbreak.
| Fleas are blood-feeding parasites involved in the transmission of several arthropod borne pathogens. Rat-fleas (Xenopsylla spp.) are known vectors of bubonic plague together with other human diseases receiving less attention such as murine typhus. This latter disease was recorded for the first time in 2011 on Reunion Island where seven human cases were further confirmed within the following year. The outbreak motivated a large survey of fleas, as these insects of major veterinary and medical importance have never been investigated on this oceanic island. We collected fleas on almost 1000 small wild mammals trapped on two altitudinal transects along the humid eastern and dry western sides of the island. Our data reveal the presence of four cosmopolitan flea species and shows an astonishing distribution pattern: 81% of all collected fleas were sampled on the western transect while not a single rat-flea was sampled on the eastern humid side of the island. Interestingly, this distribution did at least in part overlay the map of murine typhus human cases. These data stimulate the need for a diagnosis of pathogens in natural flea populations together with a comprehensive distribution map of fleas, allowing a risk assessment of flea-borne diseases in humans.
| Reunion is a small oceanic island of volcanic origin located in the Indian Ocean, Southern Hemisphere (21°6′S and 55°36′E) that forms, together with Mauritius and Rodrigues Islands, the Mascarene archipelago. This oceanic island is geographically isolated from continental landmasses and located within one of the 34 recognized world biodiversity hotspots [1]. The island lies therefore in a biogeographic context favourable to species radiation and potentially high endemism. Its dramatic relief has shaped a highly contrasted climate: the mountainous centre (>3,000 meters) separates a humid windward coast (scoring some rain world records) from a dry leeward coast, which lower part consists mainly in savannah. This peculiar situation has led to the evolution of a strong vegetal endemism with a well-described altitudinal succession of vegetal species observed on both windward and leeward coasts [2]. The diversity of terrestrial animals, specifically mammals, is clearly much less prominent: the only endemic mammal species is the insectivorous bat Mormopterus francoimoutoui [3]. Following human colonization which started in the XVIIth century, five small mammal species have been introduced, namely the insectivores Suncus murinus Linnaeus 1766 (Asiatic house shrew) and Tenrec eucaudatus Schreber 1778 (tailless tenrec) from Madagascar, and the three cosmopolitan murid rodents Rattus rattus Linnaeus 1758 (black rat), Rattus norvegicus Linnaeus 1769 (brown rat) and Mus musculus Linnaeus 1758 (house mouse). Tropical countries and especially tropical islands are known at higher risk for the emergence or re-emergence of infectious diseases [4]. Therefore, updated information on zoonotic pathogens and on the diversity and distribution of their arthropod vectors is warranted for a quicker response to outbreaks threats.
Fleas (Order Siphonaptera) form a unique group of insects comprising 15 families with a total of about 220 genera and some 2,500 described species [5]. Five families including 25 genera are ectoparasites of birds, while all other flea species specifically feed on mammals. In Madagascar, located about 800 km west of Reunion Island, flea diversity has been extensively studied, mainly because of their role as vectors of Yersinia pestis, the plague agent, especially in this country that reports most human plague cases worldwide [6], [7]. Flea diversity is high in Madagascar, with several endemic species together with a few cosmopolitan ones, which host specificity and distribution have been partly described [6], [7]. Surprisingly, Xenopsylla brasiliensis (Baker, 1904) has never been collected in Madagascar, even though this species is recognized as a main plague vector in Eastern and Southern Africa [8], [9] and has been collected in Moroni (Grande Comore) and notified on two other islands of the Southwestern Indian ocean: Mayotte (Comoros archipelago) and Mauritius [10]. By contrast, almost no data are currently available on flea diversity on the other islands of the Southwestern Indian Ocean, including Reunion. The cosmopolitan and/or tropical species possibly present in the region are Pulex irritans Linné, 1758, Echidnophaga gallinacea Westwood, 1875, Leptopsylla segnis Schönherr, 1811, Xenopsylla cheopis Rothschild, 1903 as well as Ctenocephalides spp. Hence, the recent emergence of murine typhus in Reunion Island, where ten autochthonous human confirmed cases were reported between 2011 and 2013 (Balleydier E. 2014 pers. comm.) has stimulated the investigation of fleas for vector-assessment of indigenous species for the agent Rickettsia typhi. The objective of our study was therefore to report the diversity and distribution of fleas in Reunion Island with the aim of highlighting patterns of possible epidemiological importance.
Trapping was conducted throughout a one-year period survey (November 14th 2012–November 16th 2013) in different biotopes along two altitudinal transects lying on each side of the island: the eastern transect comprised eight sampling sites and the western transect, seven. In addition, two sites located in the western coast and a few sites in the urban northern part of the island were included in the present survey (Figure 1). The sampling encompassed the two local seasons, i.e. hot-wet summer from November to June, and cool-dry season from July to October, with twelve out of the twenty sampling sites being sampled twice, i.e. during the two seasons. Trapping was conducted following a standardized protocol: wire cage live traps (29 by 18 by 12 cm) were used for rats trapping (and accidentally tenrecs), and Sherman live traps for mice and shrews. On each sampling site, forty to eighty traps were placed in line approximately 15 meters apart in the afternoon; trapped animals were collected the following morning and brought back to the laboratory for processing. Traps were baited during three consecutive nights using successively within each line cheese, coconut or a mixture of peanut butter and canned sardine oil. This baiting setup (bait A, bait B, bait C, bait A, …) was implemented in order to trap most of the prevalent mammal diversity at each sampling site. Traps were left open in the same place during the day, with productive traps being immediately replaced every morning with the same bait over the 3-days trapping session.
Animals were sacrificed by cervical dislocation without anaesthesia to avoid bleeding in accordance with guidelines accepted by the scientific community for the handling of wild mammals [11] and the institutional guidelines published by the Centre National de la Recherche Scientifique (http://www.cnrs.fr/infoslabos/reglementation/euthanasie2.htm). All animal procedures carried out in this study were approved by the French Institutional Ethical Committee “Comité d'éthique du CYROI” (No. 114).
Following sacrifice, each animal was visually examined for 10 minutes and all ectoparasites, including fleas, were manually collected either with a brush soaked in ethanol when insects jumped off the host, or forceps when eye-spotted in the fur. Collected fleas were preserved in 70% ethanol for later morphological and molecular analyses. Fleas were identified at the species level using taxonomic keys provided by Lewis [12], and Hoogstraal and Traub [13]. A subsample of Xenopsylla spp. fleas were mounted permanently on slides using Euparal medium, following a procedure adapted from Brigham Young University (http://fleasoftheworld.byu.edu/Systematics/MountingTechniques.aspx). The gender, genus, and species were recorded for each flea specimen. Xenopsylla cheopis and X. brasiliensis were mainly differentiated using the occurrence of marginal cones at the basis of the antepygidial bristle in males, and shape of spermatheca on mounted females [7]. Rodents body mass, ear, and back foot lengths, together with tail and body lengths were recorded. Rattus spp. was identified using morphological criteria including the comparison of (i) the ratio of tail to body lengths, (ii) the ear length and (iii) the hind foot length [14].
The morphological diagnosis of Rattus spp. was confirmed by molecular data through sequencing of cytB locus from 15 randomly selected animals morphologically identified as R. rattus or R. norvegicus. Briefly, DNA was prepared from 20 mg of kidney tissue as previously described [15] and used as a template with L14723 and H15915 primers set, following a previously described PCR protocol [14].
For molecular diagnosis of fleas, DNA was prepared as follows: fleas were dried individually and subsequently crushed with a TissueLyser (Qiagen, Valencia, CA) using 3 mm tungsten beads and cetyl trimethyl ammonium bromide 2%; DNA was further extracted following a previously described procedure [16]. Both nuclear and mitochondrial loci were sequenced by amplifying 28S ribosomal RNA (28S rRNA gene) and cytochrome oxidase II (COII) encoding gene using 28S A/28S rD7b1 and COII F-leu/COII R-lys primer pairs, that produce 1473-bp and 770-bp PCR fragments respectively [17]. Amplicons were sequenced on both strands by Genoscreen (Lille, France) using the same PCR primers, and sequences were edited using Geneious Pro [18]. All sequences used in this study were deposited in Genbank and are accessible under accession numbers KJ638526 to KJ638590.
All sequences were automatically aligned using MUSCLE implemented in Geneious Pro version 5.3.4 [18]. Alignments were constructed separately for the nuclear (28S) and mitochondrial (COII) datasets using sequences available in GenBank to complete our dataset. Bayesian analyses were performed to infer phylogenetic relationships between flea species. First, the best-fitting model and associated parameters were selected by jModelTest [19] and phylogenies were constructed by Bayesian inference. Two sets of four MCMCMC (Metropolis Coupled Markov Chain Monte Carlo) chains incrementally-heated were run in MrBayes 3.1.2 [20] for 20,000,000 generations. Trees and associated model parameters were sampled every 300 generations. The initial 2,000 trees were discarded as a conservative “burn-in” and the harmonic mean of the likelihood was calculated by combining the two independent runs. The 50% majority-rule consensus tree was then computed from the sampled trees in the two independent runs under the best model.
The data were entered into EPIData 3.1 and analyzed with Epi info 6.04 statistical software using the chi-squared or Fisher exact tests for observed frequencies. We used a p-value threshold of 0.001. The effect of “habitat” on fleas' diversity was measured at two scales, host and sampling region, by using the flea percentage incidence index (PII: mammals parasitized by fleas of species A/mammals caught (%)), the specific flea index (SFI: number of fleas of species A collected from host species Y/mammals of species Y parasitized by fleas of species A) and the total flea index (TFI: total fleas collected/total trapped mammals, i.e. mean number of fleas per trapped mammal) [21]. The seasonality of flea diversity was tested by comparing PII on animals trapped at each site during the cool-dry versus hot-wet seasons.
A total of 960 small mammals were trapped. They belong to the five introduced small terrestrial mammal species occurring in Reunion Island: 39 mice (Mus musculus), 168 shrews (Suncus murinus) and 25 tenrecs (Tenrec eucaudatus), all other specimens being rats (Rattus rattus: N = 554; R. norvegicus: N = 174) (Table 1). Almost 10% (95) of trapped mammals were infested with fleas (Table 1) and the TFI (mean number of fleas per host) was equal to 0.3 when based on all trapped mammals, and equal to 3 when based on parasitized mammals only. Of 288 fleas collected during the survey, 286 could be identified on a morphological basis. They were distributed within three genera and four distinct species, namely Xenopsylla cheopis (N = 171), Xenopsylla brasiliensis (N = 63), Leptopsylla segnis (N = 43) and Echidnophaga gallinacea (N = 9) (Table 2).
Rattus rattus was found to be the most parasitized host, predominantly by Xenopsylla spp. (p<10−3). Only five mice, eight shrews and one tenrec were found parasitized by fleas (Xenopsylla spp. and L. segnis) (Table 2). Rattus rattus was more heavily infested in the western side of the island (p<10−7) whereas R. norvegicus was most infested in the northern part (p<10−4) and mice in the eastern part (p<10−4). No difference according to the sampling region was found in shrews or tenrecs. When considering Xenospylla spp., X. cheopis was mainly found on Rattus spp. (p<10−5) with no difference between R. rattus and R. norvegicus but X. brasiliensis was significantly more abundant on R. rattus (p<10−4) than on any other mammal species. The number of flea species per host species ranged from one to four (Table 2), but most mammals were parasitized by a single flea species although nine R. rattus were found co-infested with two distinct species as follows: X. cheopis+X. brasiliensis (N = 1), X. cheopis+E. gallinacea (N = 3), and X. cheopis+L. segnis (N = 3).
Xenopsylla spp. were by far the most common fleas (234/286 fleas) with X. cheopis and X. brasiliensis representing 59% (171/286) and 22% (63/286) of all identified fleas, respectively (Table 2). Xenopsylla cheopis was also the most geographically widespread species, as it was present in all of the fourteen flea-positive sampling sites out of the twenty prospected ones. X. brasiliensis was collected at only two sites throughout the island, both of them being located on the western transect. Noteworthy, X. brasiliensis/R. rattus SFI index was relatively high in one of those 2 sites (Sans Soucis, SFI = 2). Leptopsylla segnis was collected on mice and both rat species in four elevated sites (>1,000 meters), and E. gallinacea was only collected on R. rattus at three distinct sites along the western transect (Figure 2).
Windward and leeward transects displayed dramatically different results, in terms of abundance of fleas and species richness (Table 3). The PII was significantly lower (p<10−7) in the eastern region compared to the northern and western regions. Indeed, 201 Xenopsylla fleas were collected out of 405 mammals trapped in the western transect while this species was totally absent on the 464 rodents trapped on the eastern transect (see Tables 1, 3); the only two X. cheopis specimens collected in the eastern side were from one tenrec trapped on the top of the eastern transect located in an elevated plateau at the centre of the island (Table 3; Figure 2). All other fleas collected in the eastern transect were identified as L. segnis (21 of 24 collected fleas; Table 3).
Lower flea species richness was recorded in animals trapped along the eastern than in the western transect: fleas were absent on six of the nine eastern sampling sites,and on the remaining sites, only seven mammals were found parasitized. The specific flea indexes (SFI) were 1.47 for X. cheopis/R. norvegicus on the northern sampling sites; 0.53 for X. cheopis/R. rattus in the western sites; and 0.26 for X. brasiliensis/R. rattus in the western sites (see Tables 4 and 5).
There is no apparent seasonality of flea abundance in the eastern region, which could be explained by the absence or very low abundance of fleas, even during the peak season observed on other parts of the Island. Seasonality is observed in the west, with greater abundance observed during the hot-wet season. Over the fourteen flea-positive sampling sites, seven were sampled during the two seasons. Two sampling sites were flea-positive during both seasons, four were flea-positive only during the hot-wet season and one was found flea-negative during the hot-wet season, and flea-positive during the cool-dry season (one R. norvegicus and one S. murinus parasitized by one X. cheopis flea each), but the difference was not statistically significant (Table 6). This seasonality was significant for X. brasiliensis on sampling site « Sans soucis » (p = 0.01; RR = 2.2 [1.1–4.3]), and for X. cheopis on sampling site « Port est » (p<10−3; RR = 11.7 [1.6–86.5]).
Sixty (28S) and seventy (COII) sequences were obtained from fleas sampled in Reunion Island. As all sequences of X. cheopis and X. brasiliensis were 100% identical, only a dozen sequences representative of each of those two species were included in the analyses. Few sequences from Genbank were added, including Parapsyllus longicornis used as an extra-group. Since no 28S or COII sequences were available on databases for X. brasiliensis, we sequenced three X. brasiliensis specimens sampled in Tanzania (KJ638557-59 in COII; KJ638585, 638589-90 in 28S: collectors Laudisoit A., Makundi R., Katakweba A., S3°58′989″ E35°21′560″, 1994 m, 10/02/2009). Models selected by jModelTest were GTR+I for 28S phylogeny (AIC weight = 0.62), and GTR+G for COII phylogeny (AIC weight = 0.85). All X. cheopis (from Reunion Island and two haplotypes from Genbank, 28S sequence) branched within a single well supported clade, while X. brasiliensis haplotypes fell within two well supported clades, one containing sequences from Tanzanian fleas, the second harboring all haplotypes from Reunion Island (Figure 3). Both clades formed a well supported monophyletic X. brasiliensis clade distinct from X. cheopis and embedded within Xenopsylla group.
The present investigation provides the first information on flea diversity and distribution on the five introduced small mammal species present on Reunion Island, where no data were available thus far. We describe the presence of three genera composed of four distinct cosmopolitan species, namely X. cheopis, X. brasiliensis, L. segnis and E. gallinacea. Morphological diagnosis of X. cheopis and X. brasiliensis was further confirmed by sequencing of 28S and COII markers: for X. cheopis, fleas sampled in Reunion Island showed 99% and 100% identity with sequences accessible in Genbank (i.e. EU336145.1 and HM188404.1 for 28S and COII sequences, respectively). As no sequences were currently available for X. brasiliensis on these 2 loci, we generated sequence data using specimens previously sampled in Tanzania and morphologically identified as X. brasiliensis by A. Laudisoit and colleagues. Again, molecular data confirmed X. brasiliensis morphological diagnosis, with 28S and COII sequences obtained from fleas sampled in Reunion Island showing respectively 99% and 94% identity with sequences obtained from Tanzanian fleas.
Phylogenetic analysis carried out with both nuclear and cytoplasmic markers provided two well resolved mostly congruent trees, suggesting that no hybridization nor introgression (two molecular events known to lead to molecular misdiagnosis [22]) has occurred within our sample. However, the analyses did reveal one incongruency for L. segnis: while 28S-based analysis was coherent with classical taxonomy, COII sequences unexpectedly clusterized L. segnis within Pulicidae. Additional and more informative markers need to be investigated in order to address this incoherence together with other more basic questions such as a previously reported paraphylly of Leptopsyllidae [23]. The absence of molecular data for L. segnis together with the overall scarcity of accessible DNA sequences for other flea species (including X. brasiliensis, see above) should stimulate an increased effort towards the release of a proper barcoding tool facilitating the diagnosis of cosmopolitan species. As for X. brasiliensis, nuclear and mitochondrial sequences from Tanzanian specimens formed a cluster separated from Reunion Island sequences, which might indicate an ongoing diversification. However, a proper investigation of eastern African and Indian Ocean X. brasiliensis populations would be required to ascertain any level of genetic structuration. Altogether, our data indicate a low diversity of fleas on small mammals from Reunion Island. In addition, all flea species were cosmopolitan and likely result from the recent introduction of their vertebrate hosts on the island, or from the importation of food stocks with preimaginal stages. This feature is not unexpected considering the low specific richness in mammal hosts, which strikingly contrasts with the neighbouring island of Madagascar where species richness and endemism of both flea [7] and small mammal hosts are high [24], flea endemism likely resulting from long host-parasite co-evolutionary processes.
Host specificity differed between fleas: E. gallinacea was only collected on R. rattus which is likely a spill over host from poultry breeding sites near the concerned sampling sites, i.e. rural areas where R. norvegicus is likely to be less common. Xenopsylla brasiliensis appeared mostly associated with R. rattus (one flea found on a shrew) a situation reminiscent to that previously described in the Canary islands [25]. On the contrary there was low host specificity for X. cheopis that was found to most commonly infest Rattus spp. (92%), but was also found on shrews and tenrecs (Tables 4 and 5), which is in accordance with previous report from Madagascar [7]. The number of collected specimens from the two other species was too low to conclude about host specificity.
This is the first report of Xenopsylla brasiliensis in Reunion Island. This species is native to continental subsaharian Africa where it is the most common plague vector in some areas, often more abundant than X. cheopis [9]. This expanding species has spread to other parts of the world such as Brazil and India [26]. This known plague vector, particularly effective in rural environments, is less tolerant to high temperatures than X. cheopis but is more resistant to drier conditions [21]. These ecological traits are in agreement with X. brasiliensis distribution in Reunion Island, where the species was restricted -in our sample- to a semi-xerophil landscape partly covered with Tamarinus indica and patches of exotic Furcraea foetida and Agave americana on the western side of the island.
The heterogeneous distribution of fleas over Reunion Island, with no Xenopsylla flea collected along the windward humid eastern side, might be related to excessive rainfall in this coast. Indeed, temperature, rainfall and relative humidity have direct effects on development and survival of fleas, and a direct effect of rainfall is supposed to occur when high intensity rainfall causes flooding of rodent burrows [27]. Seasonal abundance of fleas that has been largely reported in literature is also driven by climate variables. Warm-moist weather has been described to provide higher flea indices [27]. This is in agreement with the decrease in flea abundance observed during the cool-dry season on the two sampling sites were seasonality was significant.
Fleas are of tremendous medical and economic importance as vectors of several diseases including bubonic plague, murine typhus and tularaemia [28]. The discovery of fleas as vectors of Yersina pestis, and later of Rickettsia typhi, the ethiological agent of murine typhus, stimulated flea studies in the early 20th century. Xenopsylla cheopis is now considered as the most important cosmopolitan vector of both Y. pestis and R. typhi, and an important Bartonella spp. carrier, and X. brasiliensis is an efficient plague vector, especially in rural environments. Leptopsylla segnis is a weak vector of Y. pestis according to old standards (but no recent experimental studies have been performed to establish if the early-phase transmission apply to this species) and is a dubious vector of R. typhi [29]. Hence, our study showing that Reunion Island hosts several flea species of medical importance warrants better surveillance of potentially emerging flea-borne zoonoses.
Among flea-borne diseases, the situation of plague is of major concern for the region. Plague was introduced in Madagascar from India in 1898 and has become endemic in the highlands [30]. Xenopsylla cheopis and the endemic flea Synopsyllus fonquerniei are known as the primary vectors of Y. pestis on Madagascar [31]. In Reunion Island, plague has quite a long history: the disease was likely misdiagnosed as lymphatic filariasis until 1899 when Y. pestis was isolated by André Thiroux and formally identified by Emile Roux [32]. Thus plague was described within the same year in Madagascar, Reunion and Hawaii, but it was considered as introduced in Madagascar [32] and Hawaii [33] where foci were first described in harbors, while André Thirioux described plague as endemic in Reunion [32].
Plague is not a concern anymore in Reunion Island where the last human cases were reported in 1926 [34]. Indeed, an SFI of 0.5 to 1 is considered sufficient to maintain plague in a locality and an index ≥1 is reported to represent a potentially dangerous situation with respect to the risk of plague outbreak [8]. Some indexes reported herein (Tables 4 and 5), specifically the X. cheopis/R. norvegicus SFI measured on the north of the island may be considered of concern and should be monitored systematically. This area is close to the city of Le Port, the only international harbour of Reunion Island, and the most likely entry port for parasitized rodents and/or food. Although the risk of plague introduction from Madagascar is expected to be limited with an SFI index in this area <0.5 [7], the substantial shipping trade between Reunion and Madagascar where plague has already been described in harbours [35], [36] command a cautious control in order to prevent introduction of rodents from this plague endemic country [28]–[29]. Finally, the role of domestic cats should not be overlooked since Felidea – in contrast to Canidea in general - are sensitive to the disease, can become infected by ingesting infested rodents and develop pulmonary form of the disease, with a risk of direct respiratory transmission of infectious droplets to the people caring for them [37].
Considering other flea-borne diseases, rickettsioses represent an important concern. Interestingly, a retrospective French study (2008–2010) on travellers returning from Madagascar and Reunion reported two patients who were infected with murine typhus during their trip [38]. More recently, in 2012 and 2013, several autochthonous human confirmed cases of murine typhus were reported by hospital clinicians from the western and southern parts of the island (Balleydier E., pers. comm.). The authors were wondering if the heterogeneous distribution of human cases could be related to medical surveillance bias. Although incomplete, since the southern coast of the island wasn't sampled, the distribution of fleas reported herein is at least in part overlaid with that of human cases. This may suggest that the risk of murine typhus in Reunion Island is related to fleas' geographical distribution driven by environmental determinants. The detection of R. typhi in fleas together with the presentation of a more complete Xenopsylla sp. distribution map throughout the island may provide public health agencies with a useful tool for implementing a specific surveillance system for better risk assessment of murine typhus and other emerging flea-borne zoonoses in Reunion Island.
|
10.1371/journal.pbio.2006558 | Representational interactions during audiovisual speech entrainment: Redundancy in left posterior superior temporal gyrus and synergy in left motor cortex | Integration of multimodal sensory information is fundamental to many aspects of human behavior, but the neural mechanisms underlying these processes remain mysterious. For example, during face-to-face communication, we know that the brain integrates dynamic auditory and visual inputs, but we do not yet understand where and how such integration mechanisms support speech comprehension. Here, we quantify representational interactions between dynamic audio and visual speech signals and show that different brain regions exhibit different types of representational interaction. With a novel information theoretic measure, we found that theta (3–7 Hz) oscillations in the posterior superior temporal gyrus/sulcus (pSTG/S) represent auditory and visual inputs redundantly (i.e., represent common features of the two), whereas the same oscillations in left motor and inferior temporal cortex represent the inputs synergistically (i.e., the instantaneous relationship between audio and visual inputs is also represented). Importantly, redundant coding in the left pSTG/S and synergistic coding in the left motor cortex predict behavior—i.e., speech comprehension performance. Our findings therefore demonstrate that processes classically described as integration can have different statistical properties and may reflect distinct mechanisms that occur in different brain regions to support audiovisual speech comprehension.
| Combining different sources of information is fundamental to many aspects of behavior, from our ability to pick up a ringing mobile phone to communicating with a friend in a busy environment. Here, we have studied the integration of auditory and visual speech information. Our work demonstrates that integration relies upon two different representational interactions. One system conveys redundant information by representing information that is common to both auditory and visual modalities. The other system, which is supported by a different brain area, represents synergistic information by conveying greater information than the linear summation of individual auditory and visual information. Further, we show that these mechanisms are related to behavioral performance. This novel insight opens new ways to enhance our understanding of the mechanisms underlying multi-modal information integration, a fundamental aspect of brain function. These fresh insights have been achieved by applying to brain imaging data a recently developed methodology called the partial information decomposition. This methodology also provides a novel and principled way to quantify the interactions between representations of multiple stimulus features in the brain.
| While engaged in a conversation, we effortlessly integrate auditory and visual speech information into a unified perception. Such integration of multisensory information is a key aspect of audiovisual speech processing that has been extensively studied [1–4]. Studies of multisensory integration have demonstrated that, in face-to-face conversation, especially in adverse conditions, observing lip movements of the speaker can improve speech comprehension [4–7]. In fact, some people’s ability to perform lip reading demonstrates that lip movements during speech contain considerable information to understand speech without corresponding auditory information [1, 8], even though auditory information is essential to understand speech accurately [9].
Turning to the brain, we know that specific regions are involved in audiovisual integration. Specifically, the superior temporal gyrus/sulcus (STG/S) responds to integration of auditory and visual stimuli, and its disruption leads to reduced McGurk fusion [10–14]. However, these classic studies present two shortcomings. First, their experimental designs typically contrasted two conditions: unisensory (i.e., audio or visual cues) and multisensory (congruent or incongruent audio and visual cues). However, such contrast does not dissociate effects of integration per se from those arising from differences in stimulation complexity (i.e., one or two sources) that could modulate attention, cognitive load, and even arousal. A second shortcoming is that previous studies typically investigated (changes of) regional activation and not information integration between audiovisual stimuli and brain signals. Here, we address these two shortcomings and study the specific mechanisms of audiovisual integration from brain oscillations. We used a novel methodology (speech-brain entrainment) and novel information theoretic measures (the partial information decomposition [PID] [15]) to quantify the interactions between audiovisual stimuli and dynamic brain signals.
Our methodology of speech-brain entrainment builds on recent studies suggesting that rhythmic components in brain activity that are temporally aligned to salient features in speech—most notably the syllable rate [5, 6, 16–18]—facilitate processing of both the auditory and visual speech inputs. The main advantage of speech-brain entrainment is that it replaces unspecific measures of activation with measures that directly quantify the coupling between the components of continuous speech (e.g., syllable rate) and frequency-specific brain activity, thereby tapping more directly into the brain mechanisms of speech segmentation and coding [17].
In the present study, we used a recently developed information theoretic framework called PID (see Fig 1A and Materials and methods for details) [15, 19, 20]. We consider a three-variable system with a target variable M (here magnetoencephalography [MEG]) and two predictor variables A and V (here audio and visual speech signals), with both A and V conveying information about the target M. Conceptually, the redundancy is related to whether the information conveyed by A and V is the same or different. If the variables are fully redundant, then this means either alone is enough to convey all the information about M (i.e., obtain an optimal prediction of M), and adding observation of the second modality has no benefit for predicting the MEG signal M. The concept of synergy is related to whether A and V convey more information when observed simultaneously, so the prediction of M is enhanced by simultaneous observation of the values of A and V [15]. This means M also represents the instantaneous relationship between A and V. For example, if M is given by the difference between A and V at each sample, then observing either A or V alone tells little about the value of M, but observing them together completely determines it. The PID provides a methodology to rigorously quantify both redundancy and synergy, as well as the unique information in each modality. Unique information is the prediction of the MEG that can be obtained from observing A alone but that is not redundantly available from observing V.
The PID framework therefore addresses a perennial question in multisensory processing: the extent to which each sensory modality contributes uniquely to sensory representation in the brain versus how the representation of different modalities interact (e.g., audio and visual). The PID provides a principled approach to investigate different cross-modal representational interactions (redundant and synergistic) in the human brain during naturalistic audiovisual speech processing—that is, to understand how neural representations of dynamic auditory and visual speech signals interact in the brain to form a unified perception.
Specifically, we recorded brain activity using MEG while participants attended to continuous audiovisual speech to entrain brain activity. We applied the PID to reveal where and how speech-entrained brain activity in different regions reflects different types of auditory and visual integration. In the first experimental condition, we used naturalistic audiovisual speech for which attention to visual speech was not critical (“All congruent” condition). In the second condition, we added a second interfering auditory stimulus that was incongruent to the congruent audiovisual stimuli (“AV congruent” condition), requiring attention to visual speech to suppress the competing additional incongruent auditory input. In the third condition, both auditory stimuli were not congruent to visual stimulus (“All incongruent”). This allows us to see how the congruence of audiovisual stimuli changes integration. We contrasted measures of redundant and synergistic cross-modal interactions between the conditions to reveal differential effects of attention and congruence on multisensory integration mechanisms and behavioral performance.
We first studied PID in an “All congruent” condition (diotic presentation of speech with matching video) to understand multisensory representational interactions in the brain during processing of natural audiovisual speech. We used mutual information (MI) to quantify the overall dependence between the full multisensory dynamic stimulus time course (broadband speech amplitude envelope and lip area for auditory and visual modalities, respectively) and the recorded brain activity. To determine the dominant modality in each brain area, we statistically compared the auditory unique information to visual unique information across subjects. Note that here, auditory unique information is unique in the context of our PID analysis. Specifically, it quantifies information about the MEG response, which is available only from the auditory speech envelope and not from the visual lip area. The same is true for unique visual information. The analysis revealed stronger visual entrainment in bilateral visual cortex and stronger auditory entrainment in bilateral auditory, temporal, and inferior frontal areas (paired two-sided t test, df: 43, P < 0.05, false discovery rate [FDR] corrected; Fig 1B).
To identify a frequency band at which auditory and visual speech signals show significant dependencies, we computed MI between both signals and compared it to MI between nonmatching auditory and visual speech signals for frequencies from 0 to 20 Hz. Here, we used all the talks in the present study to delineate the spectral profile of dependencies between matching or nonmatching auditory and visual speech signals. As expected, only matching audiovisual speech signals show significant MI peaking at 5 Hz (Fig 2A), consistent with previous results based on coherence measure (see Fig 2C in [5]). Based on this, we focus our further analysis on the 3–7 Hz frequency band (5 ± 2 Hz) in the following analyses (Figs 3–5). This frequency range is known to correspond to the syllable rate in continuous speech [16] and within which amplitude envelope of speech is known to reliably entrain auditory brain activity [18–23].
Next, we investigated how multimodal representational interactions are modulated by attention and congruence in continuous audiovisual speech. Here, we focus on an “AV congruent” condition in which a congruent audiovisual stimulus pair is presented monaurally together with an interfering nonmatching auditory speech stimulus to the other ear (Fig 2B). This condition is of particular interest because visual speech (lip movement) is used to disambiguate the two competing auditory speech signals. Furthermore, it is ideally suited for our analysis because we can directly contrast representational interactions quantified with the PID in matching and nonmatching audiovisual speech signals in the same data set (see Fig 2B).
Fig 3 shows corrected group statistics for the contrast of matching and nonmatching audiovisual speech in the “AV congruent” condition. Redundant information is significantly stronger in left auditory and superior and middle temporal cortices (Fig 3A; Z-difference map at P < 0.005) for matching compared to nonmatching audiovisual speech. In contrast, significantly higher synergistic information for matching compared to nonmatching audiovisual speech is found in left motor and bilateral visual areas spreading along dorsal and ventral stream regions of speech processing [24] (Fig 3B; Z-difference map at P < 0.005). Next, we tested attention and congruence effects separately because the contrast of matching versus nonmatching audiovisual speech confounds both effects.
First, the congruence effect (“AV congruent” > “All incongruent”) shows higher redundant information in left inferior frontal region (BA 44/45) and posterior superior temporal gyrus and right posterior middle temporal cortex (Fig 4A; Z-difference map at P < 0.005) and higher synergistic information in superior part of somatosensory and parietal cortices in left hemisphere (Fig 4B; Z-difference map at P < 0.005).
The attention effect (“AV congruent” > “All congruent”) shows higher redundant information in left auditory and temporal (superior, middle, and inferior temporal cortices and pSTG/S) areas and right inferior frontal and superior temporal cortex (Fig 5A; Z-difference map at P < 0.005). Higher synergistic information was localized in left motor cortex, inferior temporal cortex, and parieto-occipital areas (Fig 5B; Z-difference map at P < 0.005).
In summary, theta-band activity in left pSTG/S represents redundant information about audiovisual speech significantly more strongly in experimental conditions with higher attention and congruence. In contrast, synergistic information in the left motor cortex is more prominent in conditions requiring increased attention. Therefore, the increased relevance of visual speech in the “AV congruent” condition leads to increased redundancy in left pSTG/S and increased synergy in left motor cortex. This differential effect on representational interactions may reflect different integration mechanisms operating in the different areas. For detailed local maps of interaction between predictors (auditory and visual speech signals) and target (MEG response), see S3 Fig.
Next, we investigated if the differential pattern of redundancy and synergy is of behavioral relevance in our most important condition—"AV congruent"—in which visual speech is particularly informative. To this end, we extracted raw values of redundancy for the location showing strongest redundancy in the left pSTG/S in Fig 5A and synergy for the location showing strongest synergy in the left motor cortex in Fig 5B for “AV congruent” condition. After normalization with surrogate data (see Materials and methods section), we computed correlation with performance measures (comprehension accuracy) across participants. Both redundancy in left pSTG/S (R = 0.43, P = 0.003; Fig 5C) and synergy in left motor cortex (R = 0.34, P = 0.02; Fig 5D) are significantly correlated with comprehension accuracy. These results suggest that the redundancy in left pSTG/S and synergy in left motor cortex under challenging conditions (i.e., in the presence of distracting speech) are related to perceptual mechanisms underlying comprehension.
In this study, we investigated how multisensory audiovisual speech rhythms are represented in the brain and how they are integrated for speech comprehension. We propose to study multisensory integration using information theory for the following reasons: First, by directly quantifying dynamic encoded representation of speech stimuli, our results are more clearly relevant to information-processing mechanisms than are differences in activation between blocks of stimulus conditions. Second, cross-modal interactions can be quantified directly within a naturalistic multimodal presentation without requiring contrasts between multimodal and unimodal conditions (e.g., AV > A + V). Third, the PID provides measures of representational interactions that address questions that are not available with other statistical approaches (particularly synergy; Fig 1A) [15, 21, 22].
We found that left posterior superior temporal region represents speech information that is common to both auditory and visual modalities (redundant), while left motor cortex represents information about the instantaneous relationship between audio and visual speech (synergistic). These results are obtained from low-frequency theta rhythm (3–7 Hz) signals corresponding to syllable rate. Importantly, redundancy in pSTG/S and synergy in left motor cortex predict behavioral performance—speech comprehension accuracy—across participants.
A critical hallmark of multisensory integration in general, and audiovisual integration in particular, is the behavioral advantage conveyed by both stimulus modalities as compared to each single modality. Here, we have shown that this process may rely on at least two different mechanisms in two different brain areas, reflected in different representational interaction profiles revealed with information theoretic synergy and redundancy.
In fMRI studies, audiovisual speech integration has been studied using experimental conditions that manipulate the stimulus modalities presented (e.g., [13, 25]). Changes in blood oxygen level–dependent (BOLD) responses elicited by congruent audiovisual stimuli (AV) have been compared to auditory-only (AO), visual-only (VO), their sum (AO + VO), or their conjunction (AO ∩ VO). Greater activation for the congruent audiovisual condition (AV) compared to others has been interpreted as a signature of audiovisual speech integration. Comparison to auditory-only (AO) activation or visual-only (VO) activation has been regarded as a less conservative criterion for integration, since even if auditory and visual stimuli caused independent BOLD activity that combined linearly, this contrast would reveal an effect. To address this, comparison to the summation of the unimodal activations (AO + VO) has been used to demonstrate supra-additive activation, which is more suggestive of a cross-modal integration process. Rather than overall activation while the stimulus is present, the information theoretic approach instead focuses on quantifying the degree to which the changing speech time course is encoded or represented in the neural signals. The MI calculated here is an effect size for the ongoing entrainment of the MEG time course by the time varying speech—i.e., it quantifies the strength of the representation of dynamic audiovisual speech in the neural activity. While the basic expression on which our redundancy measure is based (Materials and methods, Eq 1) looks similar to an activation contrast (e.g., sum versus conjunction), it is important to keep in mind that this is about the strength of the dynamic low-frequency entrainment in each modality, not simply overall activation contrasts between conditions as in the classic fMRI approach.
The PID can quantify the representational interactions between multiple sensory signals and the associated brain response in a single experimental condition in which both sensory modalities are simultaneously present. In the PID framework, the unique contributions of a single (e.g., auditory) sensory modality to brain activity are directly quantified when both are present, instead of relying on the statistical contrast between modalities presented independently. Furthermore, the PID method allows the quantification of both redundant and synergistic interactions. In the context of audiovisual integration, both types of interaction can be seen as integration effects. Redundant information refers to quantification of overlapping information content of the predictor variables (auditory and visual speech signals), and synergistic information refers to additional information gained from simultaneous observation of two predictor variables compared to observation of one. Both of these types of interaction quantify multimodal stimulus representation that cannot be uniquely attributed to one of the two modalities. Redundant representation cannot be uniquely attributed, since that part of the brain response could be predicted from either of the stimulus modalities. Synergistic representation cannot be uniquely attributed, since that part of the brain response could only be predicted from simultaneous observation of both modalities and not from either one alone.
Note that these statistical interactions are quite different from interaction terms in a linear regression analysis, which would indicate the (linear) functional relationship between one stimulus modality and the MEG response is modulated by the value of the other stimulus modality. MI is an effect size that can be interpreted, because of its symmetry, from both an encoding and decoding perspective. From an encoding perspective, MI is a measure of how much an observer’s predictive model for possible MEG activity values changes when a specific auditory speech value is observed 100 ms prior. It quantifies the improvement in predictive performance of such an observer when making an optimal guess based on the auditory speech signal they see, over the guess they would make based on overall MEG activity without observing a stimulus value. From this perspective, redundancy quantifies the overlapping or common predictions that would be made by two Bayesian optimal observers, one predicting based on the auditory signal and the other the visual. Synergy is an increase in predictive power when both signals are obtained simultaneously. That is, it is possible to obtain a better prediction of the MEG with simultaneous knowledge of the specific combination of A and V observed than it is from combining only the predictions of the previous two unimodal observers. From considering the local plots (i.e., the values that are summed to obtain the final expectation value) in S3 Fig, we can see that a better prediction of the MEG in left motor cortex is made from the joint multimodal input in the case in which the MEG signal is high (above median), and the auditory and visual signals are in opposite ranges (e.g., high/low or low/high).
Existing techniques like representational similarity analysis (RSA) [26] and cross-decoding [27] can address the same conceptual problem as redundancy but from the angle of similarity of representations on average rather than specific overlapping Bayesian predictive information content within individual samples, which the information theoretic framework provides. Techniques exploiting decoding in different conditions can show the degree to which multimodal representations are similar to unimodal representations [28, 29] and whether there is an improvement in performance when the representation is learned in the multimodal condition. However, PID is explicitly a trivariate analysis considering two explicit quantified stimulus features and the brain signal. The information theoretic definition of synergy means there is enhanced prediction of neural responses from simultaneous multimodal stimuli compared to independent predictions combined from each modality (but still presented together). This differs from typical multimodal versus unimodal contrasts, even those involving decoding, because it explicitly considers the effect of continuous naturalistic variation in both stimulus modalities on the recorded signal.
Posterior superior temporal region (pSTG/S) has been implicated in audiovisual speech integration area by functional [30–32] and anatomical [33] neuroimaging. A typical finding in fMRI studies is that pSTG/S shows stronger activation for audiovisual (AV) compared to auditory-only (AO) and/or visual-only (VO) conditions. This was confirmed by a combined fMRI-transcranial magnetic stimulation (TMS) study in which the likelihood of McGurk fusion was reduced when TMS was applied individually to fMRI-localized posterior superior temporal sulcus (pSTS), suggesting a critical role of pSTS in auditory-visual integration [14].
The redundant information in the same left superior temporal region in this study matches this notion that this region processes shared information from both modalities. We found this region not only in the congruence effect (“AV congruent” > “All incongruent”; Fig 4A) but also in the attention effect (“AV congruent” > “All congruent”; Fig 5A).
We found the left motor cortex shows increased synergy for the matching versus nonmatching audio stimuli of “AV congruent” condition (Fig 3B). However, further analysis optimized for effects of attention and congruence revealed slightly different areas—with the area that shows strongest synergy change with attention (Fig 5B; BA6) located more lateral and anterior compared to the area identified in the congruence (Fig 4B). Previous studies have demonstrated increased phase locking of left motor cortex activity to frequency-tagged stimuli during auditory spatial attention [34, 35]. We extend these findings by demonstrating attention-mediated synergistic interactions of auditory and visual representations in left motor cortex.
The motor region in the attention contrast is consistent with the area in our previous study that showed entrainment to lip movements during continuous speech that correlated with speech comprehension [5]. In another study, we identified this area as the source of top-down modulation of activity in the left auditory cortex [23]. The definition of synergistic information in our context refers to more information gained from the simultaneous observation of auditory and visual speech compared to the observation of each alone. When it comes to the attention effect (“AV congruent” > “All congruent”), “AV congruent” condition requires paying more attention to auditory and visual speech than the “All congruent” condition does, even though the speech signals to be attended match the visual stimulus in both conditions. Thus, this synergy effect in the left motor cortex can be explained by a net attention effect at the same level of stimulus congruence. This effect is likely driven by stronger attention to visual speech, which is informative for the disambiguation of the two competing auditory speech streams [5]. This notion is plausible because it is supported by directional information analysis that shows that the left motor cortex better predicts upcoming visual speech in the “AV congruent” condition, in which attention to visual speech is crucial (S2B and S2D Fig).
However, a number of open questions in need of further investigation still remain. First, the auditory speech envelope and lip area information used in our analysis only capture part of the rich audiovisual information that is available to interlocutors in a real-life conversation. Other, currently unaccounted features might even be correlated across modalities (e.g., a different visual feature that is correlated with the auditory envelope). Since our analysis is restricted to these two features, it is possible that with a richer feature set for each modality, the unique information obtained from each would be reduced. In addition, the auditory speech signal is available at a much higher temporal resolution compared to the lip area signal, leading to a potential bias in the information content of both signals. Since the analysis of speech-brain coupling is a relatively new research field, we envisage methodological developments that will capture more aspects of the rich audiovisual signals. But in the context of our analysis that is focused on syllable components in speech, it seems reasonable to use these two signals that are known to contain clear representations of syllable-related frequencies [18, 36].
Second, it should be noted that we computed PID measures on the speech signals and 100 ms shifted MEG signal as in previous analyses [5, 18, 23] to compensate for delays between stimulus presentation and main cortical responses. We have confirmed that this (on average) maximizes speech-brain coupling. However, different aspects of multisensory integration likely occur at different latencies, especially in higher-order brain areas. This highly interesting but complex question is beyond the scope of the present study but will hopefully be addressed within a similar framework in future studies.
Third, while an unambiguous proof is missing, we believe that converging evidence suggests that participants attended visual speech more in “AV congruent” condition than in the other conditions. Indeed, it seems very unlikely that participants did not attend to visual speech after being explicitly instructed to attend (especially because visual speech provided important task-relevant information in the presence of a distracting auditory input). The converging evidence is based on behavioral performance, eye tracking results, and previous studies. Previous research indicates that the availability of visual speech information improves speech intelligibility under difficult listening conditions [1, 6, 37]. The “AV congruent” condition was clearly more difficult compared to the “All congruent” condition because of the presence of an interfering auditory stimulus. One could argue that participants could accomplish the task by simply using auditory spatial attention. However, our behavioral data (see Fig 1B in [5]) argue against this interpretation. If participants had ignored the visual stimulus and only used auditory spatial attention, then we would expect to see the same behavioral performance between “AV congruent” and “All incongruent” conditions. In both cases, two different auditory stimuli were presented, and only relying on auditory information would lead to the same behavioral performance. Instead, we find a significant difference in behavioral performance between both conditions. The availability of the congruent visual stimulus (in the “AV congruent” condition) resulted in a significant increase of behavioral performance (compared to “All incongruent” condition) to the extent that it reached the performance for the “All congruent” condition (no significant difference between “All congruent” and “AV congruent” conditions measured by comprehension accuracy; mean ± s.e.m; 85.0% ± 1.66% for “All congruent,” 83.40% ± 1.73% for “AV congruent” condition). This is strong evidence that participants actually made use of the visual information. In addition, this is also supported by eye fixation on the speaker’s lip movement, as shown in S5 Fig.
In summary, we demonstrate how information theoretic tools can provide a new perspective on audiovisual integration, by explicitly quantifying both redundant and synergistic cross-modal representational interactions. This reveals two distinct profiles of audiovisual integration that are supported by different brain areas (left motor cortex and left pSTG/S) and are differentially recruited under different listening conditions.
Data from 44 subjects were analyzed (26 females; age range: 18–30 y; mean age: 20.54 ± 2.58 y). Another analysis of these data was presented in a previous report [5]. All subjects were healthy, right-handed (confirmed by Edinburgh Handedness Inventory [38]), and had normal or corrected-to-normal vision and normal hearing (confirmed by 2 hearing tests using research applications on an iPad: uHear [Unitron Hearing Limited] and Hearing-Check [RNID]). None of the participants had a history of developmental, psychological, or neurological disorders. They all provided informed written consent before the experiment and received monetary compensation for their participation. The study was approved by the local ethics committee (CSE01321; College of Science and Engineering, University of Glasgow) and conducted in accordance with the ethical guidelines in the Declaration of Helsinki.
We used audiovisual video clips of a professional male speaker talking continuously (7–9 min), which were used in our previous study [5]. Since in some conditions (“AV congruent,” “All incongruent” conditions) the auditory speeches are delivered dichotically, to ensure that there are no differences other than talks themselves in those conditions, we made all the videos with the same male speaker. The talks were originally taken from TED talks (www.ted.com/talks/) and edited to be appropriate to the stimuli we used (e.g., editing words referring to visual materials, the gender of the speaker, etc.).
High-quality audiovisual video clips were filmed by a professional filming company, with sampling rate of 48 kHz for audio and 25 frames per second (fps) for video in 1,920 × 1,080 pixels.
In order to validate stimuli, 11 videos were rated by 33 participants (19 females; aged 18–31 y; mean age: 22.27 ± 2.64 y) in terms of arousal, familiarity, valence, complexity, significance (informativeness), agreement (persuasiveness), concreteness, self-relatedness, and level of understanding, using Likert scale [39] 1–5 (for an example of concreteness, 1: very abstract, 2: abstract, 3: neither abstract nor concrete, 4: concrete, 5: very concrete). Eight talks were finally selected for the MEG experiment by excluding talks with mean scores of 1 and 5.
Questionnaires for each talk were validated in a separate behavioral study (16 subjects; 13 females; aged 18–23 y; mean age: 19.88 ± 1.71 y). These questionnaires are designed to assess the level of speech comprehension. Each questionnaire consists of 10 questions about a given talk to test general comprehension (e.g., “What is the speaker’s job?”) and were validated in terms of accuracy (the same level of difficulty), response time, and the length (word count).
Experimental conditions used in this study were “All congruent,” “All incongruent,” and “AV congruent.” In each condition (7–9 min), 1 video recording was presented, and 2 (matching or nonmatching) auditory recordings were presented to the left and the right ear, respectively. Half of the 44 participants attended to speech in the left ear, and the other half attended to speech in the right ear.
The “All congruent” condition is a natural audiovisual speech condition in which auditory stimuli to both ears and visual stimuli are congruent (V1A1A1; the first A denotes talk presented to the left ear, and the second A denotes talk presented to the right ear; the number refers to the identity of the talks). The “All incongruent” condition has three different stimulus streams from three different videos, and participants are instructed to attend to auditory information presented to one ear (V1A2A3). The “AV congruent” condition consists of one auditory stimulus matching the visual information, and the speech presented to the other ear serves as a distractor. Participants attend to the talk that matches visual information (V1A1A2 for left ear attention group, V1A2A1 for right ear attention group). Each condition represents one experimental block, and the order of conditions was counterbalanced across subjects.
Participants were instructed to fixate on the speaker’s lip throughout the presentation in all experimental conditions, and we monitored the eye gaze using an eye tracker. Furthermore, we explained the importance of eye fixation on the speaker’s lip movement during the instruction session. They were also informed that for this reason, their eye movement and gaze behavior would be monitored using an eye tracker (see eye tracker data analysis in S5 Fig).
A fixation cross (either yellow or blue color) was overlaid on the speaker’s lip during the whole video for mainly two reasons: (1) to help maintain eye fixation on the speaker’s lip movement and (2) to indicate the auditory stimulus to pay attention to (left or right ear; e.g., “If the color of fixation cross is yellow, please attend to left ear speech”). The color was counterbalanced across subjects (for half of participants, yellow indicates attention to the left ear speech; for another half, attention to the right ear speech). This configuration was kept the same for all experimental conditions to ensure the same video display other than the experimental manipulations we aimed at. However, in “All congruent” condition (natural audiovisual speech), in which 1 auditory stream is presented diotically, attention cannot be directed to left or right ear, so participants were instructed to ignore the color of the fixation cross and just to attend the auditory stimuli naturally. In addition, to prevent stimulus-specific effects, we used 2 sets of stimuli consisting of different combinations of audiovisual talks. These 2 sets were randomized across participants (set 1 for half of participants, set 2 for the other half). For example, talks for “All congruent” condition in set 1 were talks for “AV congruent” condition in set 2.
There was no significant difference in comprehension accuracy between left and right ear attention groups (two-sample t test, df: 42, P > 0.05). In this study, we pooled across both groups for data analysis so that attentional effects for a particular side (e.g., left or right) are expected to cancel out.
For the recombination and editing of audiovisual talks, we used Final Cut Pro X (Apple, Cupertino, CA). The stimuli were presented with Psychtoolbox [40] in MATLAB (MathWorks, Natick, MA). Visual stimuli were delivered with a resolution of 1,280 × 720 pixels at 25 fps (mp4 format). Auditory stimuli were delivered at a 48 kHz sampling rate via a sound pressure transducer through 2 five-meter-long plastic tubes terminating in plastic insert earpieces.
A comprehension questionnaire was administered about the attended speech separately for each condition.
Cortical neuromagnetic signals were recorded using a 248 magnetometers whole-head MEG system (MAGNES 3600 WH, 4-D Neuroimaging) in a magnetically shielded room. The MEG signals were sampled at 1,017 Hz and were denoised with information from the reference sensors using the denoise_pca function in FieldTrip toolbox [41]. Bad sensors were excluded by visual inspection, and electrooculographic (EOG) and electrocardiographic (ECG) artifacts were eliminated using independent component analysis (ICA). An eye tracker (EyeLink 1000, SR Research) was used to examine participants’ eye gaze and movements to ensure that they fixated on the speaker’s lip movements.
Structural T1-weighted MRIs of each participant were acquired at 3 T Siemens Trio Tim scanner (Siemens, Erlangen, Germany) with the following parameters: 1.0 × 1.0 × 1.0 mm3 voxels; 192 sagittal slices; field of view (FOV): 256 × 256 matrix.
Information theoretic quantities were estimated with the Gaussian-Copula Mutual Information (GCMI) method [42] (https://github.com/robince/gcmi). PID analysis was performed with the GCMI approach in combination with an open source PID implementation in MATLAB, which implements the PID [19, 20] with a redundancy measure based on common change in local surprisal [15] (https://github.com/robince/partial-info-decomp). For statistics and visualization, we used the FieldTrip Toolbox [41] and in-house MATLAB codes. We followed the suggested guidelines [43] for MEG studies.
|
10.1371/journal.pcbi.1004429 | Nonconsensus Protein Binding to Repetitive DNA Sequence Elements Significantly Affects Eukaryotic Genomes | Recent genome-wide experiments in different eukaryotic genomes provide an unprecedented view of transcription factor (TF) binding locations and of nucleosome occupancy. These experiments revealed that a large fraction of TF binding events occur in regions where only a small number of specific TF binding sites (TFBSs) have been detected. Furthermore, in vitro protein-DNA binding measurements performed for hundreds of TFs indicate that TFs are bound with wide range of affinities to different DNA sequences that lack known consensus motifs. These observations have thus challenged the classical picture of specific protein-DNA binding and strongly suggest the existence of additional recognition mechanisms that affect protein-DNA binding preferences. We have previously demonstrated that repetitive DNA sequence elements characterized by certain symmetries statistically affect protein-DNA binding preferences. We call this binding mechanism nonconsensus protein-DNA binding in order to emphasize the point that specific consensus TFBSs do not contribute to this effect. In this paper, using the simple statistical mechanics model developed previously, we calculate the nonconsensus protein-DNA binding free energy for the entire C. elegans and D. melanogaster genomes. Using the available chromatin immunoprecipitation followed by sequencing (ChIP-seq) results on TF-DNA binding preferences for ~100 TFs, we show that DNA sequences characterized by low predicted free energy of nonconsensus binding have statistically higher experimental TF occupancy and lower nucleosome occupancy than sequences characterized by high free energy of nonconsensus binding. This is in agreement with our previous analysis performed for the yeast genome. We suggest therefore that nonconsensus protein-DNA binding assists the formation of nucleosome-free regions, as TFs outcompete nucleosomes at genomic locations with enhanced nonconsensus binding. In addition, here we perform a new, large-scale analysis using in vitro TF-DNA preferences obtained from the universal protein binding microarrays (PBM) for ~90 eukaryotic TFs belonging to 22 different DNA-binding domain types. As a result of this new analysis, we conclude that nonconsensus protein-DNA binding is a widespread phenomenon that significantly affects protein-DNA binding preferences and need not require the presence of consensus (specific) TFBSs in order to achieve genome-wide TF-DNA binding specificity.
| Interactions between proteins and DNA trigger many important biological processes. Therefore, to fully understand how the information encoded on the DNA transcribes into RNA, which in turn translates into proteins in the cell, we need to unravel the molecular design principles of protein-DNA interactions. It is known that many interactions occur when a protein is attracted to a specific short segment on the DNA called a specific protein-DNA binding motif. Strikingly, recent experiments revealed that many regulatory proteins reproducibly bind to different regions on the DNA lacking such specific motifs. This suggests that fundamental molecular mechanisms responsible for protein-DNA recognition specificity are not fully understood. Here, using high-throughput protein-DNA binding data obtained by two entirely different methods for ~100 TFs in each case, we show that DNA regions possessing certain repetitive sequence elements exert the statistical attractive potential on DNA-binding proteins, and as a result, such DNA regions are enriched in bound proteins. This is in agreement with our previous analysis performed for the yeast genome. We use the term nonconsensus protein-DNA binding in order to describe protein-DNA interactions that occur in the absence of specific protein-DNA binding motifs. Here we demonstrate that the identified nonconsensus effect is highly significant for a variety of organismal genomes and it affects protein-DNA binding preferences and nucleosome occupancy at the genome-wide level.
| Binding of TFs to their target sites on the DNA is a key step during gene activation and repression. An existing paradigm assumes that the main mechanism responsible for specific TF-DNA recognition is TF binding to short (typically 6–20 bp long) DNA sequences called specific consensus motifs, or specific TF binding sites (TFBSs). It has been known for a long time, since the seminal studies of Iyer and Struhl [1], that genomic context surrounding specific TFBSs significantly influences TF-DNA binding preferences. However, general rules describing the mechanisms responsible for such influences remain unknown.
Recently, the model organism ENCODE (modENCODE) project has revealed genome-wide comprehensive maps of TF-DNA binding and nucleosome occupancy in C. elegans [2–7] and in D. melanogaster [8–10]. Remarkably, these studies have challenged the existing paradigm and revealed that a large fraction of TF-DNA binding events occurs in genomic regions depleted of specific consensus motifs. Such genomic regions with enhanced overall TF-DNA binding but depleted in consensus motifs are oftentimes of low sequence complexity, which means that they are enriched in repeated DNA sequences.
We have recently proposed that repetitive DNA sequences characterized by certain symmetries and length scales of repetitive sequence patterns (see below) exert a statistical potential on DNA-binding proteins, affecting their binding preferences [11–15]. This effect of protein binding to repetitive DNA sequences in the absence of specific base-pair recognition is different from the concept of nonspecific protein-DNA binding introduced and explored in seminal studies of von Hippel, Berg, et al. [16–21]. In particular, von Hippel and Berg defined two related mechanisms for nonspecific protein-DNA binding [19]. The first mechanism is DNA sequence-independent, and it assumes that DNA exerts an electrostatic attraction upon DNA-binding proteins, modulated by the overall DNA geometry [19]. It has been proposed that DNA-binding proteins use different conformations in specific and nonspecific binding modes [16–20, 22]. The second mechanism assumes that mutated specific DNA consensus motifs retain a reduced binding affinity for sequence-specific TFs [19]. Nonspecific protein-DNA binding might become significant since the statistical probability to find such imperfect motifs in many genomic locations by random chance is high for eukaryotic genomes [19, 23]. The importance of nonspecific protein-DNA binding has been experimentally demonstrated for a number of systems both in vivo [24, 25] and in vitro [26–31].
We demonstrated recently that repetitive DNA sequence patterns characterized by certain symmetries lead to nonconsensus protein-DNA binding that can be enhanced or reduced depending on the symmetry type [11]. We use the term nonconsensus protein-DNA binding in order to emphasize the point that the nonconsensus protein-DNA binding free energy is computed without using any experimental information on specific protein-DNA binding preferences (see below). For example, we showed that repetitive homo-oligonucleotide sequence patterns, such as repeated poly(A)/poly(T)/poly(C)/poly(G) tracts lead to statistically enhanced nonconsensus protein-DNA binding affinity [11]. Our results indicated that such nonconsensus binding significantly influences nucleosome occupancy [12], TF-DNA binding preferences [13], and transcription pre-initiation complex binding preferences [14] in yeast.
In addition, using the protein binding microarray (PBM) method, we have recently directly measured the nonconsensus protein-DNA binding free energy for several human TFs [15]. We have demonstrated that, remarkably, the magnitude of the identified nonconsensus effect reaches as much as 66% of consensus (specific) binding [15].
In this study we explore the extent and significance of the nonconsensus protein-DNA binding mechanism for a large number of proteins belonging to different structural families. First, we investigate the nonconsensus effect in more complex, multicellular organisms, using the available ChIP-seq data obtained for ~100 TFs in C. elegans [2, 3] and D. melanogaster [10, 32]. Next, we perform the analysis of high-resolution in vitro universal protein-DNA binding microarray (PBM) data obtained for ~90 eukaryotic TFs belonging to 22 different DNA-binding domain types [33–35]. In addition, we identify protein sequence features that statistically distinguish between proteins with stronger and weaker response to nonconsensus repetitive DNA sequence elements, respectively.
We stress the point that in vitro analysis is free of confounding factors present in a cell, such as nucleosomes and indirect TF-DNA binding. Our previous experimental in vitro study of nonconsensus protein-DNA binding was performed for only 6 TFs [15]. The present analysis of the vast amount of in vitro TF-DNA binding data extends this number to more than an order of magnitude, suggesting that the nonconsensus mechanism most likely represents the statistical law rather than the exception. Therefore, the results reported here strongly support our conclusion that nonconsensus protein-DNA binding is a widespread phenomenon that significantly affects protein-DNA binding preferences in eukaryotic genomes, and need not require the presence of consensus (specific) TFBSs in order to achieve genome-wide TF-DNA binding specificity.
We compared the predicted landscape of nonconsensus protein-DNA binding free energy with the genomic binding profiles of 69 transcriptional regulators in C. elegans [2, 3] and 30 transcriptional regulators in D. melanogaster [10, 32], as determined by ChIP-seq in the modENCODE project [2, 3, 8, 10]. We computed the nonconsensus binding free energy landscape using a simple approach that we developed previously [11]. Briefly, we used a set of random protein-DNA binders as a proxy for nonspecific protein-DNA interactions in a crowded cellular environment (Fig 1). Next, to each location along the C. elegans and D. melanogaster genomes, we assigned an average free energy of nonconsensus protein-DNA binding, 〈F〉TF, where the averaging is performed over an ensemble of random binders (see Methods for further details). The free energy value at each sequence location is entropy-dominated, and it is influenced exclusively by the presence of repetitive DNA sequence patterns [11] surrounding that location. We use the term DNA sequence correlations to describe the repetitive DNA patterns, and the term correlation scale to describe the length of the patterns (Methods). The larger the correlation scale, the larger the number of repetitive sequence patterns, and thus the stronger the nonconsensus protein-DNA binding effect [11]. Importantly, the genomic DNA sequence constitutes the only input for the nonconsensus binding model, i.e. the model does not have any fitting parameters (Methods).
We found that the nonconsensus protein-DNA binding free energy correlates negatively with the combined TF occupancy in both the C. elegans and the D. melanogaster genomes, i.e. the lower the nonconsensus binding free energy, the higher the combined TF occupancy (Fig 2). Fig 2a and 2c illustrate this correlation for free energy profiles, 〈〈F〉TF〉seq, averaged over genomic sequences aligned with respect to the TSS. A statistically significant correlation at the single gene level is also observed, on average, without sequence alignment with respect to the TSS (Fig 2b and 2d). In these analyses both genomes show statistically significant negative correlations, with the correlation being more pronounced in C. elegans.
We verified that the predicted free energy landscape is qualitatively robust with respect to variations in the model parameters (i.e. the sliding window width, L, and the TFBS size, M) (S1 Fig). In addition, we validated that the predicted free energy landscape is determined by the presence of repetitive sequence patterns, and not by the average genomic nucleotide content. To show this, we shuffled the DNA sequence in each sliding window along the genome to obtain random DNA sequences with a fixed nucleotide content, and we computed the normalized free energy, δF = F−Frand, where Frand is the free energy of the random, shuffled sequences, averaged over different random realizations (Methods). As shown in S2 Fig, the normalized free energy δF is robust with respect to global variations in the genomic nucleotide content.
The predicted reduction in the nonconsensus free energy upstream of TSSs (Fig 2a and 2c) stems from the enhanced level of homo-oligonucleotide sequence correlations (i.e. repetitive homo-oligonucleotide sequence patterns, such as repeated poly(dA:dT) tracts). This effect can be intuitively understood in the following way. As shown in our previous work, the presence of enhanced homo-oligonucleotide sequence correlations within a DNA region generally leads to the widening of the protein-DNA binding energy spectrum in this region [11]. For example, in the statistical ensemble of random binders interacting with DNA sequence that contains long homo-oligonucleotide tracts with two alternating types of nucleotides (such as alternating poly(dA:dT) and poly(dT:dA) tracts), the width (i.e. the standard deviation) of the binding energy spectrum, σUhomo, will be universally larger than the corresponding width for the case of entirely random DNA sequence, σUhomo≃2⋅σUrandom [11]. This result is independent of the microscopic details of the protein-DNA interaction potential, U, and it is simply the consequence of the central limit theorem [36, 37]. The wider energy spectrum, σUhomo>σUrandom, universally leads to the statistically lower free energy, Fhomo < F random [38], and therefore to a higher nonconsensus protein-DNA binding affinity. The computed probability distributions of the nonconsensus protein-DNA binding energy and the free energy in the C. elegans genome, further illustrates this mechanism (S3 Fig). Thus, the nonconsensus protein-DNA binding mechanism can significantly influence TF-DNA binding preferences in the C. elegans and D. melanogaster genomes, complementing the conventional, specific protein-DNA recognition mode.
We stress the fact that the minimum of the average nonconsensus protein-DNA binding free energy landscape does not align precisely with the maximum of the average TF occupancy profile in both C. Elegans and D. melanogaster genomes (Fig 2a and 2c). Such mismatch is also observed between the average nonconsensus protein-DNA binding free energy landscape and the average nucleosome profile (see below, Fig 3a and 3c), similar to the case as we previously observed for the yeast genome [12]. Combination of additional factors not taken into account in our model but present in vivo might explain a possible origin of such a mismatch. These factors include, first, steric constrains imposed by the presence of nucleosome particles [39]; second, steric constrains imposed by the transcription pre-initiation complex (PIC) [40]; and third, the presence of specific TFBSs [41].
We also assessed the effect of nonconsensus protein-DNA binding on nucleosome binding preferences in the C. elegans and D. melanogaster genomes. Genome-wide measurements of nucleosome occupancy show a typical nucleosome depleted region upstream of the TSSs, and a well-positioned +1 nucleosome [2, 4, 42]. In D. melanogaster, an oscillating nucleosome occupancy pattern was observed, similar to the one in yeast [43], while the C. elegans genome-wide nucleosome occupancy profile does not demonstrate such strong oscillations [4, 42].
The computed nonconsensus free energy landscapes show a statistically high, positive correlation with the nucleosome occupancy profile in both genomes (Fig 3). In particular, the average nonconsensus free energy shows a pronounced minimum in the upstream nucleosome depleted region (Fig 3a and 3c), similar to the one observed in yeast [12]. In Fig 3b and 3d we also observed, at the single gene level, statistically significant correlation between the average nucleosome occupancy and the average free energy of nonconsensus binding (Methods). Sequences with lower nonconsensus protein-DNA binding free energy have, on average, lower nucleosome occupancy.
We suggest that the observed effect stems from the competition between TFs that experience enhanced nonspecific attraction towards upstream promoter regions (i.e., reduced level of the nonconsensus free energy) and nucleosome-forming histones. It is important to stress that the presence of repetitive DNA sequence elements in promoter regions might also affect histone-DNA binding due to the nonconsensus mechanism, and as a result of it, the nucleosome formation. How exactly individual histones and histone complexes respond to different repetitive DNA sequence patterns remains an open question. This issue is further complicated by the fact that several additional mechanisms influence histone-DNA binding in promoter regions. Namely, genome-wide, in vitro nucleosome reconstruction experiments demonstrate that nucleosome-free regions (NFR) can be formed to some extend even in the mixture of purified genomic DNA with histones [44, 45]. However, intrinsic DNA sequence preferences of nucleosomes still remain an open issue [46]. In particular, it has been recently demonstrated that AT-rich sequences present in many NFRs have little effect on the stability of nucleosomes [46]. Rather it appears that ATP-dependent chromatin modifiers constitute a major factor regulating nucleosome-binding preferences in vivo [43, 46].
Here we provide an additional, highly significant validation for the proposed mechanism of nonconsensus protein-DNA binding by the analysis of the available in vitro TF-DNA binding data obtained using the protein-binding microarray (PBM) technology [35, 47–49]. The PBM technology allows to simultaneously measure binding of a TF to tens of thousands of 36-bp long DNA sequences in a single experiment [35]. The PBM method is free from the confounding factors, such as the effect of competing TFs and nucleosomes on TF-DNA binding preferences. Here, we used the currently available ‘universal PBM’ data for 91 TFs (belonging to 22 distinct DNA-binding domains) from C. elegans, D. melanogaster, and mus musculus [33, 34, 50] (Fig 4 and S1 Table). The DNA libraries used in these ‘universal PBM’ experiments were designed in such a way that they cover all possible 8-mer DNA sequences [35], giving an unbiased view of TF-DNA binding specificity. Overall, there are ~45,000 distinct DNA sequences in this library, and thus the TF-DNA binding strength was measured for each TF to all these sequences [33, 34, 50].
We computed the nonconsensus TF-DNA binding free energy, 〈f〉TF, for each 36-bp long DNA sequence in the library using the procedure described above (Fig 1). Contrary to the case of genomic sequences, here we do not move the sliding window along the DNA sequence since each sequence is short, L = 36 bp, and therefore a single value of 〈f〉TF is assigned to each DNA sequence.
Remarkably, for 69 out of 91 analyzed TFs (i.e. 76%) we detected a statistically significant, negative correlation between the nonconsensus protein-DNA binding free energy and the measured in vitro TF-DNA binding intensity. This is in agreement with the results obtained for the in vivo TF-DNA binding data (Fig 2b and 2d). Twelve TFs (i.e. 13%) did not show a statistically significant correlation, and interestingly, ten TFs (i.e. 11%) showed an opposite, positive correlation (S1 Table). The latter observation is remarkable, since it demonstrates that a non-negligible fraction of TFs can respond to DNA symmetries (represented by our free energy model) in an opposite way compared to the majority of other TFs. However, statistically, the average TF-DNA binding preferences show highly significant, negative correlation with the computed free energy of nonconsensus protein-DNA binding (Fig 4) in agreement with the in vivo results (Fig 2b and 2d).
In order to identify what structural and sequence features are responsible for the anomalous behavior of these 11% of TFs, we classified all TFs according to the DNA-binding domain (DBD) families they belong to. However, we have not identified any particular DBD families that are unique to those 11% of TFs (S1 Table and S4 Fig). We have also not identified any preference of these TFs with respect to any particular biological function, according to the gene ontology (GO) classification. Therefore, the question what sequence and structural features of proteins are responsible for the positive correlation between the free energy and the experimentally measured in vitro TF occupancy remains open.
Next, in order to identify protein sequence features that might be responsible for enhanced nonconsensus TF-DNA binding, we separated TFs (we used 82 mouse TFs for this analysis) into two groups. The first group contained 41 TFs with the strongest negative correlation between the free energy and the measured TF occupancy. The second group contained the remaining 41 TFs. We have analyzed the amino acid correlation properties in these two groups of TFs. Our working hypothesis here is that enhanced amino acid sequence correlations in TF sequences are responsible for enhanced nonconsensus TF-DNA binding. We use the term “sequence correlations” in order to describe repetitive sequence patterns. We have previously used a similar analysis in order to investigate protein sequence features responsible for enhanced level of protein structural disorder and protein-protein interaction promiscuity [36]. In particular, we have analyzed the frequency of occurrence of the following repetitive amino acid sequence patterns in each TF group: [aa], [aXa], [aXXa], and [aXXXa], where a represents each amino acid type and X represents an arbitrary amino acid (S2 Table). For example, when we compute the frequency of [Lys-X-Lys] pattern, we count the total number of the occurrence of this pattern in each protein sequence, irrespectively to the identity of X. As a result of this analysis, we have identified three patterns that demonstrated a statistically significant difference of frequencies between the two TF groups: [Lys-XX-Lys] (enriched in the first TF group; Kolmogorov-Smirnov p-value, pks ≃ 0.01), [Arg-Arg] (enriched in the second TF group; pks ≃ 0.02), and [Leu-X-Leu] (enriched in the first TF group;pks ≃ 0.05) (S2 Table). In addition the overall compositional fraction of Lys was enriched in the first TF group (pks ≃ 0.01) (S2 Table). The fact that the most statistically significant enrichment (distinguishing the two TF groups) is observed for the [Lys-XX-Lys] and [Arg-Arg] patterns is encouraging since positively charged Lys and Arg are obviously the key amino acids responsible for TF binding to the negatively charged DNA molecule.
Two conclusions can be drawn from our results. First, that the intrinsic propensity for nonconsensus protein-DNA binding is imprinted both into the DNA and the protein. Since our simple nonconsensus binding model treats proteins as random binders, it captures general trends in the binding profiles of most, but not all, TFs. Second, nonconsensus and specific (consensus) protein-DNA binding mechanisms are tightly interlinked, and both of these mechanisms cooperate in determining the overall protein-DNA binding preferences in eukaryotic genomes. The fact that our simple random-binder model (without any fitting parameters and without any protein-DNA binding specificity built in) provides such a good statistical description of the measured DNA binding strength for the majority of TFs strongly suggests that the nonconsensus mechanism is quite general and it represents the statistical law rather than the exception. However, more accurate, atomistic models describing nonconsensus protein-DNA binding interactions are necessary in order to improve the accuracy of our predictions for different proteins.
Our analyses of the effect of nonconsensus protein-DNA binding demonstrate that the combined genome-wide binding preferences of 69 TFs in C. elegans and 30 TFs in D. melanogaster are significantly, negatively correlated with the predicted nonconsensus free energy landscape (Fig 2). Our analyses also show that the experimentally derived nucleosome occupancy in C. elegans and in D. melanogaster is significantly, positively correlated with the predicted nonconsensus protein-DNA binding free energy (Fig 3). This trend is qualitatively similar to the one that we previously observed in yeast [12]. The results shown in Figs 2 and 3 strongly suggest that TFs compete with nucleosomes for nonconsensus binding to DNA. Such a competition between TFs and nucleosomes could lead to the enhanced TF binding cooperativity previously predicted by Mirny [51] and Teif et al. [52]. We suggest that nonconsensus protein-DNA binding greatly enhances such nucleosome-induced cooperativity between TFs, and most importantly, in order to achieve this enhancement, promoters do not require the presence of specific, consensus TF binding sites. We stress the important point that the predicted effect of nonconsensus TF-DNA binding most likely affects many but not all TFs. We expect for example, that stress response TFs, such as for example Msn2 in yeast [53], might be insignificantly influenced by the nonconsensus mechanism.
Our model predicts that genomic loci enriched with repetitive sequences, such as in heterochromatin, should also be enriched with TF binding. However, the ChIP-seq analysis in such regions is impeded by the fact that multi-mapping reads from long repetitive region will be filtered out by most peak-calling algorithms, therefore identifying interactions in these regions remains a challenging problem [54]. Interestingly, there are evidences that regions of heterochromatin are not actually transcriptionally inert and non-coding RNA molecules are transcribed from repeated DNA sequences in pericentromeric heterochromatin in different eukaryotic genomes [55]. A recent study even demonstrated [56] that some TFs bind directly to the major satellite repeat DNA sequences that are present in pericentromeric heterochromatin regions and might play a significant role in the mouse heterochromatin formation. Further experiments and analysis of TF binding to the heterochromatin would reveal whether nonconsensus binding play an important role in these regions as well.
Our analysis of available in vitro TF-DNA binding data from protein-binding microarray (PBM) experiments (Fig 4 and S1 Table) demonstrates that statistically, on average, in vitro TF-DNA binding preferences negatively correlate with the computed nonconsensus free energy landscape, and showed qualitatively similar behavior to the one observed in vivo (compare Fig 2b and 2d with Fig 4). This additional analysis is important for several reasons. First, the in vitro TF-DNA binding preferences are not affected by the presence of other proteins and histones, which can compete with the protein or cause an indirect binding to the DNA. Second, the TF binding intensity is measured in PBM experiments at significantly higher accuracy compared to ChIP-seq experiments. Third, the usage of non-genomic sequences that cover all possible 8-mer DNA sequences, eliminates possible sequence bias that might exist in the genomic sequences, and thus PBM measurements provide an entirely independent validation of the nonconsensus protein-DNA binding effect. Finally, the present analysis performed for ~90 TFs extends our previous analysis performed for only 6 TFs [15] by more than an order of magnitude, thus strongly suggesting the generality of the nonconsensus protein-DNA binding effect in eukaryotic genomes.
Interestingly, ten TFs (i.e. 11%) showed an opposite, positive correlation between the free energy and the measured TF-DNA occupancy (S1 Table). The latter observation is remarkable, since it demonstrates that a non-negligible fraction of TFs can respond to DNA symmetries (represented by our free energy model) in an opposite way compared to the majority of other TFs. However, we failed to identify any particular structural, sequence, or functional features unique to this set of TFs. This failure might stem from the small number of proteins that exhibited such behavior. Yet, we were able to identify repetitive amino acid sequence patterns that are responsible for enhanced nonconsensus TF-DNA binding (S2 Table). In particular, for the group of TFs characterized by the strongest nonconsensus TF-DNA binding preferences, the most statistically significant enrichment is observed for the [Lys-XX-Lys] pattern, while the frequency of [Arg-Arg] pattern is reduced in this group (S2 Table). The latter result is intuitively sound since both Lys and Arg are the key amino acids responsible for TF binding to the negatively charged DNA molecule.
Importantly, in this study, our random-binder statistical mechanics model for protein-DNA interactions does not use any experimentally pre-determined information on either low-affinity or high-affinity TF-DNA binding sites. The genomic DNA sequence constitutes the only experimental parameter of the model. In addition, our model does not have any fitting parameters. Contrary to the case of specific protein-DNA binding that requires the presence of a 6 to 20-bp long specific DNA motif (unique for each individual TF), the nonconsensus protein-DNA binding effect stems from multiple nonspecific interactions between the TF and a relatively long (few tens of bp) DNA fragments enriched with repetitive sequence patterns. The fact that different TFs are affected in a statistically similar way by entirely different DNA sequences containing similar repetitive patterns constitutes the key difference between the nonconsensus and specific protein-DNA recognition modes.
What exactly is the interplay between nonconsensus DNA repetitive sequence elements and consensus (specific) sequences and how their combination influences the overall binding of proteins to the DNA and the expression levels of genes are important questions yet to be explored. We suggest that repetitive nonconsensus sequence elements might have similar influence on TF-DNA binding and on gene expression as repeats of consensus (specific) DNA sequence elements (i.e. homotypic clusters) [57]. However, an important difference between these two types of repeated sequence elements is that nonconsensus repeats can affect many different TFs in a similar way, while homotypic clusters are more specific to a limited set of TFs.
Repetitive sequence elements located near the consensus (specific) motif, could increase the TF association rate, by inducing the one-dimension “sliding” of the TF, and improving its search for the specific binding site [20, 58]. The presence of many weaker sites flanking a strong binding site could lead to a funnel effect [59–62], where the molecules are directed to the strong binding site as depicted in S5a Fig. It could also stabilize binding sites that are not strong enough individually [63, 64] and increase the ability of binding sites to “withstand mutations” [65]. We use the C. elegans Hlh-1 protein as an example demonstrating that nonconsensus DNA sequence elements might stabilize the binding to specific consensus elements in vivo (S5b Fig). The analysis of Hlh-1 binding sites (based on the genome-wide ChIP-seq measurements [2, 3] in C. elegans) demonstrates that only 5% of the total number of Hlh-1 specific motifs in the genome is bound by Hlh-1 (S5b Fig). We sorted the genomic sequences containing the Hlh-1 motif (consensus motifs were reported in [3]) into two groups: the first group contains DNA sequences that were experimentally determined as being bound by Hlh-1, while the second group contains unbound DNA sequences. S5c Fig represents the average nonconsensus protein-DNA binding free energy computed for each of these two sequence groups. We observed that the nonconsensus free energy is reduced for the group that contains bound sequences as compared with the group that contains unbound sequences. The computed p-values show that this result is statistically significant (S5c Fig). This example supports the hypothesis that nonconsensus sequence elements might provide the funnel effect in vivo. Additional analysis and experimental measurements of the kinetics of TF-DNA binding to consensus (specific) sequence elements embedded in different nonconsensus DNA backgrounds, should shed more light on this hypothesis.
Future in vitro measurements of binding preferences for additional TFs [66], combined with high-resolution in vivo ChIP-seq and ChIP-exo analysis, will help to complete the molecular picture of design principles for nonconsensus protein-DNA binding and its functional significance.
We used the set of 23,287 C. elegans genes based on Wormbase annotation, WS228 [2, 67], and 12,188 D. melanogaster genes annotated in [10].
We used experimentally measured binding preferences of 69 C. elegans TFs (S3 Table), as determined by the Gerstein and Snyder labs [2, 3]; for computing the D. melanogaster TF occupancy we used binding preferences of 30 TFs (S4 Table) determined by the White lab [8]. TF-DNA binding preferences for both genomes were measured using ChIP-seq assays (modENCODE project). We defined TF occupancy for each genomic location as the total number of bound TFs at each location along the genome.
We used experimentally measured, genome-wide, normalized nucleosome occupancy determined by the paired-end Ilumina sequencing in C. elegans [4, 5]; we also used the genome-wide map of H2A.Z nucleosome occupancy in D. melanogaster embryos (0–12 hr) (determined in [32]).
We used experimentally measured in vitro binding intensity for the C. elegans, D. melanogaster, and mus musculus TFs (S1 Table), determined using the protein-binding microarray (PBM) technology [33, 35, 47–49].
In order to compute the nonconsensus protein-DNA binding free energy landscape, we generate an ensemble of random DNA binders as a proxy for the phenomenon of nonconsensus protein-DNA binding in a crowded cellular environment [11]. Our model does not use any experimentally pre-determined protein-DNA binding preferences in order to model protein-DNA binding. The actual DNA sequences of the C. elegans and D. melanogaster genomes constitute the only input parameter for our model. In order to compute the free energy of nonconsensus protein-DNA binding at any given location along a DNA sequence, we position the center of the sliding window of width L = 50 bp at that location. The 50 bp length is a typical sliding event distance of a protein along the DNA under physiological conditions [68, 69] (Fig 1).
We assume that a model protein (random binder) makes M bp contacts with the DNA (Fig 1b) and that the model protein-DNA interaction energy at each genomic position i is simply a sum of M interaction energies:
U(i)=−∑j=iM+i−1∑α={A,T,C,G}Kαsα(j)
(1)
where sα(j) represents the elements of a four-component vector of the type (δαA, δαT, δαC, δαG), and δαβ = 1 if α = β, or δαβ = 0 if α ≠ β. For example, if the A nucleotide is positioned at the coordinate j along the DNA, then this vector takes the form: (1,0,0,0). If, for example, the DNA sequence contains entirely poly(A) at a given genomic location, then a random binder makes all M contacts with the A nucleotide, and hence at this location the resulting energy, Eq (1), will be simply, MKA. In order to generate each model protein, we draw the values of KA, KT, KC, and KG from Gaussian probability distributions, P(Kα), with zero mean, and standard deviation σα = 2kBT, where T is the temperature and kB is the Boltzmann constant. We have shown previously that the resulting free energy is qualitatively robust with respect to the choice of model parameters [11]. The energy scale, 2kBT ≃ 1.2 kcal/mol, is chosen to represent a typical strength of a hydrogen bond, or an electrostatic bond that a protein makes with one DNA bp [16, 19].
For each model random binder, we define the partition function of protein-DNA binding within the chosen sliding window of width L bp:
Z=∑i=1Lexp(−U(i)/kBT)
(2)
and the corresponding free energy of nonconsensus protein-DNA binding in this sliding window:
F=−kBTlnZ
(3)
We then assign the computed F to the sequence coordinate in the middle of the sliding window. Next, we move the sliding window along the DNA sequence and we compute F at each sequence location. This procedure allows us to assign the free energy of nonconsensus protein-DNA binding to each DNA bp within the genome.
Next, we repeat the described procedure for an ensemble of 250 model random binders (Fig 1) and compute the average free energy, 〈FTF〉, over this ensemble, at each sequence location. We stress that the resulting free energy is qualitatively robust with respect to the choice of the sliding window size, L, within a wide range of values (S1 Fig). In addition, the free energy profiles are statistically robust with respect to a moderate variation of the value of M, within a typical range of the TF binding site size (S1 Fig). We verified that the predicted free energy landscape is dominated by DNA sequence correlations, and not by the average nucleotide composition (S2 Fig). In particular, for each random binder, in each sliding window we computed the normalized free energy, δF = F−Frand, where Frand is the free energy computed for a randomized sequence (in the same sliding window as F) and averaged over 25 random realizations.
In order to compute the p-value for S5c Fig, we first selected all the 800 bp-long sequences containing the exact binding motifs for each TF. For example, genome-wide, we have overall 9258 sequences containing the consensus Hlh-1motif. Among those 9258 sequences, 442 sequences were experimentally determined as bound by Hlh-1, while the rest of 8816 sequences were unbound. In order to compute the p-value, we compiled 105 pairs of groups containing 442 and 8816 sequences, respectively, randomly chosen from the original 9258 sequences. These 105 pairs of groups represent randomized analogs for the original groups of bound and unbound Hlh-1motifs. Second, for each of these pairs of random groups we computed the average free energies, 〈f〉, of nonconsensus binding separately for the randomized bound and unbound groups, as described above. Third, for each pair of randomized groups we computed the difference of the integrated free energy within the interval (-400,400) between the two randomized groups. Finally, we computed the probability that this difference is equal or larger than the actual value of the difference. The latter probability was taken as the p-value.
|
10.1371/journal.pntd.0006680 | The importance of dog population contact network structures in rabies transmission | Canine rabies transmission was interrupted in N’Djaména, Chad, following two mass vaccination campaigns. However, after nine months cases resurged with re-establishment of endemic rabies transmission to pre-intervention levels. Previous analyses investigated district level spatial heterogeneity of vaccination coverage, and dog density; and importation, identifying the latter as the primary factor for rabies resurgence. Here we assess the impact of individual level heterogeneity on outbreak probability, effectiveness of vaccination campaigns and likely time to resurgence after a campaign. Geo-located contact sensors recorded the location and contacts of 237 domestic dogs in N’Djaména over a period of 3.5 days. The contact network data showed that urban dogs are socially related to larger communities and constrained by the urban architecture. We developed a network generation algorithm that extrapolates this empirical contact network to networks of large dog populations and applied it to simulate rabies transmission in N’Djaména. The model predictions aligned well with the rabies incidence data. Using the model we demonstrated, that major outbreaks are prevented when at least 70% of dogs are vaccinated. The probability of a minor outbreak also decreased with increasing vaccination coverage, but reached zero only when coverage was near total. Our results suggest that endemic rabies in N’Djaména may be explained by a series of importations with subsequent minor outbreaks. We show that highly connected dogs hold a critical role in transmission and that targeted vaccination of such dogs would lead to more efficient vaccination campaigns.
| Rabies transmission between dogs and from dogs to humans can be interrupted by mass vaccination of dogs. Novel geo-referenced contact sensors tracked the contacts and locations of several hundred dogs in N’Djaména, the capital of Chad. With the data generated by the sensors we developed a contact network model for rabies transmission dynamics. The model results compared well to incidence data. The model explains the relationship between vaccination campaigns and number of cases better than previous models. Highly connected dogs play a critical role in rabies transmission and targeted vaccination of these dogs would lead to more efficient vaccination campaigns.
| The viral disease rabies, transmitted between mammals through bites, is fatal following the onset of symptoms. Although human rabies can be prevented by appropriate post-exposure prophylaxis (PEP), approximately 60,000 people die annually from rabies, mainly in Africa and Asia, [1]. The main source of exposure for human rabies is the domestic dog, so vaccinating dogs is an effective way of reducing rabies transmission among dogs and from dogs to humans [2, 3].
Rabies is endemic in N’Djaména, the capital city of Chad, with an average incidence of one laboratory-confirmed infected dog per week [4]. A deterministic model of rabies transmission predicted that mass vaccination of dogs would be sufficient to interrupt transmission for six years [2]. Vaccination campaigns in dogs were conducted in 2012 and 2013, with both campaigns exceeding 70% coverage [5]. Rabies transmission was interrupted in January 2014 after the second vaccination campaign [3], but there was a resurgence of cases nine months later. Subsequent analyses considered reasons for the quick resurgence, including spatial heterogeneity of vaccination coverage, and dog density; underreporting of cases; and importation. Simulation results from a deterministic metapopulation model suggested that importation was the most likely reason for the case resurgence [6]. Although deterministic models can predict the effect of large scale vaccination campaigns and the overall population dynamics, they do not adequately capture effects of stochasticity in low level endemic settings. This becomes important towards the end of an elimination campaign or upon re-establishment after interruption of transmission [7]. Previous models did not include fine scale heterogeneity at the individual level or the network structure of dog to dog contacts.
The importance of including host contact structure in infectious disease modelling has been highlighted in many studies [8–10]. Theoretical analysis of epidemic processes on graphs has shown that the basic reproductive ratio not only depends on the expected value but also on the standard deviation of the degree distribution of the graph [11] and that on scale-free networks diseases can spread and persist independently of the spreading rate [12]. These theoretical insights led to better understanding of disease transmission dynamics for different diseases, including pertussis [13], influenza [14], severe acute respiratory syndrome (SARS) [15], human immunodeficiency virus and acquired immune deficiency syndrome (HIV/AIDS) [16] and gonorrhea [17], and inspired novel control measures such as acquaintance immunization [18], contact tracing [19] and ring vaccination [20, 21].
Due to the substantial influence of network structure on disease transmission dynamics, many studies have collected data on host interactions. Human contact network models are generally established using contact diaries [22–24], proximity loggers [25–28], video recording [29] or mobile phones [30]. Contacts have also been studied in a wide range of animal species. The most common method for measuring animal contacts is behavioral observation, but other methods such as radio tracking, Global Positioning System (GPS) trackers, proximity loggers or powder marking are also utilized [31].
In the past decades, several rabies models with host contact structure have been published. White et al. [32] simulated fox movement pathways using home range size estimates, data from radio tracking and behavioral encounter observations to estimate contact probabilities for different seasons and fox densities. They found that the rabies front set off by an incursion of rabies into a healthy population moved more slowly than in a previous model of homogeneous fox populations. Including contact behavior in the model also resulted in a substantially higher predicted rabies control success rate. Hirsch et al. [33] used data from 30 raccoons fitted with proximity loggers to assess properties of the raccoon contact network. Unlike in earlier radiotelemetry studies, they found a highly connected population and discussed possible implications of the social network on the spread of rabies. Reynolds et al. [34] used proximity logger data from 15 raccoons to build a contact network model of 90 raccoons and simulate rabies spread. They studied the effects of seasonality, differences in vaccination coverage and impact of behavioral changes in infected raccoons on disease spread. Dürr and Ward [35] used a contact network model of rabies transmission among owned free-roaming dogs in Australia to estimate the impact of a hypothetical rabies incursion from Indonesia. They differentiated transmission within households, between households and between communities. The probability of between household transmission was based on GPS data from 69 dogs, while between community transmission was estimated using questionnaire data. Johnstone-Robertson et al. [36] developed a contact network model for rabies in the wild dog population in Australia. They constructed a function for dog contact probabilities, using a wide range of different values to generate contact networks and then implemented a rabies transmission model based on parameters from literature.
However, individual based models of dog rabies transmission in endemic settings are lacking, so this study equipped 300 dogs in N’Djaména with purpose developed geo-referenced contact sensors. This is the first study to collect contact data among dogs as well as the first to integrate contact data from such a large subset of an animal population into a rabies model. The individual based model of rabies transmission we developed includes distance between home locations and a degree distribution fit to a contact network structure of dogs in N’Djaména. We compared our model results to 2016 outbreak data from two quarters of N’Djamena. We examined the re-establishment probability of rabies over different vaccination coverage and compared outbreak probability over time with rabies incidence in N’Djaména from 2012 to 2016. Finally, we investigated the role of individual heterogeneity among dogs and the effect of targeted vaccination strategies.
Contact network data was collected in three districts of N’Djaména, Chad, using 300 geo-located contact sensors (GCS) developed specifically for this study. The devices contain Global Positioning System (GPS) modules to track the location and movements of dogs and Ultra-High-Frequency (UHF) technology sensors to measure close-proximity events between dogs. The GCS devices record locations at one minute intervals. For the contact recording, the devices broadcast beacons at one minute intervals and constantly scan for beacons ensuring that no contacts with durations of at least one minute will be missed. Close proximity events were defined as records with a received signal strength indicator (RSSI) of more than -75dBm. Static tests of the devices showed that, independently of the angle between two devices, all contacts closer than 25 cm are registered when signal strength is above that value (S1 Fig).
Collars fitted with the devices were placed on free roaming domestic dogs in three city districts (Table 1, Fig 1) with different dog densities (low, medium and high), that were easily accessible. The zones were chosen to include urban and peri-urban areas. Data were collected during the dry season in December 2016. In the selected districts, all dog-owning households in a pre-defined area of 1km2 were identified in order to capture as many of the contacts between dogs as possible, bearing in mind that only contacts between dogs that both wear a sensor can be captured. Dog owners were asked to enroll their pets. Only one dog owner refused to participate in the study. The GCS units remained on the dogs for 3.5 days. After retrieval of the GCS units, dogs were vaccinated against rabies. We excluded study zone 3 from the network analysis due to the low proportion of devices usable for analysis.
The data from the contact sensors were used to establish an empirical contact network, where the nodes correspond to the dogs and any two nodes are connected by an edge if at least one contact between the two dogs was registered. S2 Fig. shows the number of edges in the empirical network during different subintervals of the study period.
Surveillance of canine rabies in N’Djaména consists of passive reporting of cases confirmed with an immunofluorescence antibody test (IFAT). In 2012, prior to the vaccination campaign, there was, on average, one case of dog rabies per week. After the vaccination campaigns in 2012 and 2013, no rabies cases were reported for nine months. In October 2014, new rabies cases were reported in district number 9, south of the Chari River. In January 2016, the first case north of the river was reported in the Chagoua quarter of district 6 (Fig 1). An additional 6 cases of dog rabies were reported in 2016 in Chagoua and the neighboring Abena quarter.
We simulated rabies incursion into Chagoua and Abena quarters to compare the model results to the outbreak data. Dog population estimates were derived from the 2012 mass vaccination campaign coverage assessment to determine the number of nodes in the network. A total of 2775 dogs were vaccinated during the 2012 campaign in Chagoua, Abena and the neighboring Dembe quarters [5]. A capture-mark-recapture model estimated vaccination coverage in that area at 67%. In a second stage of the campaign, additional dogs were vaccinated in Chagoua, Abena and Dembe. During the latter stage, the proportion of dogs originating from Chagoua and Abena was assessed at 86% of dogs. Assuming that this proportion was the same in the first round, we estimated the dog population in Chagoua and Abena to total 3,500 dogs. This was confirmed through a household survey conducted after the vaccination campaign, which estimated the dog/human ratio to be 1/20. The proportion of ownerless dogs was between 8% and 15% [5]. The total human population in Chagoua and Abena was 72,000 people.
We developed a spatially explicit network construction algorithm to expand the empirical contact network to a synthetic network with more nodes, which allows for more realistic simulations of rabies transmission. When applied to a set of nodes of the same size as the empirical network, this algorithm generated a network with a similar degree distribution. The outbreak probability and size of a rabies transmission model on the empirical and the synthetic network were similar, meaning we captured the features of the empirical network which are relevant for disease transmission in the construction algorithm.
The steps of the algorithm to create the synthetic network are described below. We first create a graph with n nodes and zero edges. The number of nodes n corresponds to the number of nodes in the empirical network. Each node is assigned a position consisting of x and y coordinates in a square. The coordinates are sampled using Latin Hypercube sampling. Any two nodes i and j are connected with a probability pij given by pij = exp(−κΔij), where Δij is the Euclidean distance between node i and node j and κ is a scaling parameter. Next a proportion 1 − τ of the nodes are selected uniformly at random. For each node i in that subset of nodes a number m is sampled from a Poisson distribution with mean λ. The node i is then connected to exactly m other nodes out of all the nodes in the graph. The probability of selecting node j into the m nodes is given by p ˜ i j = k j ∑ l = 1 n k l, where kj is the degree of node j and ∑ l = 1 n k l is the sum of the degrees of all the nodes in the graph.
The three scaling parameters, κ, τ and λ are chosen such that the Kolmogorov distance between the degree distribution of the synthetic network and the degree distribution of the empirical network is minimal. We minimize the Kolmogorov distance by using a gridsearch and confirm the results by minimizing a second metric, the χ2 distance. The optimal values of the parameters κ, τ and λ for the two study zones are displayed in Table 2. Larger networks are constructed by choosing the desired number of nodes in the networks and following the steps described above with the optimal values for λ, τ and κ. The properties of the empirical and the synthetic networks are displayed in Table 3. When optimizing the parameters κ, τ and λ only the degree distribution of the two networks is taken into account. Therefore, other network properties such as clustering do not necessarily align between the synthetic and the empirical network.
We used an individual based transmission model to simulate the spread of rabies in a contact network. All nodes of the network are assigned a status; susceptible, exposed, infective or removed. Nodes infect adjacent nodes with a transmission rate β and progress from exposed to infectious and from infectious to removed with average transition periods σ and δ. For each infected dog the individual incubation period and infectious period is sampled from a Poisson distribution, with mean σ or δ, respectively. The model ignores birth and natural mortality. The parameter values are displayed in Table 4. The incubation period, σ, is chosen from recent literature [37] and fits with the observed time between cases in the incidence data from Chagoua and Abena. The duration of the incubation period is only marginally relevant for our simulations, because it only affects the outbreak duration and not the outbreak probability or size. The infectious period, δ, is chosen based on the assumption that a rabid dog in an urban setting would be killed earlier than a natural death from rabies. Our observation that more than two thirds of all samples tested at the rabies laboratory are positive supports the hypothesis that people are likely to recognize the symptoms of rabies since they are less likely to send non-rabid dogs for testing. If people recognise rabies they are more likely to kill rabid dogs. [4]. The transmission rate is chosen using Eq (1). We calculated the mean and variance of the empirical degree distribution, choosing the transmission rate such that R0 is smaller or equal than 1. We reasoned that rabies is endemic in N’Djaména, with a constant low number of cases and no large outbreaks observed. The transmission rate choice is further supported by the comparison of the simulation results to the outbreak data from Chagoua and Abena.
We used an individual based transmission model to test whether the properties of the empirical and the reconstructed network lead to similar outbreak probability and size for different transmission rates. The results for 1000 simulation runs of this model on the empirical and the synthetic network are shown in Fig 2. We differentiate between minor outbreaks, which are outbreaks where more than one and less than one percent of the nodes gets infected, and major outbreaks, which are outbreaks where more than one percent of the nodes get infected. Incursions denote all outbreaks where more than one node gets infected and therefore include both minor and major outbreaks. The figure suggests the construction algorithm performs well since the empirical and the simulated network yield similar results in outbreak probability and size. The values of the proportion of simulation runs with outbreaks correspond to the values of the average relative outbreak size, that is the sum of all the final outbreak sizes divided by the number of nodes in the network and the number of simulation runs. This is consistent with the theoretical result that the probability of a major outbreak and the relative size of such a major outbreak are equal [38]. This holds despite the clustering of the synthetic network being higher than in a random graph due to the spatial component of the network construction algorithm. In Fig 2 the outbreak size increases steeply for transmission rate values that are slightly larger than 0.02. This is consistent with the basic reproductive ratio R0 given by
R 0 = p ( μ + var ( D ) - μ μ ) , (1)
where p is the transmission probability given a contact and μ and var(D) are the expected value and the variance of the degree distribution [38]. In the case of the described network R0 takes the value of 1 if the transmission rate β is 0.02. Since major outbreaks are only possible when R0 is greater than one, the observed increase of the average outbreak size for values of the transmission probability greater than 0.02 aligns well with the theoretical result, even though not all conditions are met in the case of the described networks.
In study zone 1, the network consisted of 237 nodes and 1739 edges, with an average degree of 15 and and maximal degree of 64. In zone 2, the network consisted of 66 nodes and 272 edges, with an average degree of 9 and a maximum degree of 20. In both zones, nearly all dogs were part of one connected component, that is a sub-graph where any two nodes are connected by a path. The network can be divided into communities using a modularity optimization algorithm [39]. This algorithm optimizes both, the number of communities and the assignment of each node to a specific community, such that the modularity, that is the density of links within communities compared to links between communities, takes the maximum possible value. When the network of study zone 1 is divided into communities using this algorithm it becomes visually obvious that communities mainly consist of dogs which live close together and do not frequently crossrange across roads with traffic (Fig 3). This suggests, that roads with high traffic intensity constitute a functional barrier which substantially reduces contact between dogs residing on either side.
Rabies was absent from the Chagoua and Abena quarters of N’Djaména for more than a year prior to the outbreak in 2016. The 7 cases were the first to occur north of the Chari River. Chagoua and Abena are virtually separated from other quarters to the west, north and east by main traffic roads and to the south by the Chari River. The area of these two quarters is approximately 4km2, and the total number of dogs is estimated to be around 3,500. We simulate the course of the infection after the incursion of one rabid dog. We found that in 450 out of 1000 simulations the chain of transmission was longer than 1, in other words additional dogs get infected. Among these chains of transmission the median of the cumulative incidence of all simulation runs aligns well with the cases observed in Chagoua and Abena (Fig 4). This suggests that the transmission rate in our model is a reasonable choice and that our simulations yield realistic results. Since rabies is often underreported, the true number of cases is likely to be higher than the reported number of cases. We accounted for this in a sensitivity analysis on the reporting probability (S3 Fig). We found that if more than 60% of the cases are reported, the median of the simulations does not differ more from the incidence data than with perfect reporting. The final outbreak sizes are shown in S4 Fig.
To assess the impact of vaccination coverage on the outbreak probability and size after the introduction of one rabid dog, we constructed a network with a large number of nodes. We considered a 4 × 4 kilometer square and a dog population with the same density as the dog population in study zone 1, which yields a network with 4930 nodes. We ran rabies incursion simulations on that network. The outbreak probabilitiy, size and duration across different vaccination coverage are shown in Fig 5. The probability of a major outbreak, defined as more than 1% of the dog population becoming infected, is substantially reduced when vaccination coverage is above 70%. The probability of minor outbreaks also decreases with vaccination coverage, but only reaches zero with nearly complete vaccination coverage. Even though a minor outbreak, by definition, could affect up to 1% of the population (50 dogs) the simulated average outbreak size is, in fact, very low. This is consistent with the theoretical result that the final number of infected nodes converges to a two point distribution. A proportion of simulation runs stays close to zero whereas the other proportion ends up near the major outbreak size (for an example see S5 Fig). The minor outbreaks, therefore, only capture the short chains of transmission. These chains include, on average, 5 dogs and last approximately 20 weeks, yielding an average number of one infected dog per month which aligns well with the observed endemic situation in N’Djaména [4].
After the vaccination campaigns in 2012 and 2013, no rabies cases were reported north of the Chari River until October 2014. We used a deterministic model [3] to estimate vaccination coverage over time and the contact network model to calculate outbreak probability for the respective coverage. Comparing these probabilities with the incidence data (Fig 6) showed that the first case after the vaccination campaigns could not establish a chain of transmission because the probability for a major outbreak was very low at that time. Later, in February 2016, the respective probability was higher which could explain the subsequent cases.
We used the empirical contact network from zone 1 to compare different types of vaccination strategies. Dogs can be vaccinated at random or in a targeted way, based on the contact network structure among the dogs or based on the movements of the dogs. We considered four different ways of targeting dogs: (i) vaccination in order of the degree centrality of the nodes, (ii) vaccination in order of the betweenness centrality of the nodes, (iii) vaccinating each node with a probability that is linearly proportional to the average distance the corresponding dog spent away from the home location of the owner and (iv) vaccinating each node with a probability that is linearly proportional to the area covered by the corresponding dog, where the area was estimated by fitting a minimal convex polygon to the GPS locations of the dog. The outbreak probability and size for each type of vaccination and different coverages are shown in Fig 7. Consistent with previous findings [40, 41] we observed that targeted vaccination reduces the outbreak probability and size more than random vaccination. Targeting nodes by degree yields a lower outbreak probability and size than targeting nodes by betweenness. The betweenness centrality of a node i is the proportion of shortest paths between any pair of nodes in the network that pass through node i. Nodes with high betweenness centrality are therefore part of many short paths between nodes, which is why removing them affects the global network structure and reduces the size of the largest component, while targeting nodes by degree operates on a local level and reduces the total number of edges more rapidly. In our case, chains on average are short, so the local structure is more important than the global structure. Vaccination based on movement also reduces the outbreak probability and sizes.
We conducted a Partial Rank Correlation Coefficient (PRCC) sensitivity analysis [42] to assess the impact of the network construction and transmission model parameters on the model output, with ranges as displayed in S1 Table. The results are shown in Fig 8. The most sensitive parameter is τ, a scaling parameter of the network construction algorithm. For low values of τ, a large proportion of nodes are sampled to connect both to spatially close nodes and any other node in the network. These nodes have a higher degree and betweenness centrality than the other nodes in the network, resulting in an overall larger outbreak size and duration. The remaining two network construction parameters, κ and λ, do not have a large effect on the model output. Among the parameters of the transmission model the infectious period, δ, is most sensitive. Since the model ignores birth and natural mortality, the incubation period σ is only relevant for the outbreak duration and not for the outbreak size. A sensitivity analysis of the outbreak probability, size and duration for different vaccination coverages is shown in S6 and S7 Figs.
This study used empirical contact data to develop a contact network model of dog rabies transmission. We validated the simulation results with 2016 outbreak data from N’Djaména. We used the model to compare the probability of rabies establishment after incursion across different vaccination coverage. We showed that vaccination coverage above 70% prevents major outbreaks, which is consistent with previous findings [3].
In contrast to deterministic models, our individual-based model allowed us to investigate the whole possibility space of outbreak scenarios. Differentiating between minor and major outbreaks revealed that even though the probability of major outbreaks is very low for high vaccination coverage, minor outbreaks can still occur even at nearly complete vaccination coverage. These minor rabies outbreaks consist of approximately 5 dogs, which aligns well with current observations from N’Djaména [4]. The endemicity of rabies in N’Djaména could be explained as a series of rabies introductions with subsequent minor rabies outbreaks, as has been observed in Bangui [43].
We showed that targeting dogs by degree centrality, betweenness centrality or based on their movement substantially increases the impact of vaccination. Targeted vaccination based on betweenness centrality does not perform better than targeted vaccination based on degree centrality. The observation that vaccination by degree performs as well as vaccination according to other network centralities is consistent with previous findings in humans [41]. The degree or betweenness centrality can only be assessed using expensive methods like the tagging with geo-located contact sensors conducted in this study. Such methods cannot be used in routine surveillance. We have shown that vaccination based on movement also reduces the outbreak probabilities and sizes. This might indicate that oral vaccination would be an effective intervention because dogs which cover a lot of territory would be more likely to encounter oral vaccine baits. Oral vaccination has been shown to effectively prevent rabies in dogs [44] and is currently recommended by the WHO as a complementary measure to increase coverage in mass vaccination campaigns [45]. Oral vaccination must be carefully planned with regard to biosafety, for example by assuring that vaccinators retrieve unconsumed baits [46]. It has been successfully implemented to eliminate fox rabies in central Europe [47]. Further consideration of oral vaccination of dogs is warranted based on these results.
We observed a dog population where only a few dogs were not part of the largest component, similar to Hirsch et. al [33], who used proximity loggers to reveal a highly connected population in raccons.
In contrast to the raccoon rabies model of Reynolds et. al. [34], which concluded that with vaccination coverage of 65% the probability of a large outbreak remains around 60–80%, we noted a substantial drop in the probability of a major outbreak. This might be due to the fact that, while raccoons remain infectious until death from rabies, we assumed that dogs remain rabid for only two days on average because we hypothesized that in an urban setting a rabid dog would be killed by the community. Therefore, major rabies outbreaks could be prevented by rabies awareness and locally reactive interventions.
Unlike Dürr et al. [35] who found that even at a vaccination coverage of 70% approximately half the dog population dies from rabies, we found outbreak sizes of less than 1% of the population for high vaccination coverage. This might be due to the fact that Dürr et al. considered reactive vaccination after incursion rather than preventive vaccination.
There are several limitations to our study. Our simulations are based on the assumption that rabid dogs stay infective for two days on average, which does not consider the fact, that rabid dogs can be infectious for several days before they show symptoms. Previous models of rabies in wildlife indicated an effect of seasonality on outbreak sizes and durations. Collecting contact data at different times of the year is currently planned, and subsequent analyses will explore the impact of seasonality on contact rates. Dog contacts were only measured for a period of 3.5 days, the extent of battery life. While this observation window is longer than the average infectious period, we cannot be certain that the structure of the network would remain the same when measured for a longer time. Also, contacts with untagged owned dogs and unowned dogs (approx. 8% to 15% of the dog population) were not recorded. Furthermore, we did not include the change of behavior of a rabid animal. However, Reynolds et. al. [34] found that assuming a combination of paralytic and furious rabies in the population leads to little quantitative change in the outbreak size.
We found that major rabies outbreaks are unlikely when vaccination coverage is above 70%. Our results suggest that the endemicity of rabies in N’Djaména might be explained as a series of importations with subsequent minor outbreaks. Further investigation of determinants of dog roaming and contact behavior could inform potential targeted vaccination strategies.
|
10.1371/journal.pntd.0005762 | Molecular and MALDI-TOF identification of ticks and tick-associated bacteria in Mali | Ticks are considered the second vector of human and animal diseases after mosquitoes. Therefore, identification of ticks and associated pathogens is an important step in the management of these vectors. In recent years, Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) has been reported as a promising method for the identification of arthropods including ticks. The objective of this study was to improve the conditions for the preparation of tick samples for their identification by MALDI-TOF MS from field-collected ethanol-stored Malian samples and to evaluate the capacity of this technology to distinguish infected and uninfected ticks. A total of 1,333 ticks were collected from mammals in three distinct sites from Mali. Morphological identification allowed classification of ticks into 6 species including Amblyomma variegatum, Hyalomma truncatum, Hyalomma marginatum rufipes, Rhipicephalus (Boophilus) microplus, Rhipicephalus evertsi evertsi and Rhipicephalus sanguineus sl. Among those, 471 ticks were randomly selected for molecular and proteomic analyses. Tick legs submitted to MALDI-TOF MS revealed a concordant morpho/molecular identification of 99.6%. The inclusion in our MALDI-TOF MS arthropod database of MS reference spectra from ethanol-preserved tick leg specimens was required to obtain reliable identification. When tested by molecular tools, 76.6%, 37.6%, 20.8% and 1.1% of the specimens tested were positive for Rickettsia spp., Coxiella burnetii, Anaplasmataceae and Borrelia spp., respectively. These results support the fact that MALDI-TOF is a reliable tool for the identification of ticks conserved in alcohol and enhances knowledge about the diversity of tick species and pathogens transmitted by ticks circulating in Mali.
| Ticks are among the most important vectors and reservoirs of several animal and human pathogens such as viruses, bacteria and protozoa. However, very few studies have been done on ticks in Mali. At present, little information is available about tick species infesting livestock or human tick-borne diseases transmitted in Mali. The identification of tick species and the determination of pathogens associated are essential to evaluate epidemiology and risks of human and animal diseases: the One Health approach. Current identification methods are time consuming, expensive and laborious. Previous studies have shown that MALDI-TOF mass spectrometry analyses may allow accurate tick species identification. A recent study suggested that it was possible to identify ticks preserved in alcohol by MALDI-TOF MS. The aim of the present study was to improve tick leg sample preparation conditions for their identification by MALDI-TOF MS from Malian ethanol-preserved specimens collected in the field. This study provided 99.4% concordance between morphological and MALDI-TOF identification. The detection of microorganisms was also performed by molecular biology revealing the presence of the presence of Rickettsia spp., Coxiella burnetii, Borrelia spp. and Anaplasmataceae. These results support the use of MALDI-TOF MS in entomology, tick diseases epidemiology and improve the knowledge of tick species-diversity and tick-borne pathogens circulating in Mali.
| Ticks are bloodsucking arthropods that parasitize most of the vertebrates in the world and occasionally bite humans [1]. About 900 tick species have been identified and classified worldwide [2]. In Africa, the number of tick species indexed is 223, including 180 hard and 43 soft ticks [2]. Currently, ticks are considered the second most important vector of human disease after mosquitoes and can transmit bacterial [1], viral [3] and protozoan pathogens [4]. A significant number of these pathogens are of exceptional importance, as they are responsible for high morbidity and mortality in humans and animals [1]. Identification of tick species is an important step in epidemiological studies, in order to establish tick species distribution maps and to characterize tick fauna and seasonal trends [5,6].
In Mali, a West African country, livestock farming is an essential economical factor. At present, there are few studies on tick species that infest cattle or tick-borne diseases transmitted in Mali. To date, 23 tick species belonging to six genera have been categorized in Mali [7–9]. Among them, Amblyomma (Am.) variegatum, Rhipicephalus (Rh.) spp. and Hyalomma (Hy.) spp. are the main ticks monitored by Malian veterinarians for their effects on livestock healthcare and productivity [10]. Other public health problems, such as tuberculosis, AIDS or malaria, take precedence over tick-borne diseases (TBDs), which are little explored by medical doctors.
Several bacteria were detected in ticks from Mali. Spotted fever group rickettsiae were detected, including Rickettsia africae in Am. variegatum, R. aeschlimannii in Hy. marginatum rufipes, and R. massiliae in Rhipicephalus spp., all three being human pathogens [11]. An Ehrlichia sp. of unknown pathogenicity, Ehrlichia Erm58, was detected in Rh. mushamae [11]. More recently, Borrelia theileri, the agent of bovine and equine borreliosis, and B. crocidurae, agents of relapsing fever in humans, have been detected in Rh. geigyi and Ornithodoros sonrai, respectively [12–14].
To study and control ticks and TBD transmission, accurate identification of tick species and determination of their infectious status are essential [1]. Currently, tick identification is principally conducted by observing morphological characteristics. However, it is limited by entomological expertise, dichotomous keys availability, tick integrity or engorged status [9]. Molecular tools have been used as an alternative to overcome the limitations of morphological identification [15]. Sequencing of several genes has been used, including ribosomal sub-units (e.g., 12S, 16S or 18S), the cytochrome c oxidase unit I (COI), or the internal transcribed spacer [16]. These techniques are generally time-consuming, laborious and can be expensive, preventing their use in large scale studies [17–20]. Moreover, the absence of a consensus gene target sequence for tick identification and/or the comprehensiveness of genomic databases are additional factors hampering their use [16].
Recently an alternative tool based on the analysis of protein profiles resulting from matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) analysis has been explored to identify arthropods [21]. MALDI-TOF MS has been used to identify tick species [22–24] and to determine tick infectious status [25–27].
However, tick collection usually takes place far from analytical laboratories and therefore requires proper storage of samples. Ticks are generally stored either alive, at -20°C, or in alcohol. Although alcohol storage is cheaper and easier, especially in African countries, previous studies reported that the use of fresh (i.e., recently dead) or frozen specimens led to more reproducible and better MS spectra compared to the alcohol preservation mode for ticks [24] [28], and also for other arthropod families[29,30]. In a recent study, it was demonstrated that long-term tick storage in alcohol altered MS profiles, which did not provide conclusive identification following in-house MS reference spectra database-querying containing MS spectra from counterpart fresh tick species. Nevertheless, the upgrading of the in-house MS reference spectra database of specimens stored in alcohol allowed correct identification of ticks at the species level, also underlining the reproducibility and specificity of MS profiles for tick specimens stored in alcohol [31].
The goal of the present work was to determine tick population diversity and associated pathogens from alcohol stored specimens collected on cattle in Mali by using MALDI-TOF MS and molecular approaches with specimens collected in the field. First, optimized sample preparation conditions for ticks stored in ethanol for MALDI-TOF MS analysis were established. Second, based on morphological and molecular identification of ticks, an MS reference spectra database was created and tested blindly using new tick specimens. In addition, tick-associated bacterial pathogens were screened by molecular biology on half-tick body parts and leg MS spectra from ticks mono-infected or not by bacterial pathogenic agents, and they were compared to assess the efficiency of this proteomic tool for classification of ticks according to their infectious status.
Tick collection protocols were developed as of a large study under the GIRAFE programme, UMI 3189 and MSHP-MRTC HFV project. The protocols were cleared by the FMPOS IRB in 2015 and 2016. Verbal informed consent was obtained from managers of the livestock selected for tick sampling directly on mammals. The collection of ticks on domestic animals did not involve national parks or other protected areas or endangered or protected species.
Ticks were collected from three localities in Mali, including Bamako, Kollé and Bougoula Hameau, in September 2015 and August 2016 (Fig 1). Bamako, the capital city of Mali, is an urban area surrounded by hills. The climate is Sahelian-type with two distinct seasons, the dry season (i.e., from November to May) and the rainy season (i.e., from June to October). The total amount of precipitation was less than 900 milliliters in 2009. Kollé is a rural village located about 60 km southwest of the capital. Agriculture, livestock farming and small businesses are the main economic activities of the village. The village, located on a flat land with submersible and dry areas, presents a Sahelian-type climate with two distinct seasons, a rainy (i.e., from June to November with maximum rainfall in August-September of 350 to 400 milliliters) and a dry season (i.e., from December to May with a cool period in December- February and a warm period in March-May). The third site was Bougoula Hameau, a suburban village, located at 4 km of Sikasso town and it was situated at 374 km southeast of Bamako by road. The climate is of Sudanese type, under the influence of the humid forest with a rainy (i.e., from May to October) and a dry season (i.e., from November to April). The annual rainfall can vary from 1,200 to 1,800 milliliters, depending on the year. These climatic conditions are appropriate for agricultural and livestock farming.
Ticks were collected from domestic animals and cattle. Examination of all body parts was conducted from the tail to the head of the animal to detect ticks on the skin. All ticks (engorged and non-engorged) were collected manually with forceps. The ticks of the same animal were counted, pooled in the same tube and stored at room temperature in 70% v/v ethanol (ticks collected in September 2015) or frozen at -20°C (ticks collected in August 2016) until morphological, molecular and MALDI-TOF MS analyses. Ticks were transferred from MRTC (Bamako, Mali) to the URMITE laboratories (Marseille, France) for analysis.
Ticks were identified morphologically to the species level firstly by a PhD student and then checked by expert tick entomologists using previously established taxonomic identification keys [9]. Tick identification and gender determination were performed under microscope at a magnification of ×56 (Zeiss Axio Zoom.V16, Zeiss, Marly le Roi, France). The tick genera, species, gender, host and animal number, collection site and date were codified to include this information on the tube.
Each tick was dissected with a new sterile surgical blade to remove the legs, which were used for MALDI-TOF MS analyses. The rest of the tick was longitudinally cut in two equal parts. The half part with legs cut off was immediately used for molecular biology, and the second half was stored frozen as a backup sample for any additional analysis.
Each half-tick without legs was transferred to a 1.5 mL tube containing 180 μL of G2 lysis buffer and 20 μL proteinase K (Qiagen, Hilden, Germany), and incubated at 56°C overnight. DNA extraction from the half-tick was performed with an EZ1 DNA Tissue Kit (Qiagen) according to manufacturer recommendations. The DNA from each sample was eluted with 100 μL of Tris-EDTA (TE) buffer (Qiagen) and was either immediately used or stored at -20°C until use.
Standard PCR, using an automated DNA thermal cycler amplifying a 405-base pair fragment of the mitochondrial 12S RNA gene (Table 1), was used for tick identification to the species level, as described previously [31]. The 16S RNA gene was used to confirm all Rhipicepalus (Boophilus) microplus identification. DNA from Am. variegatum specimens reared at the laboratory was used as positive control. PCR products of the positive samples were purified and sequenced as described previously [31]. The sequences were assembled and analyzed using the ChromasPro software (version 1.34) (Technelysium Pty. Ltd., Tewantin, Australia), and were then blasted against GenBank (http://blast.ncbi.nlm.nih.gov).
Quantitative PCR was performed according to the manufacturer's protocol using a PCR detection system; a CFX Connect™ Real-Time (Bio-Rad) with the Eurogentec Takyon qPCR kit (Takyon, Eurogentec, Belgium). The qPCR reaction contained 10 μl of Takyon Master Mix (Takyon, Eurogentec, Belgium), 3.5 μl sterile distilled water, 0.5 μl of each of the primers and probe and 5 μl of the DNA extract. A total of 471 samples were screened using primers and probes, targeting specific sequences of the following bacterial pathogens: Rickettsia spp., Anaplasmataceae spp., Borrelia spp., Bartonella spp. and Coxiella burnetii (Table 1). For Borrelia spp we used 2 genes, the 16S Borrelia gene first and all the ticks that were positive for this gene were retested by ITS4 for confirmation. Only samples positive for both genes (16S borrelia and ITS4) were considered positive. Positive samples for Rickettsia spp. were then submitted to a qPCR system specific for detecting R. africae [32]. Negative samples for R. africae but positive for Rickettsia spp. were submitted to gltA gene sequencing to determine Rickettsia species [33]. All ticks positive either for Anaplasmataceae spp. were submitted to amplification using standard PCR and sequencing to identify the bacteria species [34,35]. Ticks that were positive for Borrelia spp for both the 16S Borrelia gene and ITS4 were submitted to amplification using standard PCR and sequencing [33]. PCR tests were considered positive when the cycle threshold (Ct) was lower than 36 [36]. The DNA from Rickettsia montanensis, Bartonella elizabethae, Anaplasma phagocytophilum, Coxiella burnetii and Borrelia crocidurae was used as positive controls and mix as negative controls in PCR, respectively. All these bacteria come from the strains of culture of our laboratory and Borrelia crocidurae was cultured in Barbour-Stoenner-Kelly (BSK-H) liquid medium supplemented with rabbit serum. Only samples considered as negative (i.e., Ct ≥ 36 for all bacteria tested), were submitted to 12 S tick gene amplification to control the correctness of DNA extraction.
The homogenized tick legs were centrifuged at 2000 g for 30 seconds and 1 μL of the supernatant from each sample was carefully dropped onto the MALDI-TOF target plate as previously described [28]. Each spot was then recovered with 1 μL of CHCA matrix solution composed of saturated α-cyano-4-hydroxycynnamic acid (Sigma, Lyon. France), 50% acetonitrile (v/v), 2.5% trifluoroacetic acid (v/v) (Aldrich, Dorset, UK) and HPLC-grade water [16]. The target plate, after drying for several minutes at room temperature, was introduced into the Microflex LT MALDI-TOF Mass Spectrometer device (Bruker Daltonics, Germany) for analysis. The loading of the MS target plate, the matrix quality, and the performance of the MALDI-TOF were performed as previously described [28].
Protein mass profiles were obtained using a Microflex LT MALDI-TOF mass spectrometer (Bruker Daltonics, Germany) using parameters previously described [31]. The spectrum profiles obtained were visualized with Flex analysis v.3.3 software and exported to ClinProTools software v.2.2 and MALDI-Biotyper v.3.0. (Bruker Daltonics, Germany) for analysis [25].
The reproducibility of spectra was evaluated by analyzing ten Am. variegatum specimens from Kollé per sample preparation protocol as previously described [29]. The selected protocol was analyzed using an unsupervised statistical test classifying specimens according to MS spectra (i.e., Principal Component Analysis, PCA test, ClinProTools v2.2 software). The Composite Correlation Index (CCI) tool (MALDI-Biotyper v3.0. software, Bruker Daltonics), was used to assess spectra variations within each sample group according to protocol tested, as previously described [37]. CCI was computed using the standard settings of mass range 3000–12000 Da, resolution 4, 8 intervals and autocorrelation off. Higher the log score values (LSVs) and correlation values (expressed by mean ± standard deviation [SD]) reflect higher reproducibility of MS spectra and were used to determine the best protocol for sample preparation.
Based on the correlation of morphological and molecular results of tick identification, two to five specimens per species were used to assess MS spectra reproducibility from specimens of the same tick species, and the MS spectra specificity from specimens of distinct tick species. These analyses were performed with the average spectral profiles (MSP, Main Spectrum Profile) obtained from the four spots of each individual tick specimen using Flex analysis v.3.3 and ClinProTools 2.2 softwares (Bruker Daltonics). Tick species exhibiting reproducible and specific MS spectra were then included in-house MS spectra reference database. To upgrade the database, MSP reference spectra were created using spectra from at least 2 specimens per species of both genders by the automated function of the MALDI-Biotyper software v3.0. (Bruker Daltonics). MS spectra were created with an unbiased algorithm using information on the peak position, intensity and frequency [38]. The spectra files are available on request and transferable to any Bruker MALDI-TOF device.
A blind test was performed with new tick specimens collected in Mali stored in 70% alcohol or frozen. A total of 451 MS spectra from tick legs, including 340 stored in alcohol and 111 frozen specimens were tested successively against the in-house MS reference spectra database (Database 1) and its upgraded version, which includes the 20 MS spectra from specimens of the 6 tick species collected in Mali and alcohol-preserved (Database 2). Among the 451 ticks tested 51 Am. variegatum and 23 Rh (B) microplus were fully engorged. Database 1 was composed of specimens of fresh or frozen arthropods (Table 2) [16,24,30,31,39]. Database 2 includes database 1 plus MS spectra of tick legs from 6 species stored in ethanol from the present study (Tables 3 and 4). The reliability of species identification was estimated using the LSVs obtained from the MALDI-Biotyper software v.3.3, which ranged from 0 to 3. These LSVs correspond to the degree of similarity between the MS reference spectra database and those submitted by blind tests. An LSV was obtained for each spectrum of the samples tested. Moreover, to decipher incoherent results obtained between morphological and MS identification, molecular identification of ticks was performed for the respective specimens.
These comparative analyses to determine the infectious status of ticks were made by ClinProTools v.2.2 software (Bruker Daltonics, Germany). Only tick leg MS spectra from species with at least five mono-infected or pathogen-free specimens were included in this analysis. The spectra of 30 specimens of A. variegatum infected by R. africae were compared to those of 12 uninfected specimens from the same species. Moreover, MS spectra of 36 uninfected specimens of Hy. truncatum were also compared with the spectra of 23 specimens of Hy. truncatum infected by C. burnetii.
A total of 1,333 ticks were collected from the three sites including 406 engorged (Fig 1). A total of 1,217 were found on 44 bovine specimens and 116 on 9 dogs. Nineteen engorged females of the Hyalomma genus (1.55% of ticks collected) were not morphologically identified to the species level. Morphologically, six distinct tick species belonging to three genera were identified among ticks collected in September 2015 (Table 3). Am. variegatum (n = 877, 71.79%) was the overall predominant tick species collected from different sites, followed by two species of the Hyalomma genus, Hy. truncatum (n = 260, 21.27%) and Hy. m. rufipes (n = 28, 2.29%). The three other tick species, Rh. (Bo.) microplus, Rh. e. evertsi and Rh. sanguineus sensus lato, represented less than 3.10% (n = 38). The five Rh. sanguineus sl [40] specimens were all collected on a dog. All 111 ticks collected in August 2016 were identified as Rh. sanguineus sl. Three hundred sixteen of the 1,222 ticks collected from three sites in 2015 and 111 ticks in 2016 had specimens of six species randomly selected for molecular and proteomic analyses (Table 3).
A total of 20 specimens, including 2 to 5 specimens per species and all specimens of Rh (Bo.) microplus, were randomly selected for molecular analysis. A GenBank query revealed that 12S gene sequences were available for the 6 tick species. BLAST analysis indicated high identity (i.e., a range from 99% to 100%) of 12S rRNA gene sequences among specimens classified per species according to morphological identification (Table 4). BLAST analysis revealed that these 6 tick species and all specimen of Rh (B) microplus had high sequence identity with their respective homolog species available in GenBank (i.e., range 96.5% to 100%; Table 4).
Among the ticks tested, 41.8% (197/471) were negative for the six bacteria tested, 37.4% (176/471) were positive for one bacterium and 20.8% (98/471) were found co-infected by two or three of the screened bacteria. Among the 274 specimens found positive for at least one bacteria tested, 76.6% (210/274) were infected by Rickettsia spp., among which R. africae was found in 87.6% (184/210) (Table 4). The amplification of the ompA fragment in the remaining ticks positive for Rickettsia spp. and negative for R. africae (n = 26) was used for identification of these Rickettsia spp. R. aeschlimannii and R. mongolitimonae were detected in 24 and 2 tick specimens, respectively (Table 5). Screening of all ticks for Coxiella burnetii revealed that 37.6% (103/274) of the specimens were positive (Table 4). Fifty-seven ticks, 20.8% (57/274) were positive in qPCR targeting the 23S rRNA of Anaplasmataceae. Among them, 23S rRNA amplification and sequencing was successful for 50 samples. The BLAST found broad agreement that 43 ticks were positive for E. ruminantium (GenBank accession number NR 077002.1), 2 ticks were positive for Ehrlichia sp. urmitei TCI148 (GenBank ACCN KT 364334.1) and 1 tick for Ehrlichia sp. rustica TCI141 (GenBank ACCN KT 364330.1). A. marginale was detected in 3 ticks and A. sp. ivoriensis TCI50 (GenBank ACCN KT 364336.1) in 1 specimen (Table 5). Borrelia spp. was detected in 1.1% (3/274) of ticks by qPCR. However, all standard PCR for determination of Borrelia species failed. No Bartonella spp. was detected in the ticks tested.
A comparison of our current reference sample preparation method (i.e., “de-alcoholization”) with the “dry” and “direct” methods was performed [31]. The best method was selected on the following criteria: reproducibility and intensity of MS spectra, low handling and simplicity of the protocol. To exclude inter-individual variability, protocols were successively compared by pairs, and then the four right legs were used for one protocol and the four left legs from the same tick for the other. Then, ten specimens of both genders tested per protocol, five males and five females, were included. For all these experiments, morphologically identified ticks from Kollé (Am. variegatum) were used. The first comparison concerned the “de-alcoholization” and “dry” protocols (Fig 2A). The visual comparison of MS profiles between these two groups using the gel view tool and the superimposition of average MS profiles in each condition using ClinProTools software (Bruker) did not reveal differences in peak position between the two protocols (Fig 3A and 3B). This reproducibility of the profiles was analyzed using an unsupervised statistical test classifying specimens according to MS spectra (i.e., Principal Component Analysis, PCA test, ClinProTools software). The mixing of both groups on the graphical representation confirmed the absence of differences between both groups (Fig 3C). Thus, the “dry” protocol was preferred compared to the “de-alcoholization” protocol, the latter considered to be more time-consuming and fastidious. The second comparison concerned the “dry” and “direct” protocols, using ten Am. variegatum specimens from both genders (Fig 2B). The comparison of MS profiles between these two groups, either by gel view, superimposition or PCA (Fig 4A, 4B and 4C), could not determine the more relevant method. The Composite Correlation Index (CCI) tool revealed a higher CCI (LSV mean±SD: 0.783±0.101; Fig 4D) for the “dry” protocol compared to “direct” (LSV mean±SD: 0.755±0.175; Fig 4D). These results were in agreement with the gel view showing a higher visual homogeneity of the MS spectra from the “dry” group. Finally, the “dry” protocol appeared consistently to be the more reproducible and low-handling procedure for the preparation of ethanol-stored ticks for MS analysis, and was chosen for the next experiments of the present study.
Twenty ticks, including several specimens per species coming from distinct localities, were identified by sequencing 12S tick gene. Their non-infected status was also controlled for the microorganisms tested in the present work by q PCR. These specimens were selected for evaluating intra-species reproducibility and inter-species specificity of MS spectra. Comparison of the MS spectra with Flex analysis software indicated reproducibility of the MS profiles between tick specimens from the same species (Fig 5A). Moreover, the visual comparison of MS profiles indicated a clear distinction of spectra according to species. To reinforce the specificity of MS profiles according to tick species, MS profiles from these 20 specimens were used to generate a dendrogram and PCA (Fig 5B and 5C). Clustering analysis revealed a gathering on distinct branches of ticks according to species. However, at the genus level, all specimens from the Rhipicephalus genus were not clustered in the same part of the dendrogram. The profile of the spectra of specimens preserved in alcohol was different from those of fresh specimens of the same species; this difference was also observed between manual sample homogenization and automated sample crushing using the TissueLyser apparatus.
To assess the efficacy of the in-house MS reference spectra database, named database 1 (DB 1) to correctly identify tick specimens preserved in alcohol, half of the MS spectra from ticks included in the present study were randomly selected. Then, MS spectra from 178 specimens including 60 Am. variegatum, 64 Hy. truncatum, 16 Hy. m. rufipes, 26 Rh. (Bo.) microplus, 7 Rh. e. evertsi and 5 Rh. sanguineus sl were queried against the DB 1 spectra database. The blind test against DB 1 revealed correct identification for some specimens of Hy. truncatum (n = 5) and Hy. m. rufipes (n = 5), with LSVs > 1.8 (Table 6). For the remaining ticks (n = 168), all LSVs were ˂ 1.8 [24]. Tick MS spectra from 20 specimens, including 6 species identified morphologically and molecularly in this work, were added to DB 1, which was then renamed DB 2 (Table 4). Thereafter, the leg spectra of the 451 morphologically-identified ticks, including 340 stored in alcohol and 111 frozen, were queried against DB 2. Among the 451 ticks tested 51 Am. variegatum and 23 Rh (Bo.) microplus were fully engorged.
The results of this second interrogation (blind test 2, BT2) showed 96.7% (325/340) concordance between morphological identification and MALDI-TOF MS identification. The percentage of concordant identification with morphology was 100% for Rh (Bo) microplus, Rh. e. evertsi and Rh. sanguineus sl stored in alcohol, with LSVs ranging from 1.89 to 2.71 (Table 6). A total of 15 specimens presented divergent identification between morphological and MALDI-TOF MS identification. To eliminate any doubt, these 15 specimens were submitted to molecular identification. Sequencing of the 12S gene confirmed the identification obtained by MALDI-TOF MS for 13 specimens (Table 6). The remaining 2 specimens identified as Hy. m. rufipes by MALDI-TOF MS were finally classified as Hy. truncatum by molecular biology, confirming morphological identification. All fully engorged ticks were correctly identified by MALDI-TOF MS. The percentage of correct MALDI-TOF MS identification for all species was 99.6% (449/451) (Table 6).
The comparison of MS profiles between 30 Am. variegatum uninfected and 12 infected by R. africae using the gel view tool and Principal Component Analysis by ClinProTools software (Bruker), revealed no differences between the two groups (S1 Fig). The same observation was made by comparing of MS profiles of 36 Hy. truncatum uninfected and 23 Hy. truncatum infected by C. burnetii (S1C and S1D Fig).
MALDI-TOF MS has revolutionized clinical microbiology by its use in the routine identification of bacteria [41,42] and archaea [43]. Even if the MALDI-TOF MS device acquisition could be expensive, its use for entomological analyzes induces low additional costs because reagents used for this high-throughput technique are economical and data analyses are simple and rapid compared to morphological and molecular methods [44]. This fast, economical and accurate proteomic tool has since been applied to the identification of arthropods: culicoides biting midges [45], mosquitoes[39,46,47], phlebotomine sand flies [48,49], fleas [30] and tsetse flies [50,51]. MALDI-TOF MS has also been proposed for identifying tick species which are laboratory-reared, collected in the field or on mammalian hosts, by analyzing whole specimens [22] or legs only [23,24]. More recently, preliminary studies have investigated the capacity of MALDI-TOF MS to differentiate ticks infected or not by Borrelia spp. or spotted fever group rickettsiae [25–27], and to detect the Plasmodium in anopheles [44]. However, tick collection is usually far from the analytical laboratories, requiring proper storage of samples. Although the alcohol storage mode is cheaper and easier, especially in African countries, previous studies reported that the use of fresh (i.e., recently dead) or frozen specimens led to more reproducible and better MS spectra, compared to the alcohol storage mode for ticks [24,28], and also for other arthropod families [29,30]. Recently, the application of MALDI-TOF MS for identification of ticks collected in the field in East Africa and preserved in alcohol has allowed reliable identification [23]. More recently, the discriminatory power of MALDI MS-TOF for the correct identification of ixodid tick specimens collected in the field in Ethiopia, which were preserved in 70% ethanol for about two years, was reported [31].
In this study, the morphological identification of ticks revealed the presence of six species, including Am. variegatum, Hy. truncatum, Hy. m. rufipes, Rh. (Bo) microplus and Rh. e. evertsi that were collected from cattle and Rh. sanguineus sl from dogs. Rh. e. evertsi was found only in Bamako, while all other species of ticks were found on cattle in the three locations. In support of these morphological identification results, several studies have reported the presence of these tick species in Mali, except for Rh. (Bo) microplus [7–9]. Rh. (Bo) microplus, which is a southeast Asian tick, was introduced in the southeast of Africa (South Africa, Zambia, Tanzania and Malawi) by cattle from Madagascar [9]. It was reported in West Africa (Ivory Coast) for the first time in 2007 [52]. The presence of Rh. (Bo) microplus has only been found in three other countries of West Africa (Mali, Benin and Burkina Faso) [53]. Biguezoton et al (2016) and Boka et al (2017) found that Rh (B) microplus represent 70% and 63.2% of ticks in Burkina Faso and Benin and Ivory Coast respectively [54, 55]. Our study confirms the presence of this species in 3 localities in Mali, which could indicate its rapid spread and its probable installation in Mali. As expected, Am. variegatum was the most prevalent species in the three sites of the present study [10].
To confirm the morphological identification of tick specimens that were used for creating the MALDI-TOF MS database, sequencing of the 12S rRNA gene was performed. The 12S rRNA gene was chosen to validate identification because this gene is known as a reliable tool for molecular identification of ixodid ticks [16]. The coverage percentages and identity between the sequences of specimens of the same species were from 99 to 100% for all species of ticks. Percentages of identity and coverage of sequences Am. variegatum, Hy. m. rufipes, Rh. (Bo) microplus, Rh. e. evertsi and Rh. sanguineus sl were 99–100% with sequences of the same species available in GenBank. Interestingly, lower sequence identities (96–97%) of Hy. truncatum compared to the corresponding reference sequence in GenBank were observed. It could be hypothesized that the sequence differences could correspond to genetic variation within ticks of the same species adapted to different geographic regions of a country or countries, as previously described [56]. The difference between the sequences of 12S rRNA genes of Hy. truncatum collected in Mali and that available on GenBank tick collected in Zimbabwe [57] could explain these genetic variations. In the future, the sequencing of a second gene target, such as 16S or COI, could be performed to further study these variations [58].
In this study, DNA from Rickettsia spp. was detected in 76.6% of infected ticks collected from cattle, among which R. africae was found in 87.6% (184/210). R. africae was detected in 92.2% of Am. variegatum, a cattle tick found throughout sub-Saharan Africa. Such high prevalence of R. africae in Am. variegatum has already been reported [59–61]. R. africae was also detected in Rhipicephalus spp. and Hyalomma spp., respectively 7.9% and 9.2%. Other recent studies have detected R. africae in other tick genera, including Rhipicephalus and Hyalomma [59,62,63]. R. africae is the etiological agent of African tick-bite fever in humans (ATBF) [64]. R. aeschlimannii have been observed in Hyalomma spp., with 9% and 52.6% respectively in Hy. truncatum and Hy. m. rufipes. These data are comparable with those of previous studies that reported 45% to 55% of Hy. m. rufipes and 6% to 7% of Hy. truncatum were DNA carriers of R. aeschlimannii in Senegal [63], and 44% and 11% in Ivory Coast [59]. The sequences of R. aeschlimannii identified in our work were identical to those of R. aeschlimannii, previously detected in Hy. truncatum collected in Senegal (GenBank accession number HM050276.1). R. aeschlimannii is an agent of spotted fever in humans [64]. R. aeschlimannii is found in sub-Saharan Africa, North Africa, Europe and Asia [11,65]. Our results confirm a large prevalence of this pathogen in Mali.
For the first time, the presence of R. mongolitimonae was identified in Hy. truncatum from Mali. It had been previously detected in Hy. truncatum from the countries bordering Mali, including Niger [11] and Senegal [63]. R. mongolitimonae 12S sequence of the present study were 99% identical with the same sequence fragment of a strain previously isolated from a patient from Algeria (GenBank DQ097081.1).
Until now, two Borrelia species have been identified in Mali, B. crocidurae in the soft tick (O. sonrai) and B. theileri in the hard tick (Rh. geigyi) [12,13]. Our results show the presence of Borrelia spp. in 2 specimen of Am. variegatum and 1 of Hy. truncatum by qPCR using 16S Borrelia and ITS4 genes. Similarly, Ehounoud et al. previously reported the presence of Borrelia spp. in the same tick species in Ivory Coast [59]. Unfortunately, no PCR products using standard amplification were obtained for any of these ticks. This failing could be explained by the higher sensitivity of qPCR compared to standard PCR [66].
In the present work, C. burnetii, the agent of Q fever, was detected for the first time in ticks in Mali, with a prevalence of 33.4% in the six tick species identified. These results differ from those of Ehounoud et al. in Ivory Coast, who found only one tick infected with C. burnetii [59]. Q fever is a ubiquitous zoonotic disease caused by C. burnetii. It is poorly documented in Africa. A recent study conducted in febrile African patients found one male adult patient (0.3%) infected with C. burnetii in Algeria and six patients (0.5%) in Senegal [67]. However, in another study conducted in Senegal, C. burnetii was detected in humans as well as in ticks [68].
The Anaplasmataceae bacteria family was previously considered to be pathogens of veterinary importance [59]. However, in recent decades, many agents of this family have been described in humans [69]. Here, we reveal the presence of A. marginale in 11.5% of Rh (Bo.) microplus. This is the first demonstration of the presence of A. marginale, the agent of bovine anaplasmosis [70] in Mali. A. marginale is an intracellular bacterium responsible for bovine anaplasmosis which manifests with anemia and jaundice [64]. Also, E. ruminantium was found in Am. variegatum, Hy. truncatum, Rh (Bo.) microplus, and Rh. e. evertsi. The prevalence of E. ruminantium was 13.9% in ticks. Potential new species of Ehrlichia and Anaplasma (E. sp urmitei TCI148, E. sp rustica TCI141 and A. sp ivoriensis TCI50) have been detected in Rh (Bo.) microplus and Hy. truncatum. These bacteria had already been detected in ticks from Ivory Coast [59].
However, co-infections have been found in the ticks in this study. The percentage of co-infected ticks was 23.1% (109/471), and we describe for the first time multiple co-infections in ticks in Mali. Recently, multiple co-infections in ticks have been reported in Ivory Coast; these co-infections systematically involved R. africae [59]. The percentage of ticks co-infected was higher in our study than that obtained in Ivory Coast [59].
To avoid bias, we choose to query the MS spectra of 178 specimens of ticks, including 6 species against DB 1 which includes several families of arthropods, including mosquitoes. We constantly improve it with new specimens collected in the field and find it more relevant to carry out a total interrogation without the knowledge without any filter on a specific family. The results of the blind test revealed correct identification in 10 specimens only with high log score values, even though this database contained the same tick species that were also preserved in alcohol. This misidentification could be attributed to several factors: (i) the method used for sample crushing (initially manually, and here an automatic apparatus was used as previously described [28], (ii) the difference in storage time (6 months here vs 3 years in the previous study), (iii) the geographical distance (Mali vs. Ethiopia), which could have consequences on MS spectra profiles, as observed also at the genetic level. This last phenomenon had already been reported in other studies of sand flies [71], mosquito immature stages [46] and ticks [31]. Conversely, when database 1 was upgraded with 20 spectra of the six tick species of our study, the blind test of all ticks revealed 95.60% (325/340) correct identification for tick species stored in alcohol. However, the remaining fifteen ticks (4.40%) with inconsistent identification between morphological and MALDI-TOF MS tools were subjected to molecular biology to determine the real identification of these specimens. The molecular biology results confirmed those of MALDI-TOF MS for 13 of these specimens. Two Hy. truncatum specimens were misidentified by MALDI-TOF MS. The reasons for the misidentification of the two specimens remain unknown. Additionally, all ticks frozenly stored were correctly identified by the blind test.
The results of this work show that MALDI-TOF MS is superior to morphological identification, as the correct identification percentage is 99.6% for all tested. It is also interesting because there are fewer entomologists able to identify ticks and the morphological identification keys are not always available. Another advantage of MALDI-TOF is that it can identify ticks that are completely engorged or damaged, for which morphological characteristics can be deformed or even disappear making morphological identification difficult or impossible. Conversely, the proteomic strategy proposed here, does not require specific skill or expertise, reagents are very cheap so the running cost is very low compared to a molecular biology. The current limiting factors of MALDI-TOF MS analysis are the small diversity of tick species included in the MS spectra reference database and the relative elevate cost to acquire the machine. Nevertheless, it high-throughput and large application for microorganisms identification either in research or medical diagnosis, do of this emerging tool a highly competitive method also for medical entomology studies. It is likely that MALDI-TOF MS will realize similar revolution in medical entomology as it was occurred in microbiology.
Our results confirm those of previous studies, according to which MALDI-TOF MS could be used for identification of ticks preserved in alcohol, but it requires the creation of a database with specimens stored in the same condition [31].
In our work, MALDI-TOF MS analysis was not able to differentiate ticks which were infected or not by the bacteria that were screened. However, preliminary studies from our laboratory seemed promising, as MALDI TOF analysis allowed differentiation of ticks infected or not by Borrelia spp. or spotted fever group rickettsiae [26,27]. The failing of bacteria-pathogen detection by MALDI-TOF MS could be attributed to several factors. The storage mode, fresh versus alcohol, might play a role. Moreover, the infectious status of these ticks was controlled against some bacteria pathogens, however, it was possible that they were infected by others pathogens not researched in the present study, which could impaired the determination of specific MS profiles for each associated pathogens. These factors could alter MS spectra profiles between uninfected and infected ticks. More studies are needed to explore the capacities of MALDI TOF to detect tick infectious status.
To conclude, the present work has confirmed that MALDI-TOF MS may represent a rapid and inexpensive alternative tool for accurate identification of ticks collected in the field and stored in alcohol. The recent demonstration of the use of MALDI-TOF MS for identification of ticks and associated pathogens requires further investigation.
|
10.1371/journal.pntd.0002492 | Exploring the Trypanosoma brucei Hsp83 Potential as a Target for Structure Guided Drug Design | Human African trypanosomiasis is a neglected parasitic disease that is fatal if untreated. The current drugs available to eliminate the causative agent Trypanosoma brucei have multiple liabilities, including toxicity, increasing problems due to treatment failure and limited efficacy. There are two approaches to discover novel antimicrobial drugs - whole-cell screening and target-based discovery. In the latter case, there is a need to identify and validate novel drug targets in Trypanosoma parasites. The heat shock proteins (Hsp), while best known as cancer targets with a number of drug candidates in clinical development, are a family of emerging targets for infectious diseases. In this paper, we report the exploration of T. brucei Hsp83 – a homolog of human Hsp90 – as a drug target using multiple biophysical and biochemical techniques. Our approach included the characterization of the chemical sensitivity of the parasitic chaperone against a library of known Hsp90 inhibitors by means of differential scanning fluorimetry (DSF). Several compounds identified by this screening procedure were further studied using isothermal titration calorimetry (ITC) and X-ray crystallography, as well as tested in parasite growth inhibitions assays. These experiments led us to the identification of a benzamide derivative compound capable of interacting with TbHsp83 more strongly than with its human homologs and structural rationalization of this selectivity. The results highlight the opportunities created by subtle structural differences to develop new series of compounds to selectively target the Trypanosoma brucei chaperone and effectively kill the sleeping sickness parasite.
| Sleeping sickness, or human African trypanosomiasis (HAT), is a deadly neglected disease for which new therapeutic options are badly needed. Current drugs have several liabilities including toxicity and route of administration limiting their efficacy to combat the disease. Our study aimed at validating a potential new drug target against Trypanosoma brucei, its heat shock protein 83 (Hsp83). The chaperone was screened against a repurposed library composed of inhibitors against the human Hsp90. The compounds were assayed in their ability to bind the T. brucei protein and to kill the parasite. Our work has identified selective and high-affinity chemical compounds targeting the parasitic Hsp83. Additionally, structural studies were conducted to explore the observed selectivity of selected inhibitors. Our work has validated T. brucei Hsp83 as a potential target for future drug discovery campaigns. It has also shown the strength of repurposing chemical libraries developed against human proteins, emphasizing the possibility to piggyback current and past drug discovery efforts for other diseases in the search for new drugs against neglected tropical diseases.
| Human African trypanosomiasis (HAT), better known as sleeping sickness is a vector-borne disease present in sub-Saharan Africa, transmitted by Glossina tsetse flies and caused by the protozoan parasite Trypanosoma brucei [1]. Two subspecies of this kinetoplastid parasite cause disease in humans and are present with different geographic distributions, T. b. gambiense and T. b. rhodesiense. The former is responsible for chronic disease in Western and Central Africa, while infection by the latter leads to the acute form present in Eastern and Southern Africa. During the initial stage of the infection, the Trypanosoma parasite lives and multiplies in the blood and tissue fluids of its human host, thanks to an elaborate mechanism for evading the host immune system. The parasite then invades the central nervous system (CNS) to give rise to the fatal stage 2 infection, during which the classic clinical symptoms of HAT occur. Currently, there are five clinically used treatments, which are prescribed based on the causative species and the stage of the disease [2], [3], [4]; however, the toxicity of existing drugs and inappropriate route of administration limit the efficacy of the current chemotherapy. Consequently, HAT is one of the most neglected tropical diseases due to the limited availability of safe and cost-effective control tools [1], [2]. New methods to treat patients are needed to treat and to eventually eliminate the disease. Among the potential new drug targets, molecular chaperones represent an interesting group already validated in other disease areas. Furthermore, a large number of inhibitors are already available [5], several of which have been proved to be effective anti-proliferatives against several parasites in vitro [6], [7], [8], [9], [10].
Among the chaperones, heat shock protein 90 (Hsp90) – alternately referred to as Hsp83 or Hsp86 because of the variable molecular weight amongst different orthologues – is a hallmark of the stress response at the cellular level, and a major component of the eukaryotic proteome [11]. Its main function is to serve as a molecular chaperone helping nascent polypeptides avoid mis-folding and limiting protein aggregation during thermal stress [12]. In eukaryotes, key regulatory proteins require Hsp90 assistance as a part of their maturation process and are known as Hsp90 clients [11]. Not surprisingly, limiting Hsp90 chaperone function either by genetic or chemical means has been found to result in a net reduction of functional client proteins within the cell. This severely impacts cell survival and homeostasis, not to mention its ability to cope with environmental stress [11]. Several of the client proteins in humans are over-expressed in cancer cells. Consequently, development of inhibitors against human Hsp90 (hereafter referred to as Hsp90) has been a highly active field in the anticancer drug therapy field [5], [13], [14]. These compounds from natural and synthetic sources have been found to be effective anti-proliferatives against cancer cells in vitro [15], with over a dozen having advanced into clinical trials [16]. Although the compounds represent diverse chemical scaffolds with unique properties, they all inhibit the enzymatic ATPase function of the chaperone by binding the ATP pocket in the N-terminal domain (NTD), which together with the middle and C-terminal domains define the highly conserved Hsp90 protein architecture [11]. During its catalytic cycle, the Hsp90 serves as a molecular clamp, binding both ATP and client protein at its NTD and middle domains respectively. The transient dimerization of the NTDs is coupled with both the ATP hydrolysis and chaperone activity [17]. While in the resting state, the Hsp90 dimerizes solely via its C-terminal domain. In addition to a multitude of client proteins, the Hsp90 also binds a cohort of co-chaperones that regulate the closing and opening cycle [11], [17].
Like most eukaryotes, all protozoan parasites feature a Hsp90 orthologue in their genomes. Plasmodium Hsp90 is essential, with geldanamycin effective at inhibiting parasite growth at sub-micromolar concentrations [18]. In Trypanosoma and Leishmania parasites, Hsp83 is implicated in thermally induced stage differentiation [19]. T. brucei Hsp83 (TbHsp83), the subject of interest herein, is 59% identical in sequence to human Hsp90α. This sequence identity increases to over 70% in the N-terminal ATPase domain, with notable sequence divergence only in a few short stretches (Fig. S1).
To explore the potential of TbHsp83 as a drug target, we employed a number of biochemical and biophysical techniques. Our approach began with the heterologous expression of the protein and the characterization of its enzymatic activity. Subsequently, a panel of known Hsp90 inhibitors was screened against the parasitic chaperone using differential scanning fluorimetry (DSF), with the results compared to their binding to the human Hsp90 isoforms. Efficacy of the compounds in inhibiting T. brucei growth correlated with the biophysical results. A chemical profile was generated from the screening results, highlighting chemical scaffolds that bind TbHsp83 more strongly than its human counterparts as well as those that have equal affinity for both. Structural analysis shows the binding modes of some of the more potent ligands, including those effective in killing the parasite in cellular assays, suggesting the possibility of developing anti-trypanosomal drug leads against TbHsp83.
The full-length coding region of T. brucei Hsp83 (gene Tb927.10.10980 - TritrypDB, http://www.Tritrypdb.org/ [20]) was cloned from genomic DNA. Full-length Hsp83 protein (Met1 to Asp704) and NTD (Met1 to Lys213) clones were obtained both including an N-terminal His6-tag. The two proteins were expressed and purified as previously described [21]. Briefly, clones were grown in TB media in a LEX bioreactor system (Harbinger Biotechnology and Engineering Corp., Ontario, Canada). Overnight starter cultures were left to grow at 37°C until reaching an OD600 value around 5, cooled to 15°C, and subsequently induced overnight with 0.5 mM IPTG. Cells were harvested by centrifugation and the pellets resuspended in 40 ml per liter of culture in 50 mM hepes pH 7.5, 500 mM NaCl, 5 mM imidazole, 5% glycerol, 1 mM benzamidine and 1 mM phenylmethyl sulfonyl fluoride (PMSF), then flash frozen in liquid nitrogen and stored in −80°C until needed.
The re-suspended pellets were pretreated with 0.5% CHAPS and 500 U of benzonase for 40 minutes at room temperature and cells were mechanically lysed using the M-110EH Microfluidizer Processor (Microfluidics Corp., MA, USA). The cell lysate was centrifuged to eliminate cells debris and the cleared lysate was loaded onto a DE52 anion exchange resin (Whatman, MA, USA) followed by a 2 mL Ni-NTA (Qiagen, MD, USA). The Ni-NTA column was then washed with 200 mL of a buffer consisting of 50 mM hepes pH 7.5, 500 mM NaCl, 30 mM imidazole and 5% glycerol. The protein was eluted with 15 mL of a buffer consisting of 50 mM hepes pH 7.5, 500 mM NaCl, 250 mM imidazole and 5% glycerol. Following elution, 1 mM EDTA and 1 mM TCEP were added to the sample. The sample was then incubated overnight with TEV protease at 4°C to cleave the His6-tag in 10 mM hepes, pH 7.5, 500 mM NaCl and 1 mM DTT. The imidazole was adjusted to 15 mM and the sample was loaded onto a pre-equilibrated 2.5 mL Ni-NTA column. The sample was allowed to bind to the nickel resin for 30 minutes after which the flow through containing the tag cleaved protein was collected. The protein was further purified by size exclusion chromatography on a Superdex 200 (GE Healthcare, NJ, USA) in the case of the full length protein or a Superdex 75 in the case of the NTD, both columns equilibrated with a buffer consisting of 10 mM hepes, pH 7.5 and 500 mM NaCl. The full length TbHsp83 peak fractions eluted at retention volumes consistent with a dimeric enzyme and monomeric for the NTD. The proteins were concentrated in an Amicon Ultra centrifugal filter device (Millipore, MA, USA) and their identities were confirmed by SDS-PAGE and mass spectroscopy.
The malachite green assay was used to characterize the ATPase activity of T. brucei Hsp83. We determined the amount of inorganic phosphate produced after incubating 95 nM TbHsp83 full-length 1 h at 37°C with its nucleotide substrate in 50 mM tris-HCl pH 7.5, 75 mM NaCl and 6 mM MgCl2. The substrate was evaluated at a concentration range from 1.9 µM to 2 mM. The kinetic parameters were obtained by fitting initial rate against substrate concentration using a nonlinear regression algorithm (SigmaPlot 2000 software; SPSS Inc., Chicago, USA).
Many of the Hsp90 inhibitors used in the assays reported herein were purchased from commercial sources (Supplemental Table S1), with six compounds commercially unavailable (Fig. 1) and synthesized using methods already reported in the literature: compound 1 (6-chloro-9-(4-methoxy-3,-dimethylpyridin-2-ylmethyl)-9H-purin-2-ylamine (BIIB021) [22]; compound 2 (9-(sec-butyl)-8-(2,5-dimethoxybenzyl)-2-methyl-9H-purin-6-amine) [23]; compound 3 – (2-amino-4-[2,4-dichloro-5-(2-diethylaminoethoxy)phenyl]thiopheno[2,3-d]pyrimidine-6-carboxylic acid ethylamide) [24]; compound 4 (4-[6,6-dimethyl-4-oxo-3-(trifluoromethyl)-4,5,6,7-tetrahydro-1H-indazol-1-yl]-2-[(trans-4-hydroxycyclohexyl)amino]benzamide) (SNX-2112) [25]; compound 5 (4-(6,6-dimethyl-4-oxo-3-(trifluoromethyl)-4,5,6,7-tetrahydro-1H-indazol-1-yl)-2-((2-(methylthio)ethyl)amino)benzamide) [25]; compound 6 (N4-[8-(5-acetylpyridin-2-yl)-8-azabicyclo[3.2.1]oct-3-endo-yl]-2-[(trans-4-hydroxycyclohexyl)amino]benzene-1,4-dicarboxamide) [26].
DSF screening was carried out using a LightCycler 480 Real Time PCR System (Roche Applied Science, Quebec, Canada). The NTDs from TbHsp83 as well as human Hsp90s α and β were buffered in 100 mM hepes pH 7.5, 150 mM NaCl and assayed in a 384-well format. The final concentration of the protein sample was optimized between 0.05 and 0.2 mg/ml for each NTD (concentration selected based on pre-testing results to avoid saturation of the fluorescence detector). The Hsp90 inhibitors were used at a final concentration of 25 µM (the full list of compounds is available as supplementary information, Fig. S2). SYPRO Orange (Molecular Probes, OR, USA) was added as a fluorescence probe in a dilution of 1∶1000. The experiments were conducted between 20°C to 95°C at a heating rate of 1°C per minute. The recorded fluorescence reads were fitted to the Boltzmann sigmoid function using Bioactive software (Harbinger Biotechnology and Engineering Corp., Ontario, Canada). The inflection point of each fitted curve is defined as the melting temperature (Tm). The observed temperature shift, ΔTm, was recorded as the difference between Tm of a sample and a reference in the same plate (i.e. Tm of the protein with ligand minus Tm of the protein without ligand). Thermal shifts above 2°C were considered significant.
ITC experiments were performed using a VP-ITC instrument (GE healthcare, NJ, USA). Injections of Hsp90 inhibitor solution were added to sample solutions of TbHsp83 NTD, with concentrations of 200 mM and 20 µM respectively. Titrations were conducted at 20°C in 100 mM hepes pH 7.5, 150 mM NaCl and 0.7% DMSO. The experimental data were fitted to a theoretical titration curve using the software package Origin (OriginLab Corporation, MA, USA), with ΔH (binding enthalpy in kcal mol−1), Ka (association constant) and n (number of binding sites per monomer), as adjustable parameters. Kd (dissociation constant) was calculated as 1/Ka. The standard free energy change and other thermodynamic parameters were calculated using the equations ΔG = −RT ln (Ka) and ΔG = ΔH−TΔS, where ΔG, ΔH, and ΔS are the changes in free energy, enthalpy, and entropy of binding, respectively.
A purified sample of the tag-cleaved TbHsp83 NTD was crystallized using the sitting drop vapor diffusion method in the presence of three different inhibitors (Fig. 1). All crystals were obtained by mixing one part of protein solution at 12 mg/ml (10 mM hepes pH 7.5, 500 mM NaCl, 2 mM TCEP and 4 mM MgCl2) containing 2 mM inhibitor with one part of reservoir solution. In the case of compound 1, the reservoir solution contained 2 M ammonium sulfate, 2% PEG400, 100 mM hepes pH 7.5; for the thienopyrimidine derivative compound it contained, 25% PEG 3350, 200 mM ammonium acetate and 100 mM hepes pH 7.5; and for the benzamidine derivative it included, 25% PEG 8000, 200 mM NaCl, 100 mM sodium cacodylate pH 5.5. Crystals appeared within three weeks and were cryo-protected in glycerol supplemented mother liquor before being flash cooled in liquid nitrogen. The diffraction data was collected using an X-ray source equipped with an R-Axis IV detector (Rigaku, TX, USA) and processed with HKL2000. The structures were determined by molecular replacement. The Leishmania major Hsp83 NTD (PDB code 3H80) served as a search model for the first complex structure of the T. brucei NTD, which was used as the model for subsequent structures. Phaser [27], [28] was used for the molecular replacement calculations, while model building was performed with COOT [29], [30] and the structures were refined with REFMAC5 [31] from the CCP4 suite of programs [32] and Buster-TNT [33]. Inhibitor coordinates and geometry restraints were created within the PRODRG topology server [34]. The stereochemistry of both models was checked by MOLPROBITY [35]. Relevant data collection and refinements statistics are shown in Table 1. The coordinates for the structure and their structure factors have been deposited with the Protein Data Bank (http://www.pdb.org [36]).
Structure superposition were calculated with LSQKAB [37] as implemented in the CCP4 package. Ligand-protein interactions were calculated and 2-D plots were generated by LigPlot+ [38]. The structure figures herein were generated with PyMOL (DeLano Scientific, Palo Alto, California, USA. http://www.pymol.org).
Assay conditions for T. brucei 427 cell based assay were followed as previously reported in the literature [39].
Kinetic constants of the enzymatic activity of TbHsp83 were determined using the standard malachite green assay. With Vmax = 37±1.6 pmoles P/min (Fig. S3), the ATPase activity of this parasitic chaperone is weak compared with other ATPases such as AAA (ATPases Associated with various cellular Activities) or P-type enzymes, but within the range observed for other Hsp90 proteins [7], [40], [41]. The Km (360 µM ATP) was also within the range of values previously determined for other Hsp90 chaperones from other organisms (Table 2).
The DSF assay results for TbHsp83, Hsp90α and Hsp90β in the presence of known or putative Hsp90 inhibitors are shown in a bar graph for NTD binders (Fig. 2A) and as a heat map for all the compounds tested (Fig. S2). Clearly, a number of compounds bind TbHsp83 with different levels of affinity, generating a chemical fingerprint (Fig. 2A). Furthermore, screening the human homologues as controls generated different fingerprints indicative of differences in chemical sensitivity between the parasite and human chaperones that might be exploited in the design of selective inhibitors. In particular, macbecin, geldanamycin and derivatives [6], and compound 5 demonstrated markedly stronger binding to TbHsp83, whereas radicicol and NVP-AUY922 had stronger affinity for human Hsp90. Overall, this experiment demonstrated the potential for designing parasite-specific inhibitors.
As a secondary assay, some of the compounds were tested with TbHsp83 using isothermal titration calorimetry. All the interactions measured were exothermic and enthalpy driven (Fig. 2B). Most of them displayed positive entropic contributions, with the exception of the interaction between the parasitic chaperone and compound 1 with a negative entropic term (ΔS = −35.6 kcal/mol). This interaction was also significantly more exothermic than the others suggesting an enthalpic compensation. Most of the measured interactions showed high affinity, with Kd values in the nanomolar range (Table 3), compounds 1 and 4, CAY10607 and NVP-AUY922 displayed sharp transitions curves in the binding enthalpies which could lead to an underestimation of the affinity (Fig. S4). Also the ITC derived affinities could have been affected in other compounds as visualized by poor baseline in the binding thermograms for compounds 2, 4, 5 and geldanamycin; and, molar rations higher than 1 for 17-AEP-GA, Macbecin I and NVP-AUY922. The compromised results were caused by limited solubility of some of the tested compounds and protein precipitation during the ITC experiments. But, the reported values represent the best values obtained from multiple repetitions. Furthermore, the Kd derived values are in agreement with previously reported affinities even in the case of compromised ITC measurements such as the one reported for NVP-AUY92 (Table 3). The highest affinity recorded for the Trypanosoma chaperone was in the low nanomolar range (Kd = 2.8 nM for a benzamide derivative, compound 4). As stated above, the recorded affinities were similar - within the same order of magnitude - to previously reported ITC measurements of Hsp90 from other species (Table 3) [42], [43], [44], [45], [46], [47]. The only exception was the 2.5 fold higher affinity displayed by T. brucei Hsp83 for macbecin I when compared to the Kd value reported for the yeast counterpart [45]. The ITC experiments generally corroborated and validated the DSF results (Fig. S5), and also provided thermodynamic insights about the interaction between TbHsp83 and the inhibitors.
Several of the compounds characterized using DSF and ITC assays were evaluated for their ability to kill T. brucei in a continuous in vitro culture system. Five out of the six compounds tested showed EC50 against the blood stream stage of T. brucei 427 in the submicromolar range (Table 3). The outlier compound, namely compound 2, also showed a non-significant ΔTm on TbHsp83 and the lowest affinity measured by ITC, consistent with correlation between binding affinity and in vitro efficacy (Fig. S5), although more data would be needed to corroborate the correlation.
We have crystallized the N-terminal domain of Hsp83 from T. brucei in complex with three different synthetic inhibitors – compounds 1, 3 and 4 (Fig. 3 & 4), but did not succeed in generating crystals of high acceptable diffraction quality in apo form or bound with ADP or ATP analogues. In general, the TbHsp83 crystals obtained were sensitive to the cryo-protection buffer and harvesting techniques. All the complex crystals were obtained by co-crystallization of the parasitic chaperone and the inhibitors. Two out of the three crystals belonged to the same space group but each one had unique cell parameters. Two of the crystals, thienopyrimidine (compound 3) and benzamide (compound 4) complexes, belonged to the orthorhombic system, containing two or three non-crystallographic symmetry (NCS) related complexes per asymmetric unit (ASU) respectively. The other crystal (compound 1 complex) was a primitive monoclinical with two copies per ASU. Despite these variations in space groups and cell parameters, no significant differences were observed between the three crystal forms in solvent content (VM values in Table 1) and most of the lattice contacts were shared. Extensive crystal contacts included an anti-parallel arrangement of the 1st and last β strands. Additional contacts included the loop regions connecting α helices 1 and 2, and helices 4 and 5.
Overall, the structure of the NTD of TbHsp83 was similar to previously determined Hsp90 nucleotide-binding domains (Fig. 3). The electron density in all three complexes was well defined for the whole polypeptide chain, facilitating the modeling of almost all of the cloned residues with the exception of the last few at the C-terminus. The final models showed good stereochemistry parameters. The resolution allowed us to explicitly model a significant number of solvent molecules. Also the electron density was unambiguous for the three different inhibitors present in each crystal form and for each molecule in the ASU (Fig. S6). Due to the absence of an apo Trypanosoma structure we compared it with a Leishmania Hsp90 in complex with AMPPNP and ADP also determined by our group (PDB codes 3H80 and 3U67, Hills et al., unpublished results). The structural differences observed between each complex and this previously determined parasitic Hsp90 were linked to the nature of the inhibitor bound to the nucleotide binding site and not to any Trypanosomes' specific structural features. Additionally, the structural superposition between each T. brucei Hsp83 complex to each other showed that each compound induced specific structural rearrangements because superposition of non-crystallographic symmetry (NCS) related molecules revealed lower r.m.s.d values than those obtained when superposing molecules with different inhibitors bound (Table 4).
From our results, the ATPase activity of the Trypanosoma chaperone is within the range of measured activities for this class of enzymes, which are characterized by a rather low turnover and high Km values (Table 2) [17]. In comparison with previously reported enzymatic constants the T. brucei chaperone has a slightly higher affinity for its substrate [7], [40], [41]. Lower Km values have also been reported for other Hsp90s, albeit at a lower incubation temperature 25°C [17]. In addition to the low affinity for its substrate, Hsp90 and its homologues tend to have slow turnover rates, ranging from 0.0001 to 0.02 s−1 [17]. Nonetheless, the T. brucei chaperone has a ten-fold higher turnover rate together with a higher affinity, which makes for a more efficient ATPase. There is no clear explanation for this enhanced ATPase activity but a similar observation has also been made for Plasmodium falciparum Hsp90 [7], another human parasite. This is an interesting trend to be evaluated as new activity reports from parasitic chaperones become available. A tantalizing hypothesis is that parasites might rely more heavily on their Hsp90 chaperones to maintain cellular homeostasis by stabilizing key cell regulators under hostile conditions. An observation that would be consistent with human Hsp90's pivotal role in some forms of cancer [49] (leading to its label as the Achilles' heel of malignant cells), a role that could also be exploited to treat parasitic infections [18]. Finally, a role has been determined for this chaperone in Candida albicans in the resistant phenotypes against commonly used antifungal drugs, then a plausible hypothesis could be that the parasitic hyperactive Hsp90 could facilitate the emergence of drug resistant parasites [50], [51].
Previous studies have determined the sensitivity of several parasitic chaperones to known inhibitors, such as geldanamycin and radicicol, and their deleterious effects on parasite growth [6], [7], [8], [9], [10], [52], [53]. To identify and characterize inhibitors of TbHsp83, we used the DSF technique to screen a collection of near 40 compounds described in the literature as known or putative Hsp90 binders. The bulk of the collection includes ATP-competitive inhibitors predicted to bind the N-terminal region but also several putative C-terminal inhibitors. Initially, our DSF assays were performed with both full-length proteins and individual domains and two observations were derived from these experiments.
In the case of the human Hsp90s, there appears to be a similar ΔTm for both the N-terminal domain and the full length protein, whilst in the case of the T. brucei Hsp83, the ΔTm appears to be greater for the N-terminal domain, with a significantly weaker shift for the full length protein. The T. brucei C-terminal domain, the middle domain and the combined C-terminal & middle domain were also tested and did not show any significant ΔTm, as would be expected for compounds which bind in the N-terminal ATP binding site (Fig. S2). Taken together, these observations suggest (i) the compounds interact primarily with the N-terminal ATPase domain, as most of the them are known to do with other Hps90 proteins; (ii) the DSF assay is an effective technique for identifying potential Hsp90 inhibitors; (iii) we can focus on the N-terminal ATPase domain for characterizing and optimizing inhibitors of TbHsp83.
We observed an inverse relationship (R2 = 0.88) existed between ITC-derived Kd values and DSF-derived ΔTm values, with the caveat that some of the ITC reported affinities might be under or overestimated affecting the correlation between the two assays (Fig. S5). Such a relationship has been reported for other proteins such as protein kinases [54] and bromodomains [55], but with even higher correlation coefficients. The observed correlation would likely be greater if the tested compounds were structurally related because the correlation is stronger for compounds sharing a common binding mode. By virtue of its operating principles, the DSF method is particularly sensitive to entropy-driven binding [56]. Therefore, when contributions of enthalpy and entropy are different, change in ΔTm may not correspond to Kd [57], [58], [59]. The lower detection limit of our thermal shift assay is in the micromolar range when using a compound concentration of 25 µM (e.g., compound 2 with ΔTm = 0.8°C and Kd = 9.3 µM.). The observed limit explains the absence of hits for the C-terminal domain since they have been known to bind with low affinity (10−3 to 10−6 M range). Nonetheless, our approach has been successful in identifying high affinity binders by screening a narrow set of Hsp90 inhibitors rather than a broad diverse library. Our results support the DSF assay as a suitable technique to screen chemical libraries originally designed against human targets, as we screened compounds originally designed as anti-human Hsp90 inhibitors.
The three-dimensional structure of the NTD of T. brucei Hsp83 in complex with three different inhibitors showed the large degree of structural flexibility of the chaperone. As was previously described, the nucleotide-binding site can be divided into the adenosine and the phosphate binding pockets [25]. All the inhibitors described here occupy both pockets but differ in how they expand or modify these binding cavities [60], [61], [62]. All of the inhibitor scaffolds recapitulate the adenosine mode of binding to the chaperone (Fig. 4 & 5). Previous structures have established that the base is recognized by its Watson-Crick face generating two hydrogen bonds with the NTD; one is a water-mediated interaction while the other one is between the base amino group and the side chain of a well conserved aspartic acid residue (Asp78 in T. brucei). All of the inhibitors described here preserve these interactions, including the water mediated one. Interestingly, the nucleotide-binding site of the NTD is highly conserved among different Hsp90 family members, with the exception of Ile171 (located in the region of sequence divergence highlighted in orange in Fig. S1) being a moderately variable position located at the bottom of the adenosine-binding pocket (Fig. S8). This side chain is in contact with residue Leu34 and indirectly with Ile35, a pair of residues directly implicated in resistance to the natural product radicicol [46]. This observation could rationalize the significant differences measured by our DSF assay between the TbHsp83 and the human α and β Hsp90s with respect to radicicol. It also highlights the large impact that small sequence differences could have in the chemical sensitivity of the parasitic chaperone.
The thienopyrimidine derivative, namely compound 3, was the only compound to establish charged contacts in the periphery of the adenosine binding pocket with residues involved in the interaction with the sugar moiety of the phosphonucleotide, such as Asn91, while the other two compounds engaged these residues in non-polar interactions (Fig. 4 & 5). As a result, this region located at the end of helix-3 between residues 87 to 93 showed the largest structural differences among the three complex structures with values exceeding 4 Å r.m.s.d. for the main chain atoms of residues Asn91 and Asn92 (Fig. 6). This structural rearrangement tilted the position of helix 4, a key structural element in the NTD, because together with helix 5 it composes the lid region. The closing of the lid over the ATP during the catalytic cycle is a preliminary step also linked to the dimerization of the NTDs, as it allows the movement of helix 1 and facilitates the exchange of the N-terminal β-strands among interacting NTDs. On the basis of the large structural movements displayed during the ATPase cycle of Hsp90, we propose the region located at the end of helix 3 plays an important role as a sensor for the binding of nucleotides a function facilitated by its structural plasticity and highlighted by recent NMR studies of the Hsp90 [63]. In our case, the region of structural flexibility allows the NTD to accommodate a diverse set of compounds generating unique binding regions for each one. The uniqueness of the conformations induced by each compound to the region encompassed by residues 87 to 93 has extensive repercussions in the lid position. Each complex structure showed a unique open conformation of the lid region (Fig. 6) with the subsequent impact on the phosphate-binding pocket (Fig. 7). This structural flexibility was already described for the NTD when binding both substrate and inhibitors and two main conclusions can be drawn from these observations. First, due to the large structural rearrangements involved upon inhibitor association, the contribution from regions other than the nucleotide-binding pocket can be quite substantial and hard to predict by analyzing the structure alone. Second, the cellular effects of Hsp90 inhibitors are related to the conformational limitations imposed on the chaperone, affecting co-chaperone binding and long-range coupled motions observed between the NTD and CTD domains [64].
Compound 5 preferentially binds TbHsp83 with high affinity but not human Hsp90, but did not yield diffracting crystals in our attempts. Therefore, we are unable to provide a definitive structural explanation for this benzamide's selective behavior. We also cannot utilize existing human Hsp90 structures and our TbHsp83 structures for this purpose because, as explained above, the chaperone's flexibility allows it to be reshaped differently by different classes of inhibitors. The chemical selectivity of compound 5 suggests that it uniquely recognizes the difference between the flexibility the parasite and human chaperones. Remarkably, the differentiation is a product of one or more of the three short regions of sequence divergence shown in Fig. S1.
This study presents a detailed structural and chemical characterization of TbHsp83 as a potential drug target against African trypanosomiasis a fatal neglected disease. We have identified a chemical scaffold with a preference for the Trypanosoma chaperone over its host counterpart. Furthermore, the assayed compounds were able to inhibit the growth of the parasite in vitro in a manner that correlate strongly to their binding affinities to TbHsp83, validating this parasite chaperone as a potential target to combat sleeping sickness. Target validation and early lead identification represent the initial stages of a potential drug development program. Furthermore, in the Hsp90 case, there is the possibility of taking advantage of the much more advanced anti-Hsp90 programs in the cancer field. We believe that tapping into other therapeutics areas such as cancer might have a great value in the development of drugs against neglected tropical diseases.
Generally speaking, a potential weakness in repurposing compound libraries designed against human targets to search for anti-parasitic compounds is that the starting chemical points are not selective. What is important, however, is to show the possibility of differentiating structure activity relationship between the human and parasite enzymes. In spite of the high degree of sequence identity with the human homologue, we have found such a chemical differentiator for the NTD of TbHsp83 in compound 5 and a number of other analogues. More research is required to substantiate and elaborate on our hypothesis that this selectivity is a result of differences in the flexibility of the human and parasite chaperones.
|
10.1371/journal.pntd.0000803 | Novel Naphthalene-Based Inhibitors of Trypanosoma brucei RNA Editing Ligase 1 | Neglected tropical diseases, including diseases caused by trypanosomatid parasites such as Trypanosoma brucei, cost tens of millions of disability-adjusted life-years annually. As the current treatments for African trypanosomiasis and other similar infections are limited, new therapeutics are urgently needed. RNA Editing Ligase 1 (REL1), a protein unique to trypanosomes and other kinetoplastids, was identified recently as a potential drug target.
Motivated by the urgent need for novel trypanocidal therapeutics, we use an ensemble-based virtual-screening approach to discover new naphthalene-based TbREL1 inhibitors. The predicted binding modes of the active compounds are evaluated within the context of the flexible receptor model and combined with computational fragment mapping to determine the most likely binding mechanisms. Ultimately, four new low-micromolar inhibitors are presented. Three of the four compounds may bind to a newly revealed cleft that represents a putative druggable site not evident in any crystal structure.
Pending additional optimization, the compounds presented here may serve as precursors for future novel therapies useful in the fight against several trypanosomatid pathogens, including human African trypanosomiasis, a devastating disease that afflicts the vulnerable patient populations of sub-Saharan Africa.
| African sleeping sickness is a devastating disease that plagues sub-Saharan Africa. Neglected tropical diseases like African sleeping sickness cause significant death and suffering in the world's poorest countries. Current treatments for African sleeping sickness either have high costs, terrible side effects, or limited effectiveness. Consequently, new medicines are urgently needed. RNA editing ligase 1 is an important protein critical for the survival of Trypanosoma brucei, the unicellular parasite that causes African sleeping sickness. In this paper, we describe our recent efforts to use advanced computer techniques to identify chemicals predicted to prevent RNA editing ligase 1 from functioning properly. We subsequently tested our predicted chemicals and confirmed that a number of them inhibited the protein's function. Additionally, one of the chemicals was effective at stopping the growth of the parasite in culture. Although substantial work remains to be done in order to optimize these chemicals so they are effective and safe to use in human patients, the identification of these parasite-killing compounds is nevertheless a valuable step towards finding a better cure for this devastating disease.
| Subspecies of Trypanosoma brucei (T. brucei) are the causative agents of human African trypanosomiasis (HAT, also known as African sleeping sickness) in sub-Saharan Africa. Neglected tropical diseases, including parasitic trypanosomal illnesses like HAT, Chagas disease, and leishmaniasis, are responsible for the loss of an estimated 56.6 million disability-adjusted life-years across several regions, particularly in the world's poorest countries [1]. Preventative measures such as vector control are effective at decreasing the incidence of HAT; however, given infection, the current treatment options are not suitable on the whole, particularly once T. brucei has infiltrated the central nervous system [2].
First-stage treatments include pentamidine and suramin, drugs developed more than half a century ago. Unfortunately, these drugs have severe side effects. Pentamidine is associated with hypoglycaemia and hypotension, while suramin is associated with anaphylactic shock, neurotoxic signs, severe cutaneous reactions, and renal failure [3]. The most common treatment for second-stage HAT is melarsoprol, a highly toxic drug with a 3%–10% fatality rate [4]. The danger of treatment is compounded by the emergence of melarsoprol-resistant parasites, particularly in central Africa [5]. Eflornithine, another HAT treatment, is less toxic but only effective against the T. b. gambiense subspecies; additionally, eflornithine is more costly to produce than melarsoprol [6]. Given the weaknesses of current treatments, new drugs are urgently needed.
Fortunately, recent studies of the trypanosomal editosome have revealed several new drug targets. In trypanosomatids, mitochondrial gene expression includes an extra RNA-editing step. As in other eukaryotes, mitochondrial DNA is transcribed into RNA. In trypanosomes and Leishmania parasites, however, a protein complex known as the editosome makes extensive uridylate (U) insertions and deletions following transcription, at times even doubling the length of the original RNA sequence [7]–[11]. After each cycle of U addition or deletion, a nick in the RNA remains; RNA editing ligase 1 (TbREL1; TriTrypDB ID: Tb927.10.8210), an essential enzyme in trypanosomes [12], is one of two ATP-dependent editosome ligases responsible for religation. As the editosome is absent in humans, the proteins of this complex, including REL1, are potential drug targets in all trypanosomatid pathogens.
Recently Amaro et al. identified several novel TbREL1 inhibitors. The relaxed complex scheme, a virtual-screening methodology that accounts for full protein flexibility [13], was used to identify five low-micromolar inhibitors from among the compounds of the National Cancer Institute Diversity Set I [14]. Unfortunately, these TbREL1 inhibitors were ineffective against whole-cell T. brucei, perhaps in part because they are too hydrophilic to cross lipid membranes (unpublished work).
Motivated by both the urgent need for novel trypanocidal therapeutics as well as the success of virtual screening against TbREL1 in the past, we here use the relaxed complex scheme to identify additional naphthalene-based inhibitors in hopes of finding compounds that can kill T. brucei. To this end, online databases of commercially available compounds were first searched for compounds similar to the inhibitors previously characterized. Following virtual screening, the most promising of these compounds were subsequently tested experimentally, revealing four novel TbREL1 inhibitors with unique naphthalene-based scaffolds, two of which have ALogP values that suggest reasonable lipophilicity. Analyses of the predicted binding modes of these active compounds, performed using an ensemble-based approach and coupled with computational fragment mapping experiments, suggest that receptor flexibility may play an important role in ligand binding.
To generate a library of compounds similar to the TbREL1 inhibitors characterized previously [14], we performed online substructure searches of several databases of commercially available compounds, including Hit2Lead (Hit2Lead.com, ChemBridge), the NCI/DTP Open Chemical Repository (dtp.cancer.gov), Sigma-Aldrich (sigmaaldrich.com), and ZINC [15]. Searches were performed using three structures similar to the core naphthalene scaffolds of known inhibitors: naphthalene-2-sulfonic acid, 2-naphthoic acid, and 2-nitronaphthalene (Figure 1).
The compounds identified via online substructure searches were each docked into a 1.20-Å resolution crystal structure of the TbREL1 catalytic domain (PDB ID: 1XDN) [16] using AutoDock 4 [17]. Ligand files were processed with AutoDockTools 1.4.5 to merge nonpolar hydrogens with parent heteroatoms and to assign Gasteiger charges. AutoGrid affinity grids contained 86×72×78 points spaced 0.375 Å apart, centered on the TbREL1 active site, the ATP-binding pocket. Grid files were created for the following ligand atom types: A (aromatic carbon), C, F, I, N, NA (hydrogen-bond accepting N), Cl, OA (hydrogen-bond accepting O), P, S, SA (hydrogen-bond accepting S), Br, HD (hydrogen-bond hydrogen), and e (electrostatic).
AutoDock parameters similar to those published previously by Amaro et al. [14] were used: population size 200; 5,000,000 evaluations; 27,000 generations; 100 runs; and cluster tolerance of 2.0 Å. All other AutoDock parameters were set to the default values. The correct docked pose was judged to be the lowest-energy pose of the most populated cluster.
With the intent of rescoring the top hits from the initial crystal-structure screen in a way that accounts for full protein flexibility, we drew upon a previous study of TbREL1 molecular motions [18]. In brief, molecular dynamics (MD) simulations of TbREL1 were performed using NAMD 2.6 [19]. Four hundred receptor conformations were extracted from the MD simulations, one every 50 ps. QR factorization [20] was used to eliminate conformational redundancy, thereby reducing the number of representative structures from 400 to 33 [14]. These 33 TbREL1 structures are said to constitute an ensemble representative of the many protein conformations sampled during the MD simulation.
The relaxed complex scheme (RCS) was subsequently used to rescore the top compounds from the initial crystal-structure screen [13]. AutoDock was used to dock each of the top inhibitors into the 33 protein conformations of the receptor ensemble using the same docking parameters described above. The ensemble-average binding energy of each ligand was computed by taking the simple mean, and the ligands with the best mean predicted binding energy were subsequently tested experimentally.
To partition the ATP-bound trajectory [18] into a set of structures representing regions of decreasing conformational population density, RMSD clustering, distinct from the QR factorization described above, was performed [21]–[23] as implemented in the rmsdmat2 and cluster2 programs of the GROMOS++ analysis software [24]. Four hundred receptor conformations were extracted from the 20 ns ATP-bound MD trajectory, one every 50 ps. Clustering was performed on a subset of 24 residues that line the ATP binding cleft: 87–90, 155–162, 207–209, 283–287, and 305–308. These residues constitute the 5 conserved motifs of the nucleotidyltransferase superfamily [25], [26] to which TbREL1 belongs. The trajectory frames were first aligned by minimizing the RMSD between the alpha carbons of the 24-residue subset of each frame and the corresponding alpha carbons of the first frame. This least-squares alignment removed external translational and rotational motion so that subsequent RMSD calculations could focus on the internal conformational variability of the 24-residue subset. After varying the RMSD similarity criterion from 0.06 to 0.12 Å, a value of 0.085 Å was chosen, as this cutoff produced 8 clusters of protein conformations. The three most populated clusters comprised 93.5% of the trajectory.
Computational fragment mapping (FTMap, http://ftmap.bu.edu) was utilized to identify druggable regions on the surface of TbREL1. The FTMap algorithm [27] determines the energetically favorable binding regions of sixteen fragments along a protein surface (Figure S1) via the following steps: (1) rigid body docking of fragments using a fast Fourier transform approach, (2) minimization and rescoring of fragment-protein complexes, (3) clustering and ranking of low-energy fragment-protein complexes, and (4) determination of consensus sites. Consensus sites are regions of the protein surface where low-energy fragment clusters of multiple fragment types co-localize; in previous studies using FTMap and its predecessor CSMap [28], highly populated consensus sites were shown to correlate strongly with ligand binding hot spots identified via biophysical methods [27], [29], [30].
The top ranked compounds from the relaxed complex screen were obtained for testing in experimental assays. Compounds were provided by the Developmental Therapeutic Program at the National Cancer Institutes (NCI) of Health, Hit2Lead.com, and Sigma-Aldrich (Table S1). Compounds V1, V2, and V3 (Figure 1) were provided by the NCI, and compound V4 was purchased from Sigma. All compounds were dissolved in DMSO or DMSO/H2O.
The protocols for recombinant TbREL1 expression, purification, and assaying have been described previously [14]. In brief, recombinant full-length TbREL1 was expressed in Sf9 insect cells after infection with recombinant baculovirus and purified via a C-terminal tandem affinity purification (TAP) tag. To measure enzyme inhibition, 0.1 pmol TbREL1 was incubated with 1.8 µCi (30 nM) [α-32P]ATP in assay buffer (25 mM KCl, 12.5 mM HEPES pH 7.9, 5 mM Mg acetate, 0. 25 mM DTT, 0.1% Triton X-100) for 5 min at room temperature and in the presence of varying concentrations of the potential inhibitors. The extent of protein adenylylation (and therefore competition with ATP for binding to the active site) was subsequently measured by SDS/PAGE and phosphorimaging. All reactions were done in at least triplicate, and IC50 values were calculated using the GraphPad Prism 5 software.
The effect of the identified REL1 inhibitors on parasite growth was determined using the Alamar Blue assay, essentially as described by Räz et al. [31]. Briefly, T. brucei brucei cells (strain s427) were seeded in 96-well plates at a density of 1×104 cells per ml in a volume of 200 µl, in the presence of varying concentrations of predicted inhibitors or DMSO alone. After 48 hours, 20 µl Alamar Blue (Invitrogen) were added to the cells and incubation continued for an additional 24 hours. Absorbances at 540 and 595 nm were measured using an ELx808 Microplate Reader (BioTek), and EC50 values were calculated using the GraphPad Prism 5 software.
RNA editing ligase 1 (REL1) is a key component of the trypanosomatid editosome. In trypanosomatid parasites (i.e. species of Trypanosoma and Leishmania), mitochondrial mRNA requires editing following transcription; after each round of U addition or deletion, REL1 and the related protein REL2 religate the RNA in an ATP-dependent reaction. REL1 is a noteworthy drug target because it is required for the survival of T. brucei [12], [32] and presumably other trypanosomatids as well. Additionally, no close human homologues have been identified [14]. The heavy disease burdens caused by human African trypanosomiasis (HAT), Chagas disease, and leishmaniasis, as well as the urgent need for novel trypanocidal therapeutics and the success of virtual screening against TbREL1 in the past, have motivated the current work, wherein we identify novel TbREL1 inhibitors with naphthalene-based scaffolds.
Previously, Amaro et. al identified several micromolar inhibitors of TbREL1 [14]. The top three inhibitors identified were all based on a naphthalene-2,7-disulfonate (NDS) scaffold. In silico docking provides insight into why this scaffold is amenable to TbREL1 inhibition (Figure 2). Similar to the adenine moiety of ATP (the native co-factor), the NDS naphthalene group is able to form π-π stacking interactions with F209. Additionally, one of the negatively charged NDS sulfonate groups interacts electrostatically with the positively charged R111 guanidino group at the active-site periphery; R111 also participates in electrostatic and hydrogen-bond interactions with the ATP polyphosphate tail. A hydrogen bond is formed between NDS and N92, similar to the hydrogen bond formed with the O2' oxygen atom of the ATP ribose. Finally, docking suggests that the second of the two NDS sulfonate groups is buried deep within the binding pocket, displacing a water molecule that normally mediates a hydrogen-bond network between the ATP adenine N1 atom and R288. This water displacement allows the sulfonate group to interact with the charged R288 residue directly.
Unfortunately, these previously identified TbREL1 inhibitors were ineffective against whole-cell T. brucei, likely because they are too hydrophilic to cross lipid membranes (A. Schnaufer, unpublished work). Interestingly, these compounds show similarities to the anti-trypanosomal drug suramin, which, although much larger, also has a negatively charged polysulfonated naphtyl group [14]. How suramin enters the cell is unclear, but both fluid-phase endocytosis and receptor-mediated uptake have been suggested [33], [34]. Suramin both binds various serum proteins, which may facilitate uptake by the trypanosome cell [34], –and inhibits a considerable number of enzymes, including dehydrogenases and kinases in various organisms and glycolytic enzymes in T. brucei [35]. This promiscuous binding may be in part attributable to the negatively charged sulfonate groups [35]. Additionally, the hydrophilicity these sulfonates impart likely impedes both suramin and the previously identified REL1 inhibitors from passively crossing inner cellular membranes to reach organellar targets such as mitochondrial proteins.
In an attempt to identify additional naphthalene-based TbREL1 inhibitors with improved pharmacological properties, we searched several online databases of commercially available compounds for similar structures: naphthalene-2-sulfonic acid, 2-naphthoic acid, and 2-nitronaphthalene (Figure 1). These searches identified 588 compounds: 61 compounds from Hit2Lead (Hit2Lead.com, ChemBridge), 394 from the NCI/DTP Open Chemical Repository (dtp.cancer.gov), 87 from Sigma-Aldrich (sigmaaldrich.com), and 46 from ZINC [15]. In all, the search identified 376 naphthalene-2-sulfonic acid compounds, 130 2-naphthoic acid compounds, and 85 2-nitronaphthalene compounds.
Given its previous successful identification of TbREL1 inhibitors, AutoDock 4.0 was utilized for docking. Although the AutoDock scoring function sacrifices accuracy for speed as compared to more rigorous methodologies such as thermodynamic integration [17], [36], single-step perturbation [37], and free energy perturbation [38], AutoDock performs well [39] when compared to other docking programs such as DOCK [40], FlexX [41], and GOLD [42].
The 588 compounds identified through online substructure searches were first docked into a 1.20-Å resolution crystal structure of the catalytic domain of TbREL1 [16]. AutoDock placed 14% of the naphthalene compounds in the expected pose (26% of the 2-naphthoic acid compounds, 10% of the naphthalene-2-sulfonic acid compounds, and 8% of the 2-nitronaphthalene compounds), with the naphthalene portion of the ligand buried deep in the ATP-binding pocket and the electronegative group at the two position either interacting with R288 or with R111 at the active-site periphery.
The preliminary docking to the TbREL1 crystal structure, while useful for eliminating those structures that were grossly incompatible with the TbREL1 active site, did not account for full protein flexibility. Aside from the inaccuracies inherent in docking scoring functions themselves, docking accuracy decreases further when protein and/or ligand flexibility are ignored. When a ligand approaches a protein receptor in solution, it does not encounter a single static protein conformation, but rather an ensemble of many different conformations. Often, a given ligand may only bind to a certain subset of all protein conformations sampled, depending in part on the varied side-chain positions of active-site residues. When multiple protein conformations are incorporated into a virtual-screening protocol, the hit rate can drastically improve; ligands that do not bind to the crystal structure may bind to other related protein conformations. Screening against these other conformations in principle reduces the false negative rate.
Of the top-ranked 100 binders from preliminary crystal-structure screens, 45 shared significant structural similarity with the most potent compound identified previously by Amaro et al. [14]. In order to account for full protein-receptor flexibility, these 45 compounds, roughly corresponding to the top 7.5% of the library, were docked into 33 protein receptor conformations extracted from a MD simulation of TbREL1 [14] using QR factorization [20]. The 45 ligands were then reranked by their respective ensemble-average scores, and 12 of the top compounds (Table S1) were subsequently tested experimentally.
Prior to RNA ligation, a key TbREL1 lysine must first be adenylylated. To measure the inhibition of this first step of the reaction pathway, the formation of TbREL1-[32P]AMP was monitored via SDS/PAGE and autoradiography in the presence of predicted inhibitor. Triton X-100 (0.1%) was added in order to prevent aggregate-based inhibition. Four compounds, V1, (E)-7-benzamido-4-hydroxy-3-((5-hydroxy-7-sulfonaphthalen-2-yl)diazenyl)naphthalene-2-sulfonic acid; V2, (E)-7-amino-4-hydroxy-3-((5-hydroxy-7-sulfonaphthalen-2-yl)diazenyl)naphthalene-2-sulfonic acid; V3 (Di-J acid); and V4 (Mordant Black 25), inhibited TbREL1 activity with IC50 values of 2.16±1.20 µM, 1.53±1.17 µM, 8.36±1.71 µM, and 1.59±1.1 µM, respectively (Table 1, Figure 1). Additional information about the predicted binding poses of these four validated inhibitors can be found in Table S1. An additional four compounds inhibited TbREL1 adenylylation with IC50 values between 10 and 100 µM; the exact values in these cases were not determined (Table S1). All other compounds did not show significant inhibition at 100 µM.
Interestingly, the crystal-structure protein conformation used for the initial docking is likely itself suboptimal for the binding of the four inhibitors identified, as evidenced by the improvement in rank when an ensemble-average AutoDock score was used (RankEnsemble) instead of the crystal-structure score (RankCrystal, Table 1). In fact, only one of the four compounds, V1, scored in the top twelve when all 588 compounds were docked into the crystal structure alone. V2, V3, and V4, which ranked 20th, 31st, and 25th against the crystal structure, respectively, may not have been tested had the ligand set not been reranked by an ensemble-average AutoDock score. A direct comparison of the predicted binding energy of the four indentified inhibitors docked into the crystal structure (AutoDockCrystal) and docked into the optimal protein conformation from the ensemble (AutoDockEnsemble/Best) likewise demonstrates the importance of accounting for full protein flexibility; in all four cases, predicted energies of binding improved several kcal per mol when the optimal structure was used rather than the crystal structure (Table 1).
In order to investigate why binding to the crystal structure was suboptimal, the crystal structure was compared to the optimal receptor conformation for each of the four ligands. By aligning the best-scoring MD-generated receptor structures to the crystal structure and visualizing both proteins and ligands, it is evident that in all four cases the crystallographic position of E60 prevented optimal binding. During the molecular dynamics simulation, however, E60 extends its contact with R111 (initial contact distance 5.35 Å; final interaction distance greater than 11 Å). This movement opens a wide cleft that is favorably occupied by all four of the novel inhibitors (Figure 3). This unique binding mode, described in more detail below, would not have been identified had protein-receptor flexibility been ignored.
To further explore the role that receptor flexibility plays in inhibitor binding, we grouped the frames of the MD trajectory into sets of geometrically similar conformations using an RMSD-based clustering algorithm. Each cluster contains a central structure, or centroid, whose structural characteristics and binding properties are representative of all cluster members. Similar to QR factorization [14], [20], RMSD clustering reduces the MD ensemble to a representative set of (centroid) conformations. However, unlike QR factorization, RMSD clustering provides an approximate idea of the probability of sampling a set of geometrically similar conformations based on the fraction of conformations contained within each cluster [21].
Assuming the conformations sampled along the inhibitor-bound trajectories are similar to those observed during the ATP-bound trajectory, the receptor-inhibitor interactions characteristic of the most populated clusters, which represent the most frequently visited system conformations, should contribute most to ligand affinity. Indeed, the representative protein structure that best accommodates V1, V3, and V4 from the QR-factorization ensemble, as judged by the AutoDock score, belongs to the most populated RMSD-based cluster. The protein conformation that best accommodates V2 belongs to the second most populated RMSD-based cluster.
The conformations sampled by the MD trajectory were grouped into 8 clusters when an RMSD similarity cutoff of 0.085 Å was used; 93.5% of the trajectory conformations were contained in the three most populated clusters. The conformational variability among the centroids of the top three clusters suggests two dynamically distinct active-site regions. Deep within the inhibitor-binding cleft, where F209 forms π-π stacking interactions with the sulfonated naphthalene moiety of each inhibitor, the conformational differences among the centroids are modest, consisting of only subtle amino-acid side-chain shifts (Figure 4A). Given the rigidity of this region and the similarity between naphthalene and the adenosine of ATP, the native TbREL1 substrate, we hypothesize that the naphthalene scaffold is highly complimentary to the modest conformational fluctuations observed at the deep end of the binding pocket.
In contrast, conformational variability at the binding-site periphery near the solvent interface is much larger (Figure 4B). As the predicted binding modes of the validated inhibitors initially suggested, the varied positions of E60 relative to R111 are particularly notable. In the centroid conformation of the first and second most populated clusters, a cleft is again seen between E60, which is directed into bulk solvent, and R111, which is directed toward the inhibitor binding site. The distances between E60(OE2) and R111(NH1) are 9.01 Å and 10.96 Å, respectively. These two open-cleft clusters represent 83% of the entire trajectory. In the centroid conformation of the third most populated cluster, representing 11% of the entire trajectory, the cleft is narrowed; E60 is directed downward, toward R111, and the distance between E60(OE2) and R111(NH1) is only 7.14 Å, closer to the closed-cleft crystal-structure distance of 5.35 Å. As noted previously, all four novel inhibitors are predicted to occupy this previously uncharacterized cleft, suggesting that it is pharmacologically important.
This new cleft also presents an opportunity to develop compounds with improved specificity over the related human DNA ligases. A structural and sequence alignment of the superfamily members [18] reveals key sequence differences in relative positions between REL1 and human DNA ligase (PDB: 1X9N). In REL1, residues I59-E60-I61-D62 line the newly revealed cleft and make contact with several of the bound inhibitors. In human DNA ligase 1 (PDB 1X9N), the equivalent residues are M543-L544-A545-H546. The strategic design of REL1 inhibitors to take advantage of the variable contacts in this area, particularly the exposed side chains of the residues lining the cleft, may present novel avenues to design compounds with increased selectivity for the trypanosomal enzymes.
To explore the pharmacological importance of the E60-R111 cleft in greater depth, computational fragment mapping was carried out on both the centroids of the three most populated clusters as well as the crystal structure (Figure 4C). Computational fragment mapping estimates the binding affinity of fragment-sized organic groups and clusters them into consensus-binding regions. These consensus-binding regions (a.k.a. hot spots) represent regions of receptor sites that are the principal contributors to the ligand-binding energy. Importantly, these computationally predicted sites have been shown to correlate well with fragment-binding hot spots as determined via biophysical experiments in numerous studies [27], [29], [30].
Fragment mapping confirmed that the TbREL1 active site can be divided into two regions, as two consensus sites were apparent. The first site, conserved among the centroids of the three most populated clusters as well as the crystal structure, is found deep in the inhibitor-binding cleft, where both the adenine of the native ATP substrate and the sulfonated-naphthalene moieties of the novel inhibitors bind. The conservation of this solvent cluster supports the pharmacological importance of this region and is in harmony with the predicted docking poses of the four novel inhibitors.
The second consensus site is found in the previously uncharacterized E60-R111 cleft. Notably, while conserved among the three most populated clusters, this site is entirely absent in the crystal structure, likely because the closed E60-R111 cleft of that structure occludes solvent-probe binding. Naphthalene-based inhibitors docked into the crystal structure are predicted to interact only with the high-affinity region deep in the binding pocket; at the active-site periphery, binding to the high-affinity region in the E60-R111 cleft is impossible, and so the predicted binding affinity is less favorable. Hence, the fragment-mapping approach supports the presence of an additional pharmacologically relevant feature of the ATP binding pocket. It also helps to explain why those compounds eventually confirmed as genuine inhibitors were not initially ranked among the top-scoring candidates.
While fragment mapping did reveal a high-affinity region in the E60-R111 cleft of the centroid representing the third most populated cluster, this region does not extend as far into the cleft as the corresponding clusters of the top two centroids. This fact, together with the narrower cleft width, may partly explain why none of the four novel inhibitors was predicted to bind to receptor conformations of the third most populated cluster.
In order to analyze the predicted binding mode of the four confirmed TbREL1 inhibitors, the protein conformation from the ensemble generated by QR-factorization that gave the best AutoDock-predicted binding energy (i.e. the “optimal receptor”) was visualized together with the associated docked ligand. In all cases, the electronegative group at the naphthalene C2 position was buried deep within the active site, forming interactions with R288, as expected. Additionally, three of the four ligands, similar to the three most potent TbREL1 inhibitors identified previously [14], had hydroxyl groups in the naphthalene 4 position, suggesting that the hydrogen bonds formed with E86 and V88 are also critical to ligand binding (Figure S2, upper rows). A fourth ligand, V4, had a hydroxyl group in the naphthalene 6 position, were it could form hydrogen bonds with the backbone carbonyl oxygen atom of V88 and the side-chain amino group of K87.
At the active-site periphery, all four of the confirmed inhibitors had secondary sulfonate groups that docked near the more positively charged side of the active-site periphery, opposite the R111 residue (Figure S3), where they interact with K307, R309, and K87 (Figure S2, bottom rows). In contrast, the peripheral, negatively charged sulfonate groups of previous NDS inhibitors, substituents of the naphthalene core itself, interacted principally with R111. The new inhibitors do not entirely neglect R111, however; all four compounds are predicted to participate in π-cation interactions with this residue.
In addition to these electrostatic interactions, the four novel inhibitors are predicted to interact with other protein residues at the active-site periphery (Figure S2, bottom rows). In some ways, these interactions mimic the interactions between TbREL1 and its native substrate, ATP. V1 forms a hydrogen bond with the R111 guanidinium group, similar to the bond formed between R111 and the ATP gamma phosphate. V1 also forms a hydrogen bond with the E159 side-chain carboxylate group, similar to the bond formed with the ATP 2′ ribose hydroxyl group. V1 forms unique interactions with TbREL1 as compared to the substrate; V1 forms a hydrogen bond with the backbone carbonyl of Y58, a residue that does not participate in ATP binding (Figure S2A).
V2 is predicted to participate in only one hydrogen bond at the active-site periphery. This bond is formed with the E60 side-chain carboxylate group, a group that does not participate in ATP binding (Figure S2B). V3 and V4 are likewise predicted to form only one hydrogen bond at the active-site periphery, a bond with the I59 backbone carbonyl. This same backbone carbonyl forms a hydrogen bond with the 3′ hydroxyl group of the ATP ribose (Figure S2C).
Unfortunately, first-stage HAT treatments such as pentamidine and suramin have harsh side effects [3], and second-stage treatments such as melarsoprol can be fatal. The pharmaceutical industry has been slow to develop novel trypanocidal therapeutics because HAT infections occur primarily in developing countries with little market appeal; indeed, the only novel trypanocidal therapeutic registered in the last 50 years is eflornithine [43], a drug that is likely only available because it can also be sold as a topical cosmetic cream for the treatment of hirsutism in developed countries.
Given the hesitancy of the pharmaceutical industry, in recent years academia has played an increasing role in HAT drug-discovery efforts (e.g. [44]). Amaro et al. recently identified inhibitors based on a 4,5-dihydroxynaphthalene-2,7-disulfonate scaffold that target T. brucei RNA editing ligase 1 (TbREL1), a validated drug target in these organisms [12]. Unfortunately, these inhibitors, while effective against the TbREL1 protein, were ineffective in whole-cell assays. As Schrodinger's LigPrep software [45] suggested that at pH 7.0 the sulfonates of these compounds are negatively charged, we hypothesize that they are too hydrophilic to cross cellular and organellar T.-brucei lipid membranes and thus cannot reach their physiological target. The ALogP values of Amaro's S5, V1, and S1 compounds were −1.043, −0.292, and −0.778, respectively (Discovery Studio, Accelrys), likewise suggesting excessive hydrophilicity. Indeed, two of these three compounds, S5 and S1, are too hydrophilic to be considered druglike [46].
Building on the previous work of Amaro et al., we have developed additional TbREL1 inhibitors based on novel naphthalene scaffolds. The compounds proposed in the current work are also sulfonated naphthalenes; however, some of them are more hydrophobic than the naphthalene-based inhibitors identified previously. The ALogP values of V1, V2, V3, and V4 are 0.492, −1.039, −1.112, and 1.835, respectively (Discovery Studio, Accelrys), suggesting that two of the novel inhibitors, V1 and V4, may even prefer a lipid environment. Indeed, V4 was effective against cultured T. brucei with an EC50 of 2.16 µM (Table 1). To what extent this trypanocidal effect can be attributed to inhibition of REL1 is currently under investigation.
The hydrophobicity and specificity of these compounds, and their ability to reach the mitochondrial matrix, could be further improved by eliminating the charged sulfonate groups. In the virtual screen presented here, naphthalenes with carboxylic acids and nitro groups were included to see if the sulfonate groups could be replaced with less electronegative functional groups. Unfortunately, none of the compounds with carboxylate groups scored well enough to justify experimental testing, and the few nitro-group containing compounds that were tested failed to inhibit TbREL1. Rather than replacing the sulfonate groups, a better strategy may therefore be to modify those groups in order to neutralize their charge. For example, replacing the sulfonate groups with sulfonamides, a similar functional group that is not charged, may decrease hydrophilicity while preserving important protein-ligand interactions.
Both molecular docking and computational fragment mapping indicate that a new cleft revealed by the molecular dynamics simulations may play a role in the favorable binding of these four novel TbREL1 inhibitors. Furthermore, RMSD-based clustering indicated that this previously uncharacterized cleft persists for a majority of the MD trajectory.
In the future, further drug optimization is needed. Three of the four novel compounds contain diazene linkers that may be hydrolysable in vivo. Furthermore, the nitrogen atoms of these linkers are not predicted to participate in hydrogen bonds with the protein; replacing one or both of them with carbon atoms may therefore decrease hydrophilicity without sacrificing compound potency. Additionally, some of the compounds contain other moieties like hydroxyl and amino groups that are not predicted to contribute to inhibitor binding. Perhaps these groups could likewise be eliminated.
|
10.1371/journal.pntd.0003177 | QTL Mapping of Genome Regions Controlling Temephos Resistance in Larvae of the Mosquito Aedes aegypti | The mosquito Aedes aegypti is the principal vector of dengue and yellow fever flaviviruses. Temephos is an organophosphate insecticide used globally to suppress Ae. aegypti larval populations but resistance has evolved in many locations.
Quantitative Trait Loci (QTL) controlling temephos survival in Ae. aegypti larvae were mapped in a pair of F3 advanced intercross lines arising from temephos resistant parents from Solidaridad, México and temephos susceptible parents from Iquitos, Peru. Two sets of 200 F3 larvae were exposed to a discriminating dose of temephos and then dead larvae were collected and preserved for DNA isolation every two hours up to 16 hours. Larvae surviving longer than 16 hours were considered resistant. For QTL mapping, single nucleotide polymorphisms (SNPs) were identified at 23 single copy genes and 26 microsatellite loci of known physical positions in the Ae. aegypti genome. In both reciprocal crosses, Multiple Interval Mapping identified eleven QTL associated with time until death. In the Solidaridad×Iquitos (SLD×Iq) cross twelve were associated with survival but in the reciprocal IqxSLD cross, only six QTL were survival associated. Polymorphisms at acetylcholine esterase (AchE) loci 1 and 2 were not associated with either resistance phenotype suggesting that target site insensitivity is not an organophosphate resistance mechanism in this region of México.
Temephos resistance is under the control of many metabolic genes of small effect and dispersed throughout the Ae. aegypti genome.
| The mosquito Aedes aegypti is the principal vector of dengue and yellow fever flaviviruses. Due to a lack of effective drugs or vaccines, if an epidemic of dengue fever occurs in the near future, the first line of defense will involve the use of insecticides to suppress adult populations of Ae. aegypti. Unfortunately, the species has become resistant to most of the insecticides that can be safely applied. The authors have worked extensively on the mechanisms of resistance to the various insecticides commonly used for suppression of Ae. aegypti populations. Temephos is an organophosphate insecticide used globally to suppress Ae. aegypti larval populations but resistance has evolved in many locations. In this study we show that temephos resistance is under the control of many metabolic genes of small effect and dispersed throughout the Ae. aegypti genome. This information will be of general interest to field workers involved in the suppression of field populations of Ae. aegypti.
| Aedes aegypti is the principal vector of Dengue Fever (DENV) and Yellow Fever (YFV) flaviviruses throughout tropical and subtropical regions of the world and 2.5 billion people are at risk for DENV infection [1]. Currently DENV vaccines have low efficacy [2], [3] so that vector control remains the only option to reduce or prevent DENV transmission. Adult control depends largely on the use of pyrethroid insecticides. However, resistance to pyrethroids has been rising globally [4], [5], [6], [7], [8], [9]. More sustained control can potentially be achieved through the placement of insecticides in water containers that are known to harbor developing Ae. aegypti larvae in and around human habitations. For larval control, the three most widely used compounds are Bacillus thuringiensis israelensis (Bti), methoprene, and temephos. Globally, temephos is the most widely used of these three due to its very low vertebrate toxicity, relatively low cost, the fact that methoprene is a growth regulator with greatest effectiveness against older (third and fourth instar) larvae [10] and, because Bti must be ingested to be effective, it does not affect late larval or pupal stages when active feeding has ceased. Temephos is one of a few organophosphates registered to control Ae. aegypti larvae, and is the only organophosphate with any appreciable larvicidal use.
Temephos was first registered in the United States for mosquito control in 1965. It was quickly adopted as a larvicide because it was effective in polluted water, had a long residual activity, was available in several use-specific formulations, had a different mode of action than alternatives, and could be used on any larval instar. Temephos is toxic to many mosquito vector species that grow in a diversity of stagnant, saline, brackish and temporary water bodies. It remains an important management tool for mosquito abatement programs. The most widely used commercial preparation of temephos is Abate (EPA Registration No. 8329-60, Clarke Mosquito Control Products, Inc., Roselle, IL).
Temephos was used for 30 years before initial reports of resistance appeared in 1995. Initial studies reported less than a 5-fold resistance ratio (RR) in Ae. aegypti collections from Falcon and Aragua states of Venezuela [11]. In 1995, larvae from 34 strains of Ae. aegypti from 17 Caribbean countries were bioassayed and there were fairly high levels of temephos resistance in Tortola, British Virgin Islands (RR = 10–12) and Antigua (RR = 6–9) [12]. In 1999 a Tortola collection of Ae. aegypti was tested and a RR = 47 was identified [13]. After 13 generations of temephos laboratory selection, the RR increased to 181 fold [13]. Since 2000, temephos resistance has been reported from Cuba and Venezuela [14], [15], Thailand [16], the Brazilian states of Sao Paulo [17], Espirito Santo, Rio de Janeiro [18], Sergipe, Alagoas, [19], Ceara [20], and Paraiba [21]. Most recently reports have appeared from El Salvador [22], Martinique Island in the French West Indies [23], Argentina [24], [25], India [26], Colombia [27], and Trinidad [28], [29]. Although resistance to temephos has been demonstrated in many areas of the world, it is the only remaining organophosphate larvicide with any appreciable use. As such, it is an important tool in resistance management programs that depend on alternative larvicides. Alteration in the registration status or availability of temephos would have a large negative impact on our ability to control DENV transmission globally.
The purpose of the present study was to develop a better understanding of the genetics underlying temephos resistance in Ae. aegypti using QTL mapping in recently collected strains. A strain previously established from Solidaridad, Mexico was selected to have 290 fold higher temephos resistance than another strain that had been established from Iquitos, Peru. Parents from these two strains were reciprocally crossed to generate F1 siblings which were then intercrossed to generate an F2. The F2 generations were not large enough to assay for temephos resistance and so an F3 was generated through additional sib mating. F3 larvae were exposed to a discriminating dose of temephos and then checked every two hours up to 16 hours. Dead mosquitoes were preserved for DNA isolation at each time point and those surviving longer than 16 hours were considered resistant.
Two strains of Aedes aegypti were used. A F3 strain collected from Iquitos, Perú was kindly provided by Dr. Amy Morrison (University of California, Davis). A second strain raised during two generations in the lab was collected by the authors from the neighborhood of Solidaridad, in the city of Chetumal, in the state of Quintana Roo, México. Eggs were hatched in deoxygenated water from egg papers and then fed brewer's yeast. Adults were provided 10% (w/v) sucrose solution and were blood fed on citrated sheep blood in an artificial membrane feeder every three days. Incubators were set to a 14∶10 photoperiod, 30°C water temperature for larvae and 28°C for adult with a relative humidity of 85%.
F2 or F3 offspring from the field constituted the FS0 generation in the selection experiments. FS0 larvae were bioassayed to estimate the concentration of temephos (Chem Service, West Chester, PA) necessary to kill 50% of larvae (LC50). Bioassays were performed in plastic cups containing 100 ml of water with five different concentrations of temephos in 1 mL ethanol as a solvent. Approximately 25 3rd-instar larvae were gently pipetted into each cup. Mortality was recorded every 15 minutes up to two hours. All larvae were then transferred into clean water and mortality was scored at 24 hours. Each bioassay was performed in triplicate to obtain ∼75 larvae per concentration. LC50 and confidence limits were calculated using the IRMA quick calculator software (http://sourceforge.net/projects/irmaproj/files/Qcal/beta/QCal_ver_0.1_rev190.msi/download) which performs logistic regression [30].
Selection proceeded in three replicate lines for three generations. In the first round of selection 40–100 third instar larvae from each of the three replicates were exposed to an LC50 of 30 ng temephos/mL for two hours. Larvae were then transferred to clean water and mortality was recorded at 24 hours. Surviving larvae were transferred to 1 cubic foot rearing cages (BugDorm-1, Mega View Science, Co.) and raised to adults who were then blood fed to obtain FS1 eggs. We performed an initial bioassay with ∼75 larvae in each of the subsequent FS1–FS3 generations of selection to calculate the new LC50. From 40–100 larvae from each replicate were then exposed to the new LC50.
For the P1 mapping family, we crossed Solidaridad (SLD) FS3 and Iquitos (Iq) adults. Twenty P1♀SLD FS3×♂Iq and twenty reciprocal P1♀Iq×♂ SLD FS3 crosses were made. Larvae from each line were hatched and at the pupal stage, a female (larger size) from one strain was transferred to plastic cups in cardboard containers with a male pupa from the other strain. After adults emerged, they were allowed to mate for 3 days and the P1 male was frozen and held at −80°C. Females were blood fed three times with an artificial membrane feeder over the next ten days and the P1 female was then frozen and held at −80°C. Egg batches were maintained at room temperature for 7 days and then hatched by submersion in water followed by feeding them on Brewer's yeast ad libidum. For the F1 intercross families, one female and one male pupa from the same P1 family were allowed to emerge, mate and blood fed to eventually generate F2 progeny. F2 eggs from the largest F1 families were hatched and siblings were intercrossed in a single cage.
Third instar larvae (200 total) were exposed to 250 ng temephos/mL. After 2 hours, larvae that were unresponsive to prodding with a pipette tip were individually transferred to a labeled 1.5 mL microcentrifuge tube and frozen at −80°C. This was repeated every two hours for the next 16 hours. After 16 hours all remaining larvae were recorded as resistant.
The DNA of the P1 and F1 parents, and the two sets of 200 F3 offspring was individually isolated following the salt extraction method [31] and then suspended in 200 µl of TE buffer (10 mM Tris-HCl, 1 mM EDTA pH 8.0). The DNA was divided into 2–100 µl aliquots and stored at −80°C.
A total of 23 single copy genes [32], [33] and 26 microsatellite loci from [34] were amplified and analyzed. Each of these 49 genes has a known physical and linkage map position in the Ae. aegypti genome. A PCR mixture sufficient to perform 100 25-µl reactions was made by mixing 2,114 µL ddH2O, 250 µL 10×Taq buffer (500 mM KCl, 100 mM Tris-HCL pH 9.0), 25 µL of 20 mM dNTPs, and 2,500 pm of each of the primers. This reaction mixture was set under a UV light source (302 nm) for 10 min, after which 20 µl of Taq DNA polymerase was added. The mixture was then dispensed into a 96-well plate. Template DNA (∼100 ng) was then added to each well, followed by a drop of sterilized mineral oil. Each set of reactions was checked for contamination by the use of a negative control containing all reagents except template DNA. Samples were stored at 4°C before electrophoresis. The contents of each well were tested for the presence of amplified products by loading 5 µl from each well onto a 1.5% (w/v) agarose gel made with Tris-Borate-EDTA buffer. DNA fragments were size fractionated by electrophoresis for 15–20 min at 112 V. Fragments were visualized by staining with Syber Green and viewing the gel over a UV transilluminator. SSCP analysis and silver staining procedures were previously published [31].
Polymorphic SSCP-markers were sequenced in the four P1 and F1 parents to test for SNPs and to determine the inheritance patterns of SNP alleles. Sequences were aligned using CLUSTALW [35]. Allele specific primers were designed at those loci in which genotypes were fully or partially informative in the P1 and F1 parents. Design of primers for melting curve PCR is previously published [36]. Allele specific fragments were detected by melting curve PCR in a CFX-96 Real time PCR detection system (Bio-Rad, Hercules, CA). Table S1 provides previously unpublished oligonucleotide sequences for allele specific detection.
Associations between genotypes at each marker locus and hours until death (HTD) phenotype were initially assessed with ANOVA using summary (glm(HTD∼“Marker locus name”)) in R2.15.2 [37]. Our null hypothesis was that HTD was equal in each genotype. Associations between death (scored 0) or survival (1) (DOA) after 16 hours were initially assessed with Fisher's exact test (table (DOA, “Marker locus name”)) in R2.15.2. The null hypothesis was that the proportions of surviving larvae were equal in each genotype class. When the ANOVA or Fisher's exact test yielded a probability below 0.05, we examined the inheritance of the alleles at that locus. Our a priori hypothesis was that an excess of F3 individuals with an allele inherited from the SLD P1 parent would be resistant while an excess of F3 individuals with an allele inherited from the Iq P1 parent would die.
Multiple Interval mapping (MIM) [38] was then performed using QTL Cartographer 2.5 [39]. Two separate MIM were done. First, mosquitoes were scored as 2, 4, 6, 8, 10, 12, 14, 16 or 24 corresponding to hours until death. Second, F3 mosquitoes were scored as one if they survived to 16 hours or as zero if they died before 16 hours. In either case we created an initial model containing QTL map positions for markers at which ANOVA or Fisher's exact tests were significant. This model was then refined in MIM by 1) searching for new QTL, 2) estimating QTL effects, 3) obtaining and recording a summary, 4) optimizing QTL position, 5) searching for new QTL interactions, 6) testing for existing QTL main effects, 7) testing for existing QTL interaction effects, and 8) obtaining and recording a final summary. In addition, we used QTL Cartographer 2.5 to perform an initial MIM model selection on all markers using forward and backward selection with a significance level criterion of 0.01. We then compared this model with the model based upon markers identified as significant by ANOVA or Fisher's exact tests. The models agreed in all four cases: (1) ♀ SLD FS3×♂Iq –HTD (2) ♀ SLD FS3×♂Iq –DOA, (3) P1 ♀ Iq×♂ SLD – HTD and (4) P1 ♀ Iq×♂ SLD – DOA.
The concentration of temephos sufficient to kill 50% of larvae (LC50) was 50 ng temephos/mL water for the Iquitos strain. The Solidaridad FS0 strain initially had an LC50 of 27 ng temephos/mL water. Following three generations of temephos selection, the LC50 increased to 7.9 ug temephos/mL water in the Solidaridad strain. Thus the selected Solidaridad strain had ∼160 fold higher temephos resistance than the Iquitos strain. Among the SLD×Iq F3 larvae the LC50 was 6.5 ug temephos/mL water and was 1.9 ug temephos/mL water among the IqxSLD F3 larvae.
The genetic markers used in constructing maps in both the SLDxIq and IqxSLD crosses are listed along with their linkage positions in Table S2. Results of the ANOVA to test the null hypothesis that time until death is equal among genotypes are presented in Table 1. Results of Fisher's Exact Test on proportions of surviving larvae among genotype classes appear in Table 2. Loci with significant results are shown for all three chromosomes in Figure 1.
In the SLDxIq cross there were five QTL on chromosome 1 associated with HTD, four on chromosome 2 and four on chromosome 3. In the same cross there were four QTL on chromosome 1 associated with DOA, four on chromosome 2 and four on chromosome 3. In the IqxSLD cross there were three QTL on chromosome 1 associated with HTD, four on chromosome 2 and five on chromosome 3. There was one QTL on chromosome 1 associated with DOA, two on chromosome 2 and three on chromosome 3. The two families shared common QTL at loci 192TAAA1 and 88GAA1 on chromosome 1, at loci 462GA1 and 1132CT1 on chromosome 2 and at locus 86AC1 on chromosome 3. Between the two families there were six, six and nine QTL affecting HTD on chromosomes 1, 2, and 3, respectively or 21 loci in total. In the two families there were four, five and six QTL affecting DOA on chromosomes 1, 2, and 3, respectively or 15 loci in total.
When the ANOVA or Fisher's exact tests yielded a probability below 0.05, we examined the inheritance of the alleles at that locus. The last columns of Tables 1 and 2 indicate when the allele inherited from the SLD FS3 P1 parent were associated with resistance while the allele inherited from the Iq P1 parent was associated with susceptibility. Figure 2 plots HTD among larvae with the three possible genotypes. The first column of plots correspond to chromosomes 1, 2, and 3 in the SLDxIq cross. SLD alleles conferred slightly greater longevity for the first three marker loci on chromosome 1 but Aegi22 Iq homozygotes had greater longevity than heterozygotes (Fig. 2A). In contrast, SLD alleles confer greater longevity for all marker loci on chromosome 2 (Fig. 2B) and the effects appear to be additive. On chromosome 3, no general trend is evident (Fig. 2C). Iq homozygotes confer slightly greater longevity at marker loci 69TGA1 and para. The opposite trend is seen in markers 766ATT1 and 86AC1.
The second column in Figure 2 corresponds to chromosomes 1, 2, and 3 in the Iq×SLD cross. Again, SLD alleles confer slightly greater longevity on chromosome 1 (Fig. 2D). In contrast, on chromosome 2 SLD alleles at markers 328CTT1, 462GA1, and Arc4 confer only slightly greater longevity (Fig. 2E) while SLD alleles at the 1132CT1 locus appear to act as recessives in conferring much greater longevity. A similar pattern is seen in SLD alleles at 301ACG1 on chromosome 3 (Fig. 2F). However, Iq homozygotes confer slightly greater longevity at marker loci CCEae2D, vitg, 201TTA1 and Apyr1.
Figure 3 plots proportion surviving past 16 hours among larvae with the three possible genotypes. In the SLDxIq cross SLD alleles conferred greater survival at the first three marker loci on chromosome 1 but Aegi22 Iq homozygotes had greater longevity than heterozygotes (Fig. 3A). Note that these are the same markers as in Figure 2A, but with markers 192TAAA1, and 88GAA1. SLD alleles confer a 50% increase in survival. On chromosome 2 (Fig. 3B), with the exception of Arc4, SLD alleles at markers, 462GA1, Carbox and 1132CT1 all greatly increase survival. SLD alleles at 462GA1 appear to act additively in increasing survival from zero in Iq homozygotes to 50% in heterozygotes to 100% in SLD homozygotes. Resistant alleles at markers Carbox and 1132CT1 are recessive with 75–80% greater survival in SLD homozygotes. As with HTD, on chromosome 3 there is no general trend (Fig. 3C). Iq homozygotes confer slightly greater survival at marker loci 69TGA1 and para but the opposite trend is seen in markers 766ATT1 and 86AC1. In the Iq×SLD cross (Fig. 3D) SLD alleles at marker 88GAA1 increase survival by 50% and SLD alleles appear recessive. Similarly, alleles at the 1132CT1 marker increased survival by 90%. Identical patterns were seen in the SLDxIq cross (Fig. 3B). On chromosome 3, Iq homozygotes confer slightly greater survival at marker loci CCEae2D, vitg, and 86AC1.
The results of Multiple Interval Mapping with the HTD and DOA phenotypes are shown for both crosses in Table 3. Eleven QTL were identified in the SLD×Iq cross and these accounted for 68% of the phenotypic variance in HTD. There were nine QTL that accounted for 63% of the phenotypic variance in DOA. These nine were also all associated with HTD. The QTL that accounted for most (48%) of the genetic variation in HTD were at 47 cM and 70 cM on chromosome 2. The QTL that accounted for the most variation in DOA was at 62 cM on chromosome 2. QTL at 30 cM and 70 cM on chromosome 1 affected both phenotypes.
Genetic factors accounted for less of the variation in HTD and DOA phenotypes in the Iq×SLD cross. Eleven QTL were identified that accounted for 58% of the phenotypic variance in HTD. There were only two QTL that accounted for 31% of the variance in DOA and these were also associated with HTD. The QTL that accounted for most of the variation in HTD were at 57 cM on chromosome 1, 64 cM on chromosome 2 and 43 cM on chromosome 3. The only QTL that accounted for negligible variation in DOA was at 62 cM on chromosome 2. QTL at 57 cM on chromosome 1 affected both phenotypes. QTL at 30 and 57 cM on chromosome 1, and at 23.5 and 70 cM on chromosome 2 were common to both families
QTL mapping indicates that resistance to temephos is conditioned by many regions of the Ae. aegypti genome and therefore appears to behave as a classic quantitative genetic trait that is controlled by many loci each of minor effect. This pattern is supported by a recent parallel study in which we tracked changes in transcription of metabolic detoxification genes using the Ae. aegypti ‘Detox Chip’ microarray [40] during five generations of temephos selection [41]. We selected for temephos resistance in three replicates in each of six collections, five from México, and one from Peru. We used the esterase inhibitor DEF (S.S.S-tributylphosphorotrithioate) to show that esterases were the major metabolic source of resistance. However, the microarray data indicated that expression of many esterase genes increased with selection and that no single esterase was consistently upregulated among the six selected lines.
Target site resistance in acetylcholine esterase genes is a very common mechanism of resistance to organophosphate and carbamate insecticides [42]. We therefore tested for a significant genotype -phenotype interaction with SNPs in the AChE-2 gene (AAEL012141) at 40.7 cm on chromosome 1 and the AChE-1 gene (EF209048) at 3p1.2 (30.4 cM) on chromosome 3 [43]. Results in Table 1–3 show that no significant associations were detected. Similar studies of temephos resistance in field populations of Ae. aegypti also failed to detect insensitive acetylcholine esterase [44] despite the fact that these authors were able to generate recombinant clones that produced Ae. aegypti insensitive acetylcholine esterases in the laboratory [45]. Another possibility is that temephos in particular fails to select for insensitive acetylcholine esterases. Cuban investigators were able to select Ae. aegypti with 13-fold increase in insensitive acetylcholine esterase but using the carbamate insecticide propoxur [46].
Previous studies of esterase isozyme loci identified two genetically mapped loci associated with resistance to the organophosphate insecticide malathion. Elevated activity staining of Esterase-5 located at 57 cM at the base of Chromosome 1 [47] was reported [48]. This may correspond to the 57 cM QTL on chromosome 1 associated with marker 88GAA1 in both families in the current study. Similarly elevated activity staining of Esterase-6 located at 83 cM at the base of Chromosome 2 in the map of [47] was reported [49], [50]. This may correspond to the QTL at 70 cM on chromosome 2 associated with marker 1132CT1 found in both families in the current study. We have no means to formally check these associations because neither the nucleotide nor amino acid sequences of proteins Esterase-5 and 6 are known.
There are 49 currently identified carboxy/choline esterase genes [40]. With the recent publication of a physical map that contains 45% of the Ae. Aegypti genome [51], [52] we had hoped to learn the physical locations of many of these esterases. However, other than AChE-1 and AChE-2, there were only six other esterase genes that occurred in mapped supercontigs. These were CCEbe2o (AAEL008757) on 2p3.4 (also mapped in the present study see Figure 1), CCEjhe2o (AAEL004323) on 2q2.4, and four (CCEjhe1F (AAEL005200), CCEjhe2F (AAEL005198), CCEjhe3F (AAEL005210), and CCEjhe4F (AAEL005182)) all located in supercontig 1.145 at 2p4.4. Whether these four are associated with the QTL at 5.8 cM on the top of Chromosome 2 in the Qi×SLD cross (see Tables 1–2) is unknown at this time.
Even though the selected Solidaridad strain had overall ∼160 fold higher temephos resistance than the Iquitos strain, this pattern wasn't uniform across the entire genome. This could have affected the locations and relative contributions of QTL. There are many instances in Tables 1 and 2 wherein the mosquitoes homozygous for markers from the “susceptible” Iquitos strain were more resistant than heterozygotes or homozygous for markers from the “resistant” SLD strain (note especially the bottom of chromosome 3 for both HTD and DOA). This counterintuitive outcome is probably a result of using Iquitos mosquitoes taken directly from the field without selecting for a more susceptible phenotype. However, it could also be associated with negative fitness effects associated with resistance alleles in the SLD strain that became concentrated during selection.
In our previous QTL mapping study [36] we found resistance to permethrin to be principally (91.8% of genetic effect in MIM) under the control of target site insensitivity in the voltage gated sodium channel gene (orthologue of paralysis in Drosophila [53]). We have shown that the genetic architecture underlying temephos resistance to be completely different with both families having up to 11 QTL affecting the HTD phenotype in both families and from 2–9 QTL affecting DOA. The practical implications of these findings are that selection for temephos resistance in the field is likely to involve many (principally esterase) loci. It is unlikely that the same genes will be involved in all field populations and that genetic drift may play a large part in determining which combinations of the 49 currently identified carboxy/choline esterase genes [40] become upregulated and assume responsibility for metabolic detoxification of temephos.
|
10.1371/journal.pgen.1002000 | Targeted Sister Chromatid Cohesion by Sir2 | The protein complex known as cohesin binds pericentric regions and other sites of eukaryotic genomes to mediate cohesion of sister chromatids. In budding yeast Saccharomyces cerevisiae, cohesin also binds silent chromatin, a repressive chromatin structure that functionally resembles heterochromatin of higher eukaryotes. We developed a protein-targeting assay to investigate the mechanistic basis for cohesion of silent chromatin domains. Individual silencing factors were tethered to sites where pairing of sister chromatids could be evaluated by fluorescence microscopy. We report that the evolutionarily conserved Sir2 histone deacetylase, an essential silent chromatin component, was both necessary and sufficient for cohesion. The cohesin genes were required, but the Sir2 deacetylase activity and other silencing factors were not. Binding of cohesin to silent chromatin was achieved with a small carboxyl terminal fragment of Sir2. Taken together, these data define a unique role for Sir2 in cohesion of silent chromatin that is distinct from the enzyme's role as a histone deacetylase.
| Replication of chromosomes in each cell cycle produces pairs of identical sister chromatids that are held together by a protein complex known as cohesin. At mitosis, cohesin is dismantled, permitting segregation of one full set of chromosomes to each daughter cell. Cohesin binds at discrete sites along chromatids, including domains that are commonly referred to as silent chromatin in budding yeast. Silent chromatin, like heterochromatin in higher eukaryotes, is a repressive structure that hinders a variety of DNA-based events. We discovered that a single silent chromatin constituent, Sir2, was both necessary and sufficient for cohesion of silent chromatin domains. Sir2 is the founding member of the sirtuin family of NAD-dependent protein deacetylases that exist in most organisms. Substrate deacetylation by sirtuins has been linked to numerous pathways that promote health and survival in humans, including lifespan extension. Enrichment of cohesin at silent chromatin domains in yeast, however, is the first example of a role for Sir2 that does not explicitly require the protein deacetylase activity.
| Proper segregation of chromosomes at mitosis and meiosis requires sister chromatid cohesion. The process ensures that newly replicated chromatids bi-orient on spindle microtubules such that a single copy of each chromosome transfers to progeny cells. Defects in the sister chromatid cohesion pathway lead to certain developmental diseases, and chromosome segregation defects like those seen in cancer [1]–[4].
Cohesion of sister chromatids is mediated by a protein complex known as cohesin [5], [6]. The core complex consists of a hetero-dimer of SMC proteins Smc1 and Smc3, as well as non-SMC proteins Scc3/Irr1 and Mcd1/Scc1/Rad21 (hereafter referred to as Scc3 and Mcd1, respectively). The subunits form a large protein ring with a striking central void. Thus, a prominently held view is that cohesin holds sister chromatids together by single complexes embracing both chromatids. Elegant protein-crosslinking studies showed that single cohesin rings can indeed hold together two partially purified minichromosomes [7]. Other data raises the possibility that cohesin might hold sister chromatids together by a different mechanism [8]–[10].
Cohesin binds discrete sites on chromosomal DNA. Most non-centromeric sites in budding and fission yeasts lie within the AT-rich regions between convergently transcribed genes [11]–[13]. Transcriptional elongation redistributes complexes from intragenic to intergenic regions, suggesting that cohesin enrichment is maintained dynamically. In contrast to the situation in these fungal systems, cohesin maps along the lengths of actively transcribed genes in Drosophila and to sites within transcribed genes in humans [14]–[16]. Thus, cohesin binding and transcription are not always mutually exclusive.
Cohesin is also found within pericentric heterochromatin regions where transcription is suppressed but not extinguished. In fission yeast, the complex is retained at these locations by Swi6, a homolog of heterochromatin protein HP1, which interacts with cohesin subunit Psc3 (Scc3 in budding yeast) [17], [18]. During meiosis, Swi6 also interacts with shugoshin, a protein that protects centromeric cohesin from being dismantled [19]. In heterochromatin mutants, cohesin does not bind pericentric domains and mitotic chromosomes fail to mount properly onto spindle microtubules.
Budding yeast lacks Swi6 and pericentric heterochromatin but it does contain transcriptionally silenced domains that nevertheless bind cohesin. Using the HMR locus as one representative example, we found that silencing mutations selectively disrupted cohesin binding and correspondingly abolished cohesion of sister chromatid DNA bearing the locus [9]. A search to understand why cohesin accumulates at HMR served as the impetus for this study.
Based on the chromatin-mediated mechanism of regional DNA inactivation, transcriptionally silenced domains in budding yeast are referred to as silent chromatin [20]. Like heterochromatin domains in other organisms, silent chromatin is packaged with histones that bear a distinct signature of post-translational modifications. Specifically, acetylation and methylation of lysines are absent. Silent chromatin domains associate with a complex of non-histone silencing factors known as the Sir proteins (Sir2, Sir3 and Sir4). Sir2 is a member of the evolutionarily conserved class of NAD+-dependent protein deacetylases known as sirtuins. The enzyme creates and maintains histone deacetylation within silent chromatin. Sir3 and Sir4 associate with the suitably deacetylated histones. The complex of Sir proteins is first recruited to sites of action by cis-acting elements known as silencers, which bind ORC, as well as Abf1 and Rap1 in various combinations. Following recruitment, cycles of histone deacetylation and histone binding allow the Sir proteins to spread over kilobases. A tRNA gene acts as a barrier element on the right side of HMR that blocks silent chromatin from spreading further downstream [21]. The element also augments HMR with sufficient cohesin for cohesion [22], probably through recruitment of the Scc2/4 cohesin loading complex [23], [24].
We considered two competing hypotheses to account for retention of cohesin at HMR. The first, based on a simple recruitment model, posits that a silent chromatin component interacts directly with cohesin or some factor associated with the complex. A second hypothesis stems from the ability of silent chromatin to impede a broad-range of DNA-based events, such as DNA replication, repair and transcription [20]. If silent chromatin also suppresses an activity that mobilizes cohesin, the complex would accumulate at silenced loci. To distinguish between these possibilities, we developed assays to determine whether silencing or silent chromatin components were required for cohesion of HMR. Our studies show that Sir2 is sufficient for cohesion, even in the absence of silencing.
Our principle assay for cohesion at HMR utilizes a strain in which the locus is tagged with lac-GFP and flanked by target sites for a site-specific recombinase [9]. Inducible excision after arrest in M phase converts HMR loci on sister chromatids into a pair of extrachromosomal circles that produce one bright fluorescent focus if they are held together and two foci if they are not (Figure 1A and 1B).
To test whether silent chromatin components can mediate cohesion we tethered individual silencing factors directly to the DNA circles (Figure 1C). To this end, the E silencer of HMR was replaced with a synthetic construct (6lexopssEB) that includes binding sites for Rap1, Abf1 and the bacterial protein lexA. The I silencer was deleted. These modifications were previously shown to eliminate silencing of the locus [25]. Individual silencing factors were then targeted to HMR-6lexopssEB as lexA-linked fusion proteins. Cell cycle arrest in M phase, recombinase induction and fluorescence microscopy were performed as described previously [9].
Tethering silent chromatin components to DNA often nucleates silent chromatin assembly and restores transcriptional repression [25], [26]. In these situations, it would be impossible to determine whether the tethered protein, a co-recruited protein or the silenced state was responsible for cohesion. Therefore, tethered proteins were also examined under conditions that abolish silent chromatin assembly to evaluate their precise roles in cohesion.
Pilot experiments showed that excised circles bearing the HMR-6lexopssEB construct colocalized infrequently [9]. When lexA was expressed, only 22% of the nuclei contained the single bright fluorescent spot (Figure 1D). Strikingly, cohesion of the circles increased to 67% when lexA was fused to Sir2 (designated lexA-Sir278–562). Tethering Sir2 to DNA was essential. In a strain lacking lexA binding sites at HMR, the chimera failed to produce cohesion (Figure S1).
LexA-Sir278–562 lacks the first 77 amino acids of Sir2 that are dispensable for transcriptional repression [27]. We confirmed that lexA-Sir278–562 nucleates silencing at HMR using a strain that contains lexA binding sites and a TRP1 reporter gene at the locus (Figure 1E). Taken together, these initial findings demonstrate that tethered Sir2 confers both silencing and cohesion at HMR.
The Sir2 polypeptide consists of a conserved catalytic core, as well as N and C terminal domains that help target the deacetylase to sites of action [28]. An allele lacking the N-terminal 198 amino acids confers little transcriptional repression, even when tethered to DNA [29]. To generate a lexA chimera with similar characteristics, we eliminated the entire N-terminal domain (amino acids 1–242). Surprisingly, this construct (lexA-Sir2243–562) yielded a measurable degree of silencing in a strain with intact SIR genes (for comparisons, see lexA-Sir278–562 and lexA alone in figure 2A). Deletion of either the SIR2 or SIR4 genes eliminated silencing by lexA-Sir2243–562, indicating that 1) the chimera operates within the Sir pathway, and that 2) the chimera requires the endogenous full-length Sir2 for transcriptional repression. The reliance of lexA-Sir2243–562 on other SIR genes, including SIR2, for silencing made the chimera an ideal candidate for further study.
Figure 2B shows that lexA-Sir2243–562 produced cohesion in over 60% of the nuclei examined. Importantly, cohesion of the excised HMR circles persisted in strains that lacked SIR2, SIR3 or SIR4. We conclude that tethered Sir2 can mediate cohesion in the absence of transcriptional silencing and without the aid of endogenous Sir proteins.
Sir3 was also examined directly with the targeted cohesion assay. When the protein was linked to lexA, HMR circles colocalized in over 60% of wild-type cells (Figure 2C). The tethered protein also conferred transcriptional repression in the wild-type reporter strain (Figure 2A). Both cohesion and silencing by Sir3-lexA were significantly diminished by deletion of Sir2. Elimination of Sir4, on the other hand, disrupted silencing but not cohesion. A simple explanation for the requirement of Sir2 but not Sir4 is that tethered Sir3 recruits Sir2, which in turn mediates cohesion of the locus.
We note that in the absence of Sir2, Sir3-lexA yielded a slightly higher level of cohesion than lexA alone (Sir3-lexA = 34% vs lexA = 26.3%). This difference is sufficiently small (p = 0.03) that we cannot conclude equivocally whether Sir3 possesses a subtle intrinsic cohesion activity. Given the strong cohesion signals afforded by Sir2, we focused our attention on this Sir protein for the remainder of the study.
If Sir2 mediates cohesion at HMR then the protein ought to impart cohesion when tethered at other genomic positions. We explored this possibility by targeting the protein to the LYS2 locus. LYS2 is situated near the center of chromosome II, hundreds of kilobases away from silent chromatin domains at the chromosome ends [30]. The locus had previously been modified to contain lexA-binding sites, as well as lac repressor sites and recombinase sites for the DNA excision assay [31]. When lexA alone was expressed, LYS2 DNA circles colocalized in 37% of the cells (Figure 3). This value is higher than the baseline for HMR cohesion under similar conditions (Figure 1D). The value is sufficiently low, however, to detect increases in cohesion due to tethered Sir2 fragments. Indeed, cohesion of LYS2 circles increased to 60% when lexA-Sir2243–562 was expressed. Pairing persisted in a strain lacking SIR3 indicating that cohesion was due to the tethered protein and not due to formation of silent chromatin at LYS2. These findings indicate that Sir2 can impart cohesion at chromosomal locations other than HMR.
We next asked whether the deacetylase activity of Sir2 was responsible for Sir2-mediated cohesion. To address this question, we introduced a well-characterized active site mutation (H364Y) into lexA-Sir2243–562. Previous studies showed that this mutation abolishes Sir2 deacetylase activity, silencing and silent chromatin formation [32], [33]. We found here that the mutated polypeptide conferred as much cohesion to HMR circles as the unaltered polypeptide (Figure 4A). This experiment was performed in a sir2 null strain to eliminate contributions of the endogenous gene (see figure 2A). Furthermore, acting on the remote possibility that tethered Sir2 mediates cohesion by recruiting one of the other yeast sirtuins (Hst1-4), we repeated the experiment in media supplemented with nicotinamide, a generic sirtuin inhibitor [28]. No decrease in cohesion was observed (64% vs. 61% with nicotinamide; p = 0.5). Collectively, these results show that the enzymatic activity of Sir2 or other sirtuins is not required for cohesion by tethered Sir2.
To map the Sir2 domain responsible for cohesion we generated a set of truncation mutants. Figure 4A shows that all but one of the constructs yielded HMR cohesion levels significantly above background. The exception is a lexA chimera that bears just the conserved catalytic core of Sir2 (residues 243–499). All of the other cohesion-proficient chimeras share in common a small C-terminal domain of Sir2 spanning amino acids 525–547. The picture that emerges is that Sir2 contains a discrete motif within the non-catalytic, C-terminal region of the protein that mediates cohesion.
Hst1 is a yeast sirtuin that bears considerable amino acid similarity to Sir2 in the C-terminal region (Figure 4B). The deacetylase represses middle sporulation genes in vegetative cells, as well as genes involved in NAD+ and thiamine biosynthesis [34]–[36]. Hst1 differs from Sir2 in that the protein acts locally to repress specific promoters rather than by forming an extended repressive domain [37]. We tethered the C-terminal domain (amino acids 440–503) to HMR in the same Δsir2 strain used to evaluate the lexA-Sir2 constructs. Figure 4A shows that lexA-Hst1440–503 imparts a comparable degree of targeted cohesion. These results indicate that Sir2-mediated cohesion is not limited to just one member of the sirtuin family.
Both Sir2 and its yeast paralog Hst2 can form homotrimers [38], [39]. Thus, one explanation for DNA colocalization is that tethered Sir2 fragments on different DNAs associate with one another directly. To explore this possibility we performed a two-hybrid analysis using lexA-Sir2243–562(H364Y) as a bait protein. The experiments utilized a HIS3 reporter strain that lacks the endogenous SIR2 gene. A weak positive interaction signal was obtained with a prey vector bearing full length Sir2 fused to the Gal4 activation domain (Figure 5A). Importantly, no interaction was seen with a prey vector bearing the shorter Sir2243–562 fragment. Given that all of our critical experiments were performed with this fragment in strains lacking full length Sir2, colocalization of HMR circles is not likely attributable to Sir2 self-association.
Cohesin mediates cohesion of the native HMR locus [9]. We therefore anticipated that cohesin genes would also be required for cohesion by tethered Sir2. To test this possibility we crossed temperature sensitive alleles of MCD1/SCC1 and SMC3 into our DNA circle-producing strain. The scc1-73 and smc3-42 mutants and a wild-type counterpart were arrested in mitosis at permissive temperature (24°C). After recombining the HMR locus, the cultures were divided: half was maintained at the permissive temperature while the other half was shifted to 37°C, the non-permissive temperature for these mutants (see Figure 5 legend for details). In the wild-type strain, cohesion of the HMR circles by lexA-Sir2243–562 was unaffected by the temperature shift (Figure 5B). By contrast, both mutant strains displayed a significant reduction in HMR cohesion at the non-permissive temperature. This data indicates that cohesin is responsible for cohesion of DNA circles bound by Sir2.
Chromatin immunoprecipitation (ChIP) was used previously to show that cohesin associates with HMR in a silencing-dependent manner [9]. We showed that Mcd1-TAP binding was lost when silent chromatin assembly was blocked by 1) deletion of SIR3, or 2) inhibition of the Sir2 deacetylase (see ChIPs of chromosomal templates in figures 5A and 7D of [9]). In the current study, a similar ChIP protocol was used to test whether lexA-Sir2243–562 retained cohesin at HMR-6lexopssEB. Unexpectedly, we could not obtain reproducible enrichment of the targeted locus. A variety of conditions and reagents were tested, and the procedure was validated with native silent chromatin (see below). We suspect that the level of cohesin necessary for colocalization in the targeted cohesion assay falls below the detection limit of this ChIP experiment.
We turned instead to a protein chimera approach we recently developed to study other aspects of transcriptionally silent chromatin [40]. In that study, Sir3 was fused to the Rpd3-family deacetylase Hos3 to show that the roles of Sir2 in silencing could be bypassed entirely. We demonstrated that the Sir3-Hos32–549 chimera 1) spread throughout HMR, 2) deacetylated histones across the locus and 3) required both silencers and Sir4 to mediate repression. Silencing could also be achieved by fusing Sir3 to a fragment of Sir2 that possessed enzyme activity but that lacked domains necessary for targeting. Here these Sir3 chimeras (Sir3-Sir2243–562 and Sir3-Hos32–549) were used to investigate the role for Sir2 in binding cohesin at a silenced domain. An additional chimera (Sir3-Hos32–549-Sir2499–562) was constructed to study the contribution of the 64 amino acid, cohesion-proficient fragment of Sir2.
Mating assays were used to evaluate the silencing potential of each chimera. In this assay, loss of HMR silencing in MATα cells creates a pseudo-diploid state that blocks mating and thus subsequent growth on SD indicator plates. Figure 6A confirms previous findings that Sir3-Sir2243–562 and Sir3-Hos32–549 mediate silencing of HMR in the absence of endogenous Sir2. The figure shows that Sir3-Hos32–549-Sir2499–562 also conferred silencing of HMR, albeit at a reproducibly reduced level. This functional assay indicates that Sir3-Hos32–549-Sir2499–562 delivers the Sir2499–562 fragment to the site where cohesion and cohesin binding were to be tested.
Cohesion by the Sir3 chimeras was evaluated in a sir2 null strain that produces GFP-tagged HMR circles with wild-type silencers (Figure 6B). In the absence of a chimera, HMR cohesion occurred in 33% of the cells. When Sir2 or the Sir3-Sir2243–562 was expressed, HMR cohesion levels rose to 69% and 61%, respectively. By contrast, expression of Sir3-Hos32–549 did not increase cohesion above background levels. Remarkably, addition of Sir2499–562 to the Sir3- Hos32–549 chimera restored colocalization to the level obtained with Sir3- Sir2243–562. This analysis indicates that the C-terminal fragment of Sir2 must be present within silenced chromatin for cohesion to occur.
ChIP of TAP-tagged Mcd1 was used to measure the ability of the chimeras to position cohesin at the HMR a2 gene. A cohesin-associated region of chromosome V (designated 549.7) that is not influenced by the SIR genes was used as a point of comparison [22]. Reference SIR2 and Δsir2 strains in figure 6C confirmed earlier findings: binding of Mcd1-TAP at HMR is hindered when silent chromatin is disrupted by loss of a single Sir protein, in this case Sir2. Expression of the Sir3-Sir2243–562 chimera restored cohesin binding at HMR to within 20% of the native level. By contrast, expression of the silencing-proficient Sir3-Hos32–549 chimera did not raise cohesin binding above the background sir2 null level. Importantly, the addition of 64 amino acids of Sir2 to the end of the Sir3-Hos32–549 chimera increased cohesin binding substantially. Taken together, these data indicate that Sir2 must be present within silent chromatin for cohesin to accumulate at silenced loci, and that a small C-terminal portion of Sir2 is sufficient for this activity.
Sir2 associates with the cluster of tandemly repeated ribosomal RNA genes known as the rDNA array. In this context the protein suppresses recombination between the repeated elements and suppresses RNA polymerase II transcription within each element [41]–[43]. Sir2 has been implicated in cohesin binding at the rDNA [44], [45]. It therefore seemed prudent to test whether rDNA-specific, protein partners of Sir2 modulate cohesion by the tethered protein. We first considered Net1, which along with Sir2 and Cdc14 forms the RENT complex [46], [47]. This protein is required for Sir2 binding at the rDNA and it has been found at HMR when over-expressed [46], [48]. A 15 amino acid C-terminal truncation of Sir2 disrupts the Net1-Sir2 interaction, abolishing rDNA silencing but not silencing of telomeres or the HM loci [49]. Figure 4A shows that deleting these 15 residues (lexA-Sir2243–547) did not interfere substantially with cohesion of HMR circles. We conclude that the RENT complex is not necessary for cohesion by tethered Sir2.
Transcriptional silencing by Sir2 at the rDNA recombinational enhancer requires a set of interacting proteins that includes Tof2 and a pair of bifunctional factors Csm1 and Lrs4. During meiosis I, Csm1 and Lrs4 form the monopolin complex that orients sister chromatid pairs towards the same spindle pole [50], [51]. Csm1 interacts with both Mcd1 and Smc1 prompting Huang and Moazed to hypothesize that these proteins link cohesin to the rDNA [52]. We tested whether these genes were required for cohesion by lexA-Sir2243–562. Figure 7 shows that neither TOF2, CSM1 nor LRS4 were required for colocalization of HMR circles. Collectively, the findings indicate that these rDNA silencing and stability proteins do not contribute to cohesion of HMR by tethered Sir2.
In this study we examined the mechanistic basis for Sir-dependent cohesion of a silenced chromosomal domain in budding yeast. We developed a protein-targeting assay and found that the evolutionarily conserved Sir2 deacetylase was both necessary and sufficient for pairing DNA circles. Cohesin was required but other silencing factors like Sir3 and Sir4 were not. Through the use of mutants we showed that transcriptional silencing and cohesion are separable events: tethered Sir2 conferred cohesion in the absence of silencing and the Sir3-Hos3 chimera generated silencing in the absence of cohesion. Importantly, fusing a small fragment of Sir2 to Sir3-Hos3 was sufficient to restore cohesin binding and cohesion of the locus. We conclude that, in addition to deacetylating histones for silent chromatin assembly, Sir2 also orchestrates cohesin-dependent cohesion of silent chromatin domains on sister chromatids.
Although our studies here focused on HMR we expect that the relationship between Sir2 and cohesion extends to other loci where Sir proteins assemble. Indeed, preliminary evidence indicates that the HML mating-type locus is also cohered in a silencing-dependent manner (Campor and Gartenberg, unpublished results). Why do silent chromatin and cohesion converge? Initial studies suggested a role in regulating transcriptional repression. Donze and Kamakaka first showed that silencing spread beyond HMR barrier elements in cohesin mutants [21]. Steve Bell and colleagues followed by showing that cohesin delayed establishment of silencing in cells that were walked step-wise through the cell cycle [53]. A parsimonious explanation for these observations is that cohesin impedes the conversion of active chromatin to silenced chromatin. Numerous studies in higher eukaryotes have further linked cohesin to gene regulatory phenomena (see [54] for a review). Intriguingly, cohesin was recently shown to form loops between enhancers and promoters by interacting with a transcriptional coactivation complex known as mediator [55]. Similarly, cohesin forms loops between distant sites by binding the mammalian CTCF, a protein that associates with insulators as well as other gene regulatory elements [56]–[59]. In yeast, silent chromatin domains fold-back upon themselves and interact with one another over great distances [60], [61]. Thus, one possibility is that cohesin facilitates long interactions to regulate silent chromatin domains.
The rationale for Sir2 mediating cohesion might alternatively be related to its role in genome stabilization at the rDNA. Binding of the deacetylase is necessary for binding of cohesin, which in turn is thought to block unequal sister chromatid exchange by maintaining the register between rDNA elements of opposing sister chromatids [44]. Exactly how Sir2 retains cohesin at the locus in not entirely clear. In one model, the deacetylase modulates cohesin levels indirectly by silencing a conserved RNA polymerase II promoter element near the rDNA recombinational enhancer [45]. According to the model, transcription by RNA polymerase II displaces cohesin when Sir2 is absent. A competing model by Huang and Moazed suggests that direct interaction between cohesin and one of the components of the rDNA silencing pathway, Csm1 specifically, could account for recruitment of the complex [52]. Our work with tethered fragments of Sir2 suggests an even more direct possibility: the polypeptide, not its capacity to silence, mediates cohesin recruitment directly. We note that a direct recruitment model for Sir2 need not be mutually exclusive with models based on transcriptional inhibition, or with other factors that contribute to rDNA stability [62].
Acetylation and deacetylation of cohesin subunits plays a newly appreciated role in regulating cohesion during the cell cycle. Cohesion is established during S phase when the Eco1 protein acetyltransferase acetylates Smc3 [63]–[66]. This modification persists until cohesin complexes disassemble at anaphase onset. Deacetylation is a prerequisite for Smc3 to be re-used in the next cell cycle. Recently, the Rpd3-family member Hos1 was identified as the principle Smc3 deacetylase in yeast [67]–[69]. That residual deacetylation persists in the absence of Hos1 suggests that additional Smc3 deacetylation activities remain to be discovered [67]. Following DNA double strand breaks, Eco1 similarly acetylates Mcd1 to establish damage-induced cohesion [70]. Presumably, a parallel pathway exists for Mcd1 deacetylation. Whether Sir2 or a combination of sirtuins is involved in Mcd1 deacetylation or the residual deacetylation of Smc3 has not been determined.
The catalytic activity of Sir2 accounts for all other known functions of the enzyme. By contrast, cohesion by tethered Sir2 fragments does not even require the conserved catalytic core. Instead, we found that a small domain at the carboxyl-terminus was responsible. We anticipate that this domain retains cohesin at silenced loci by interacting directly with a cohesin subunit or with other proteins involved in cohesin utilization. Conversely, such an interaction could be important in situations where Sir2 might be recruited to sites where cohesin binds.
Significant homology exists between the C-terminal domain of Sir2 and Hst1. That this Hst1 domain also mediates cohesion when tethered to DNA suggests that cohesion occurs at the numerous promoters where Hst1 binds to regulate gene expression [34]–[36]. The lack of a homologous C-terminal domain in mammalian sirtuins thwarts a simple extrapolation to a cohesion connection in higher eukaryotes. However, the characteristics of two mammalian sirtuins, SirT1 and SirT6, warrant consideration (see [71], [72], and references therein). Like Sir2, these mammalian enzymes deacetylate histones (and other protein targets) to regulate gene expression. Additionally, SirT1 plays multiple roles in heterochromatic repression and SirT6 localizes to heterochromatin domains. Double strand breaks may represent sites of particular interest. Cohesin is recruited to these sites in yeast and in humans, as are Sir2, Hst1, SirT1 and SirT6 [73], [74]. The mammalian enzymes have been shown to suppress genomic instability, in part, by modifying DNA repair factors (see [75] for most recent example). Whether SirT1 and SirT6 also link sister chromatid cohesion to these chromosome-based events has not yet been tested.
Table S1 provides a complete list of strains used in this study. Cohesion of HMR by tethered proteins was measured with variants of strain CSW19 (RS::6lexopssEB-a2a1-256lacop-TRP1-ΔhmrI::RS (LEU2::GAL1-R)2::leu2-3,112 ADE2::HIS3p-lacGFP::ade2-1). Recombinase target sites are designated as RS. Cohesion of the LYS2 gene was measured with variants of strain CSW91 (RS::lys2-TRP1-4lexop-256lacop::RS). Cohesion of HMR by Sir3 chimeras was measured with native silencers in strain CSW84 (RS::HMRE-a2a1-HMRI-TRP1-256lacop::RS Δsir2). Silencing assays were performed with strains derived from GA-2050 (Aeb4lexop-TRP1-HMRI) and two hybrid assays were performed with strain JCY13 (LYS2::(4xlexop)-HIS3 Δsir2). ChIP assays were performed in variants of strain CSW116 (MCD1-TAP Δsir2).
Complete ORF deletions were generated by PCR-mediated gene replacement using purified plasmids or extracted yeast DNA as PCR templates. All modifications were confirmed by combined gain and loss of diagnostic PCR products. Strains CSW18 and CSW19 are segregants of a cross between CSW10 and YCL49. Strains CSW47 and CSW48 are segregants of crosses between CSW36 and either K5832 or CRC85. Strain CSW91 is a segregant of a cross between CSW19 and GA-2627. Plasmids pRS403-lexA-Sir278–562 and pRS403-lexA were integrated in single copy at the HIS3 locus of CSW19 to yield strains CSW36 and CSW37, respectively.
Tables S2–S3 provide detailed information about the plasmids used and how they were constructed. In addition to using traditional bacterial cloning techniques, plasmids were constructed in yeast by PCR-mediated plasmid gap repair (P-MPGR) or fragment-mediated plasmid gap repair (F-MGPR) using restriction digestion products. SIR2 truncations were generated by oligonucleotide-mediated plasmid gap repair (O-MPGR). All modifications within the gene chimeras were confirmed by sequencing. Plasmid sequences are available upon request.
Colocalization of excised DNA circles in M phase was measured as described in Chang et al. [9], unless specified otherwise. To retain plasmids, selective media was used for pregrowth on dextrose and subsequent growth on raffinose overnight. When the cultures reached mid-log phase the following morning, an equal volume of YPA media plus raffinose was added. Twenty minutes later, nocodazole (stock 1 mg/ml in DMSO, Cf = 10 µg/ml) was added to initiate M phase arrest. After three hours, galactose (Cf = 2%) and benomyl (stock 1 mg/ml, Cf = 10 µg/ml) were added. Two hours later, cells were harvested, fixed and mounted on microscope slides with agar pads. Serial sections were obtained by fluorescence microscopy and GFP-foci/nucleus were counted manually. All measurements (reported as the percentage of cells with single dots) are based on at least three independent trials, which were pooled because they satisfied χ2 tests of homogeneity of proportions. Error bars represent the standard error of proportion. In each data panel, values were compared to an appropriate control by χ2 tests and judged as significant using a 95% confidence interval.
To measure silencing of TRP1 inserted at HMR or two-hybrid interactions with the HIS3 reporter construct, plasmid-bearing strains were grown to saturation in selective media to retain plasmids and spotted in 10-fold serial dilutions. One set of selective plates was used to measure reporter gene expression and a second set was used as a loading control.
Strain YFC9 (MATα Δsir2) bearing Sir2-substitution plasmids was grown to saturation in selective medium, diluted 10-fold and then spotted on a lawn of mating tester strain K125 (MATa) on YPDA plates. After at least 5 hr at 30°C, the cells were replica plated to SD agar to measure mating and SC-trp as a loading control.
Nocodazole was added to mid-log cultures that were either grown in YPDA overnight or that were sub-cultured in YPDA for 3 hours after overnight growth in selective media to retain plasmids. Three hours later, cross-linking and subsequent ChIP procedures were performed according to [22] using anti-TAP antibody (Open Biosystems) and Protein A beads (Invitrogen). PCR reactions were run in multiplex using primer sets listed in Table S4. Simultaneous amplification of a cohesin-free site (534) was included as an internal negative control of the immunoprecipitation reaction (Figure S2). Gels were stained with EtBr and destained in water before digital photography (Alpha Innotech). All bands were found to be non-saturating and within the linear range. Reported values were calculated as (a2/549.7)IP/(a2/549.7)In.
|
10.1371/journal.pbio.2003611 | Nitric oxide-mediated posttranslational modifications control neurotransmitter release by modulating complexin farnesylation and enhancing its clamping ability | Nitric oxide (NO) regulates neuronal function and thus is critical for tuning neuronal communication. Mechanisms by which NO modulates protein function and interaction include posttranslational modifications (PTMs) such as S-nitrosylation. Importantly, cross signaling between S-nitrosylation and prenylation can have major regulatory potential. However, the exact protein targets and resulting changes in function remain elusive. Here, we interrogated the role of NO-dependent PTMs and farnesylation in synaptic transmission. We found that NO compromises synaptic function at the Drosophila neuromuscular junction (NMJ) in a cGMP-independent manner. NO suppressed release and reduced the size of available vesicle pools, which was reversed by glutathione (GSH) and occluded by genetic up-regulation of GSH-generating and de-nitrosylating glutamate-cysteine-ligase and S-nitroso-glutathione reductase activities. Enhanced nitrergic activity led to S-nitrosylation of the fusion-clamp protein complexin (cpx) and altered its membrane association and interactions with active zone (AZ) and soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor (SNARE) proteins. Furthermore, genetic and pharmacological suppression of farnesylation and a nitrosylation mimetic mutant of cpx induced identical physiological and localization phenotypes as caused by NO. Together, our data provide evidence for a novel physiological nitrergic molecular switch involving S-nitrosylation, which reversibly suppresses farnesylation and thereby enhances the net-clamping function of cpx. These data illustrate a new mechanistic signaling pathway by which regulation of farnesylation can fine-tune synaptic release.
| One way neurons communicate with each other and with other tissues, such as muscle, is by releasing chemical compounds known as neurotransmitters at sites of interaction known as synapses. This synaptic transmission can be finely regulated by both the releasing neuron and the receiving neuron or muscle cell. Many signaling molecules and pathways are involved in neurotransmitter release. In this study, we have investigated one of such pathways and its role in modulating neurotransmitter release at the neuromuscular synapse of the larva of the fruit fly Drosophila melanogaster. This regulation involves nitric oxide, a freely diffusible reactive molecule that can be generated in response to activity in the motor neuron. Several neuronal proteins can be modified by nitric oxide, and our study identified a specific target molecule that regulates neurotransmitter release. This protein, called complexin, undergoes a posttranslational modification in response to increased levels of nitric oxide, changing its localization and function at the synapse and modulating neurotransmission. Our findings can explain how neurons may modulate communication in an activity-dependent manner utilizing nitric oxide signaling.
| Throughout the central nervous system (CNS), the volume transmitter nitric oxide (NO) has been implicated in controlling synaptic function by multiple mechanisms, including modulation of transmitter release, plasticity, or neuronal excitability [1–3]. NO-mediated posttranslational modifications (PTMs) in particular have become increasingly recognized as regulators of specific target proteins [4]. S-nitrosylation is a nonenzymatic and reversible PTM resulting in the addition of a NO group to a cysteine (Cys) thiol/sulfhydryl group, leading to the generation of S-nitrosothiols (SNOs). In spite of the large number of SNO-proteins thus far identified, the functional outcomes and mechanisms of the underlying specificity of S-nitrosylation in terms of target proteins and Cys residues within these proteins are not clear.
Synaptic transmitter release is controlled by multiple signaling proteins and involves a cascade of signaling steps [5]. This process requires the assembly of the soluble N-ethyl-maleimide-sensitive fusion protein Attachment Protein Receptor (SNARE) complex and associated proteins, the majority of which can be regulated to modulate synapse function. Regulatory mechanisms include phosphorylation of SNARE proteins [6] as well as SNARE-binding proteins such as complexin (cpx), which have been reported at different synapses such as the Drosophila neuromuscular junction (NMJ) [7] or in the rat CNS [8].
Several contrasting effects on transmitter release are induced by NO-mediated PTMs [9]. Other forms of protein modification to modulate cellular signaling include prenylation, an attachment of a farnesyl or geranyl-geranyl moiety to a Cys residue in proteins harboring a C-terminal CAAX prenylation motif. This process renders proteins attached to endomembrane/endoplasmic reticulum (ER) and Golgi structures until further processing, as shown for Rab GTPases [10–12]. Farnesylation also regulates mouse cpx 3/4 [13] and Drosophila cpx function [14–17]. The Cys within CAAX motifs can also undergo S-nitrosylation, which interferes with the farnesylation signaling [18]; however, direct evidence in a physiological environment is lacking. Cpx function has been studied in many different systems and there is controversy regarding its fusion-clamp activity. Cpx supports Ca2+-triggered exocytosis but also exhibits a clamping function [19–24]. Analysis of mouse cpx double-knockout neurons lacking cpx 1 and 2 found only a facilitating function for cpx on release, and different D. melanogaster and Caenorhabditis elegans cpx mutant lines exhibit altered phenotypes in clamping or priming/fusion function [14–17, 24–27], illustrating the controversial actions of cpx.
Here, we investigated the effects of NO on synaptic transmission and found that NO reduces Ca2+-triggered release as well as the size of the functional vesicle pool, which was reversed by glutathione (GSH) signaling. At the same time, spontaneous release rates were negatively affected by NO. We confirmed that cpx is S-nitrosylated and that NO changes the synaptic localization of cpx, as also seen following genetic and pharmacological inhibition of farnesylation. Thus, we propose that the function of cpx is regulated by S-nitrosylation of Cys within the CAAX motif to prevent farnesylation. This increases cpx-SNARE-protein interactions, thereby rendering cpx with a dominant clamping function, which suppresses both spontaneous and evoked release.
Previously, we found that enhancing endogenous nitric oxide synthase (NOS) activity induced by overexpression of D. melanogaster NOS (DmNOS) caused a reduction in synaptic strength at the Drosophila NMJ synapse [28]. To examine the effects of NO on glutamatergic transmission in more detail, we exposed wild-type (WT) w1118 control (Ctrl) larvae to NO donors, which provide an estimated NO concentration of about 200 nM [29]. When recording evoked excitatory junction currents (eEJCs) up to 70 min during NO incubation, the amplitudes started to decline significantly after 35 min (Fig 1A and S1 Data, p < 0.05; n = 3 each). Mean eEJC amplitudes and quantal content (QC) at 50 min for Ctrl (122 ± 7 nA, QC: 200 ± 15, n = 20–22) and NO treatment (59 ± 7 nA, QC: 93 ± 10, n = 14) are shown in Fig 1B. As the canonical NO-cGMP pathway is active in Drosophila [30] and potentially responsible for this observation, we blocked the soluble guanylyl cyclase (sGC) with 1H-[1,2,4]oxadiazolo[4,3-a]quinoxalin-1-one (ODQ, 50 μM). Interestingly, ODQ did not prevent the effects of NO, suggesting a cGMP-independent mechanism (amplitudes: Ctrl + ODQ: 127 ± 5 nA, NO + ODQ: 70 ± 7 nA, QC: Ctrl + ODQ: 200 ± 22, NO + ODQ: 130 ± 11, Fig 1B, n = 10–16). As Drosophila has endogenous NO signaling and produces neuronal NO in a Ca2+/calmodulin-dependent manner [31, 32], we used NOS knockout-like (NOS “null”) larvae to assess endogenous NO modulation of release. We used two different lines with strongly reduced DmNOS showing NOS “null” activity (NOSC and NOSΔ15 [33, 34]) and we would expect that lack of endogenous NO generation has the opposite effects on release. When recording eEJCs, both genotypes exhibited a tendency towards larger eEJC amplitudes and QC (Fig 1C) and, in addition, we detected an increased presynaptic release probability (pvr) in NOSC NMJs, as indicated by the reduced paired pulse ratio (PPR) at 20 ms ISI (0.80 ± 0.03 [n = 11], p = 0.002, Student t test) compared to WT Ctrls (0.93 ± 0.03 [n = 17]), indicating endogenous nitrergic effects on release probabilities.
To further understand the effects of NO on release, we analyzed miniature EJCs (mEJCs) under the same conditions. NO had no effect on mEJC amplitudes or decay kinetics; however, the frequency was reduced following NO and NO+ODQ incubation (Ctrl: 2.0 ± 0.2 nA [n = 25], NO: 1.1 ± 0.1 nA [n = 16], NO+ODQ: 1.0 ± 0.2 nA [n = 8], ODQ: 1.7 ± 0.2 nA [n = 11], Ctrl versus NO: p < 0.01, Ctrl versus NO+ODQ: p < 0.05, Fig 1D). This suggests that NO is unlikely to affect synaptic vesicle filling or composition/activity and density of postsynaptic D. melanogaster glutamate receptors (DmGluR) [35]. We tested miniature events in the NOS “null” mutants and confirmed a further inhibitory role of NO signaling on release, with mEJC frequencies being significantly enhanced in NOSΔ15 (3.5 ± 0.5 s−1 [n = 4], p = 0.001) and NOSC (3.5 ± 0.4 s−1 [n = 16], p = 0.04) larvae compared to Ctrl (Fig 1E), without affecting mEJC amplitudes (NOSΔ15: 0.8 ± 0.1 nA [n = 13], NOSC: 1.1 ± 0.3 nA [n = 3] Fig 1E) or decay kinetics (NOSΔ15: 8.9 ± 0.6 ms [n = 12], NOSC: 9.4 ± 0.3 ms [n = 4], p > 0.05 versus Ctrl). Thus, reduction of endogenous NOS activity shows opposite effects to elevation of NO levels, confirming the inhibitory action of NO on evoked and spontaneous vesicle release.
As the data imply cGMP-independent signaling, we wanted to confirm that cGMP levels are not altered following NO stimulation. Thus, we measured cGMP directly in isolated larval brains. NO application did not raise cGMP levels (at 50 min: Ctrl: 2.4 ± 0.5 pmol/mg, NO: 3.0 ± 0.6 pmol/mg, p > 0.05 [n = 30 each], Fig 1F). Cyclase inhibition in the presence of NO did not significantly reduce cGMP levels, confirming lack of NO-induced neuronal cGMP accumulation. We found that any generated cGMP was broken down by phosphodiesterase DmPDE5/6 [36], as cGMP increased following NO stimulation only with PDE inhibition (20 μM zaprinast [Zap]; NO+Zap: 50.2 ± 8.3 pmol/mg, p < 0.0001), while Zap alone had no effect (Zap: 4.6 ± 2.0 pmol/mg, p > 0.05).
To assess whether NO is produced endogenously to induce modulation of synaptic function as observed above, we expressed FlincG3 presynaptically and stimulated NMJs at 20 Hz (for 10 s every minute for 20 min). As shown in Fig 1G, 20 Hz stimulation induced a significant increase in fluorescence, confirming endogenous presynaptic generation of NO (Ctrl: 62 ± 4 arbitrary units [a.u.’s], Stim: 96 ± 8 a.u.’s, Fig 1H [n = 13–15 boutons], p < 0.01). Importantly, addition of the NO donor did not further increase the fluorescence, indicating that activity-induced synaptic NO concentrations reach similar levels (NO: 93 ± 7 a.u.’s). A potential target of NO signaling is mitochondria [37], which are required for the energy to maintain vesicle recycling and synaptic transmission [38]. Thus, we measured mitochondrial activity in third instar larvae under the same conditions (50 min NO incubation) and found that mitochondrial activity was unaffected by NO (S1 Fig and S9 Data), suggesting that the effects of NO on synaptic transmission are not due to ATP depletion. Together, these data suggest that NO has a presynaptic effect on transmitter release, which is independent of cGMP signaling.
Several mechanisms contribute to the regulation of synaptic strength [39], including altered pvr, alterations in the number of readily releasable vesicles and release sites (N) or quantal size (q). Alterations in q are likely not involved in the NO-induced effects observed based on our mEJC data above (Fig 1). We next assessed additional release parameters, including pvr, N, vesicle pool size, and Ca2+ dependency of release in NOS “null” and WT NMJs following nitrergic signaling.
We determined pool size via a method successfully applied at the Drosophila NMJ, by analyzing the cumulative QC of trains of higher frequency stimulation [40]. Stimulation at 50 Hz for 500 ms in 1.5 mM extracellular calcium concentration ([Ca2+]e) retrieves vesicles from the readily releasable pool (RRP) [41]. This stimulation pattern induced mild depression in Ctrls and strong initial facilitation of trains under NO conditions (Fig 2A and S2 Data). Cumulative QC analysis revealed a pool size of 453 ± 37 (n = 17) in Ctrl and 185 ± 18 in NO-exposed NMJs (n = 16, p < 0.01), suggesting a strong reduction in ready-releasable/recycling vesicles (Fig 2A–2D). Supporting the above data, pool size estimation in the presence of ODQ confirmed cGMP independence (NO+ODQ: 310 ± 33 [n = 9], p < 0.05 versus Ctrl; ODQ alone: 501 ± 34 [n = 9], p > 0.05 versus Ctrl, Fig 2A–2D). And importantly, analysis of the vesicle pool sizes in NOS “null” mutants revealed a strong 2-fold increase compared to w1118 Ctrl and an over 5-fold increase compared to NO application (NOSC: 975 ± 161 [n = 11]; NOSΔ15: 958 ± 139 [n = 4], p < 0.001 versus Ctrl, Fig 2A–2D). To exclude any potential developmental effects caused by NOS deficiency that could account for these strong increases in release, we assessed NMJ morphology and ultrastructure. We analyzed the total volume of NMJs (horseradish peroxidase [HRP] signal) and the number of Bruchpilot (Brp) puncta/NMJ volume of z-stack confocal images (S2A and S2B Fig and S9 Data) and measured the number of AZs, T-bars per Ib bouton, and vesicles within a 250-nm semicircle around the AZ (S2C and S2D Fig). These data indicated that reduced NOS activity has no developmental impact on the structure of NMJs and synaptic boutons and can therefore not explain the physiological differences observed above.
In addition to changes in release, NO could also exert its effects indirectly via modulating transmitter uptake and pool recovery. To exclude this possibility that altered recovery from depression affected the above pool estimations, we examined eEJC recovery. Following depletion of vesicle pools during a 50-Hz train (1 s), we measured the time course of recovery over the following 60 s. NO did not show any effects on the time constant of recovery (S3 Fig and S9 Data).
In order to test whether NO acts specifically on RRP or also affects the availability of other pools, we stimulated the NMJ for longer periods (8 s) at 50 Hz. This prolonged stimulation leads to recruitment of vesicles from the reserve pool (RP) [42, 43]. Analysis revealed that NO also caused a strong reduction of release from the RP (Fig 2E–2H, Ctrl: 11,160 ± 1,645 [n = 6]; NO: 5,286 ± 798 [n = 7], p = 0.0062).
One important protein that regulates vesicle clustering and release of neurotransmitter is the phosphoprotein synapsin (syn), which regulates recycling of RP vesicles in Drosophila NMJs [43]. We tested whether modulation of syn could be responsible by employing larvae deficient in this protein from the Syn97-null mutation [44]. These larvae did not exhibit any reduction in single-stimulus QC compared to Ctrls, but prolonged recruitment (500 ms at 50 Hz) showed reduced vesicle availabilities. Importantly, incubation of Syn97 larvae with NO led to further reduction of both parameters (S4 Fig and S9 Data), suggesting that NO effects are via a different signaling route.
Based on these data, we suggest that NO decreases release of vesicles from the RRP and RP but does not affect the rate of vesicle pool recovery from depletion. The NO-mediated effects appear to be independent of syn, suggesting an event downstream of vesicle recruitment per se.
We next applied an independent approach to estimate the synaptic parameters: fluctuation analysis [45] to estimate the number of functional release sites N. eEJCs were elicited at varying calcium concentrations ([Ca2+]e: 0.5–3 mM, 0.2 Hz) and amplitudes were plotted over [Ca2+]e (Fig 3A and 3B and S3 Data). NO exposure led to reduced release across different Ca2+ concentrations (0.75–3 mM). N was estimated from parabolic fits to the variance-mean plots for each NMJ (Fig 3C). This analysis revealed a strong reduction in N following NO exposure (Fig 3D, NCtrl: 630 ± 104 [n = 5], NNO: 117 ± 32 [n = 6], p = 0.0006). The estimation of N from the fluctuation analysis (about 600) in Ctrl is in accordance with previously reported electron microscopy (EM) data showing a number of about 500 vesicles per NMJ [46]. These data confirm that NO most likely reduces the number of releasable vesicles by preventing vesicle fusion at individual release sites.
The reduced QC seen following NO exposure can also be attributable to a change in the Ca2+ dependency of release, so we determined whether the reduced transmitter release is due to altered Ca2+ cooperativity of release [47]. The Hill slope was strongly reduced by NO (Ctrl: 3.2 ± 0.4 [n = 6], NO: 1.8 ± 0.7 [n = 5], p = 0.0024, Fig 3E); however, the half maximal effective Ca2+ concentration (EC50) was unaltered (Ctrl: 1.0 ± 0.03, NO: 1.0 ± 0.09, p > 0.05, Fig 3E), indicating that sensitivity to Ca2+ was not affected by NO.
To further assess nitrergic effects on pvr, we used the PPR approach by delivering two pulses with interspike intervals (ISIs) between 10 and 200 ms at two different [Ca2+]e (1 and 1.5 mM, Fig 3F and 3G) in Ctrl and NO-treated NMJs. Analysis showed that Ctrl NMJs only exhibit slight potentiation at low Ca2+ and high ISI, indicative of low pvr. In contrast, pvr in the presence of NO was decreased, as shown by an increased PPR (potentiation at all ISI at 1 mM Ca2+ and 20 and 40 ms ISI at 1.5 mM Ca2+, p < 0.05, Ctrl versus NO at each ISI), which is also in agreement with elevated pvr in NOS “null” larvae. With about 500 release sites per NMJ and a QC of 200 (Ctrl) and 90 (NO), our data present estimated pvr values of 0.33 (Ctrl) and 0.16 (NO), with Ctrl values similar to estimates made previously in WT larvae [40].
Previously, we have shown that NO signaling can suppress mammalian P/Q and N-type Ca2+ channels [48]. In order to test whether altered Ca2+ influx could cause the observed effects on evoked release at the NMJ, we tested whether NO application for 60 min changed presynaptic Ca2+ levels during a train of synaptic stimulation. GCaMP5 was expressed presynaptically and activity-evoked Ca2+ influx in type 1b NMJ boutons was imaged at different extracellular Ca2+ concentrations (0.25–3 mM). Our data showed that NO had no effect on stimulated Ca2+ levels at any concentration tested (ΔF/F0, myrGCamP5: 3 mM Ca2+: Ctrl: 0.70 ± 0.09, NO: 0.78 ± 0.12 [n = 13–18 boutons from 4–6 NMJs each], p > 0.05; Fig 3H and 3I; GCaMP5: 0.25 mM Ca2+: 0.24 ± 0.03, NO: 0.14 ± 0.03, 0.5 mM Ca2+: Ctrl: 0.42 ± 0.08, NO: 0.50 ± 0.07, 1.5 mM Ca2+: Ctrl: 1.13 ± 0.14, NO: 1.18 ± 0.21 [n = 28–46 boutons from 7–11 NMJs each], p > 0.05; S5 Fig and S9 Data). Together, the data suggest that NO reduced evoked release and the frequency of spontaneous release, likely due to reduced release probability and Ca2+ cooperativity, which manifests itself in reduced vesicle fusion. We showed that the Ca2+ dependence of release, but not Ca2+ entry per se, was reduced by NO, which indicates a possible modulation of SNARE (-associated) protein interactions via NO-mediated PTMs.
S-nitrosylation is a reversible non-enzymatic protein modification, the levels of which can be regulated via S-nitrosoglutathione reductase (GSNOR), the sole alcohol dehydrogenase 5 (ADH-5) isozyme in vertebrate brains [49], which has a homologue in Drosophila (encoded by the formaldehyde dehydrogenase [fdh] gene). This de-nitrosylation process requires GSH. GSH is produced from L-glutamate and Cys via the enzyme glutamate-cysteine ligase (GCL), the rate-limiting step in GSH synthesis in fly [50]. The Drosophila GCL holoenzyme is heterodimeric, consisting of a catalytic (DmGCLc) and a modifier (DmGCLm) subunit, each encoded by a unique gene, and overexpression of either subunit increases cellular GSH levels [50].
In order to assess the contributions of SNO formation to the physiology at the NMJ, we investigated the effects of altering neuronal GSH levels. If NO mediates its observed actions via SNO formation, we should be able to prevent/reduce the effects on transmitter release by providing elevated GSH levels by (i) GSH supplementation, (ii) overexpression of GSNOR (fdh), or (iii) overexpression of GCL (DmGCLm/c) and, inversely, enhance NO effects by using RNA interference (RNAi) expression of the above proteins. We tested first the recovery of NO-mediated reduction of eEJC amplitudes following NO exposure for 50 min by washing out NO. eEJC amplitudes recovered slightly (Fig 4A, green and S4 Data); however, when washing in GSH (150 μM), the amplitudes recovered to control levels after 15 min (GSH [blue] versus NO at 50 min [red], p < 0.05), indicating a GSH-mediated reversal.
To characterize effects of endogenous GSH formation, we used elav-Gal4-driven UAS-fdh31, UAS-DmGCLm, and UAS-DmGCLc overexpression. It has been shown that overexpression of either DmGCLc or DmGCLm results in enhanced enzyme activity and elevated GSH levels [50], GSNOR overexpression (elav > UAS-fdh31) reduces global S-nitrosylation in fly, and conversely, GSNOR-RNAi expression (elav > UAS-fdhri34) elevates SNO protein levels [51]. Overexpression of GSNOR and GCLm/c (Fig 4B–4D) prevented NO effects on QC (GSNOR: 238 ± 20 [n = 11], DmGCLm: 197 ± 32 [n = 7], DmGCLc: 215 ± 39 [n = 6], GSNOR+NO: 223 ± 24 [n = 8], DmGCLm+NO: 177 ± 10 [n = 7], DmGCLc+NO: 329 ± 26 [n = 3], p > 0.05) and vesicle pool sizes (GSNOR: 438 ± 51 [n = 10], DmGCLm: 400 ± 99 [n = 7], DmGCLc: 496 ± 93 [n = 6], GSNOR+NO: 472 ± 34 [n = 8), DmGCLm+NO: 360 ± 40 [n = 7], DmGCLc+NO: 685 ± 148 [n = 3], p > 0.05, Fig 4B–4D). These data confirm that by enhancing GSNOR and GCL activities, thereby elevating intracellular GSH levels, the effects of NO on pool size and pvr (PPR at 20 ms ISI; w1118 Ctrl [0.93 ± 0.03] versus NO [1.2 ± 0.07], p < 0.0001, GSNOR overexpression [0.88 ± 0.02], +NO [0.84 ± 0.05]/GCLm overexpression [0.87 ± 0.04], +NO [0.92 ± 0.03]/GCLc overexpression [0.99 ± 0.07], +NO [0.96 ± 0.02], p > 0.05, Fig 4E) were precluded, suggesting that this was due to reduced SNO formation. Furthermore, overexpression of GSNOR, DmGCLm, and DmGCLc prevented the reduction in mEJC frequency following NO exposure (fGSNOR: 2.4 ± 0.3 s−1 [n = 13]; fDmGCLm: 3.0 ± 0.3 s−1 [n = 13]; fDmGCLc: 1.5 ± 0.42 s−1 [n = 5]; fGSNOR+NO: 1.7 ± 0.2 s−1 [n = 7]; fDmGCLm+NO: 2.9 ± 0.4 s−1 [n = 13]; fDmGCLc+NO: 0.4 ± 0.1 s−1 [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F) without affecting mEJC amplitudes (GSNOR: −0.6 ± 0.07 nA [n = 13]; DmGCLm: −0.7 ± 0.07 nA [n = 13]; DmGCLc: −0.5 ± 0.07 nA [n = 5]; GSNOR+NO: −0.6 ± 0.07 nA [n = 7]; DmGCLm+NO: −0.6 ± 0.08 nA [n = 13]; DmGCLc+NO: −0.6 ± 0.07 nA [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F) or decays (GSNOR: 7.5 ± 0.2 ms [n = 13]; DmGCLm: 9.7 ± 0.4 ms [n = 13], DmGCLc: 6.7 ± 0.3 ms [n = 5], GSNOR+NO: 6.2 ± 0.2 ms [n = 7], DmGCLm+NO: 7.9 ± 0.5 ms [n = 13], DmGCLc+NO: 6.2 ± 0.4 ms [n = 3], p > 0.05 versus w1118 Ctrl and versus each Ctrl, Fig 4F). Furthermore, the reduction of endogenous GSNOR and DmGCLm activities (elav > UAS-RNAi) caused partial electrophysiological phenotypes, such as a decrease in eEJC amplitudes, QC, or vesicle pool size compared to w1118 Ctrl, with NO having no further major negative effects (S6 Fig and S9 Data).
We next asked which signaling routes and PTMs are involved in NO modulation of release. The SNARE-binding and fusion-clamp protein cpx regulates not only the Ca2+ cooperativity of evoked release but also spontaneous release [14] as well as release probabilities [52], thereby presenting a strong candidate for mediating the observed NO-induced changes.
Cpx acts by binding to the SNARE complex, thereby promoting the clamping of release, and only when replaced by synaptotagmin 1 in response to Ca2+ influx will vesicle fusion be initiated [14, 20]. Dmcpx function can be regulated by protein kinase A (PKA) phosphorylation of serine126 (Ser126) [7] or by prenylation at the C-terminus [15, 16]. In order to test whether cpx is required to exert NO effects, we first used cpx null mutants (cpxSH1, cpx-/-) [14]. In these animals, we detected a strong reduction in evoked release and QC (22.6 ± 3.2 [n = 11], p < 0.0001 versus Ctrl), which was unaffected by NO (13.8 ± 2.0 [n = 4], p > 0.05 versus cpx-/-, p < 0.0001 versus Ctrl, Fig 5A–5D and S5 Data). Similarly, when comparing the vesicle pool size, cpx-/- NMJs showed a strong reduction (26 ± 6 [n = 11], p < 0.0001 versus Ctrl), which again was unaffected by NO (22 ± 5 [n = 4], p > 0.05 versus cpx-/-, p < 0.0001 versus Ctrl, Fig 5A–5D). These data confirm that cpx is required for NO to induce suppression of evoked release and available vesicle pool size and suggest that NO might enhance the clamping function of cpx in WT larvae. We next tested the impact of NO on the clamping ability of cpx by characterizing spontaneous release. Interestingly, the frequency of spontaneous events inversely correlates with endogenous cpx levels [14]. We analyzed mEJCs in cpx-/- muscle 6 (m6), which exhibited an extremely high frequency [14] (>40 × w1118, Fig 5E). NO did not reduce the mEJC frequency in those preparations, although a precise analysis is difficult due to strong overlap of single mEJCs [14]. In order to allow more accurate frequency measurements in cpx-/- animals, we used neighboring muscle 5 (m5), posessing a synapse with approximately 4-fold fewer release sites compared to m6. Similar to m6, cpx-/- increased mEJC frequencies >10-fold compared to Ctrl (m5: w1118: 0.8 ± 0.2 s−1 [n = 3], cpx-/-: 11.6 ± 0.8 s−1 [n = 6], p < 0.0001); however, following NO exposure, this preparation did not show any change in mEJC frequency (m5 cpx-/- + NO: 9.7 ± 0.9 s−1 [n = 5], p > 0.05 versus m5 cpx-/-, Fig 5E and 5F), suggesting the requirement of cpx for the observed nitrergic effects. Nevertheless, we recorded from m6 of heterozygous animals, which exhibit higher frequencies than w1118 but are still accurately quantifiable (m6 cpx+/-: 4.6 ± 0.8 s−1 [n = 5]). Here, NO induced a strong reduction in the frequency (m6 cpx+/- + NO: 0.6 ± 0.2 s−1 [n = 5] #p < 0.05 versus m6 cpx+/- Ctrl, Fig 5E and 5F), similar to that seen in w1118. These data confirm that NO only modulates spontaneous release frequencies in the presence of cpx.
Together, these data show that in the absence of cpx, NO causes no electrophysiological phenotypes. The NO-mediated reduction of eEJC amplitudes, QC, pool size, and mEJC frequency all require the presence of cpx, suggesting that its modulation might be responsible for the observed nitrergic effects, which could be explained by a gain-of-clamping function [53]. This potential effect was further investigated by using the established paradigm of activity-induced enhancement of spontaneous release at the Drosophila NMJ [7]. We assessed whether NO modulation of release also affects this activity-dependent signaling, which would strengthen the role of cpx as a target for nitrergic regulation and a general regulatory mechanism. PKA has been reported to modulate mEJC frequency potentiation in a cpx overexpression model (Dmcpx 7B, [7]). We confirmed that high frequency stimulation (50 Hz for 3 s) led to an enhanced mEJC frequency in w1118 NMJs relative to baseline (Ctrl: 1.9 ± 0.2-fold [n = 13], Fig 5G and 5H). Interestingly, repeating this protocol in larvae exposed to NO showed a lack of frequency potentiation (NO: 0.8 ± 0.1-fold [n = 14], p < 0.05 versus Ctrl), which was also ODQ independent (NO + ODQ: 1.0 ± 0.1-fold [n = 7], p > 0.05 versus NO, Fig 5G and 5H). To test whether the manipulation of PTMs also affects nitrergic suppression of frequency potentiation, we used larvae overexpressing GCLm and GSNOR and NOS “null” larvae. We found that GCLm and GSNOR overexpression occluded nitrergic effects on suppression of mEJC frequency potentiation, whereas the lack of NO signaling led to enhanced potentiation (GCLm + NO: 2.3 ± 0.4-fold [n = 7], GSNOR + NO: 2.5 ± 0.5-fold [n = 7], NOS “null” [comprised of n = 5 NOSC and n = 3 NOSΔ15]: 3.8 ± 0.3-fold, **p < 0.01 versus Ctrl, ##p < 0.01 versus NO, ####p < 0.001 versus NO, Fig 5G and 5H). These data show that NO suppresses the activity-mediated increase in mEJC frequency and suggest that, similar to phospho-incompetent cpx mutants [7], nitrergic modulation of WT cpx produces an inhibitory action on spontaneous release. The lack of PTM signaling leads to an enhanced frequency potentiation, strengthening the notion that NO-mediated effects are responsible for suppression of synaptic release and our data point towards modulation of cpx as a key signaling mechanism.
Having shown that cpx signaling is involved in NO-mediated effects on spontaneous and evoked release, we next considered if S-nitrosylation of the Cys residue within the C-terminus of cpx possessing the CAAX motif could explain the observed results. Importantly, prenylation has been studied in several genetically modified cpx proteins in which the CAAX motif was eliminated [15, 16]. These studies suggest that deletion of final parts of the C-terminus/final amino acid affects cpx localization, interactions with SNARE-proteins, and, subsequently, its function. To explore the effects of cpx farnesylation more in detail, we made use of Drosophila lines expressing green fluorescent protein (GFP)-tagged WT and mutant cpx (cpx1257, lacking the final amino acid [16]), referred to as CpxΔX. This mutant has been shown to exhibit altered co-localization with syntaxin at the dorsolongitudinal flight muscle (DLM) neuromuscular synapse. We assessed localizations of WT and mutant cpx at the NMJ (elav > UAS-cpx-GFP, elav > UAS-cpx1257-GFP) with respect to their interaction with the AZ protein, Brp. WT cpx exhibits diffuse localization within boutons (as previously reported [15]) with little co-localization with Brp (Fig 6A and 6B and S6 Data). In contrast, the mutant form, lacking farnesylation, is highly co-localized with Brp, as indicated by the increase in Pearson’s coefficient (Fig 6A and 6B; WT cpx: 0.35 ± 0.30 [n = 9], CpxΔX: 0.65 ± 0.03 [n = 9], p < 0.0001). These data confirm that preventing cpx farnesylation results in enhanced co-localization with AZ. To further support these data, we conducted high-resolution stimulated emission depletion (STED) microscopy [54] and analyzed the Pearson’s coefficient for the co-localization of Brp with cpx. This experiment verified the confocal data showing enhanced co-localization of CpxΔX with Brp versus WT cpx (WT cpx: 0.13 ± 0.02 [n = 25], CpxΔX: 0.27 ± 0.02 [n = 23], p < 0.0001, Fig 6C and 6D). As Dmcpx possesses a predominant clamping function [23], we propose that NO could lead to a reduction in farnesylation, a consequent stronger interaction with the SNARE complex at the AZ, and thereby enhance its clamping function upon transmitter release. To specifically confirm co-localizations, we used the high-resolution proximity ligation assay (PLA), with which we imaged interactions of Brp with cpx. We used both lines, WT cpx-GFP and CpxΔX-GFP expressing larvae, and found that PLA signals are strongly enhanced at NMJs expressing the mutant cpx (Fig 6E and 6F; WT cpx: 0.04 ± 0.004 [n = 9], CpxΔX: 0.12 ± 0.02 [n = 9], p = 0.009). As the co-localization data may depend upon expression of GFP-tagged cpx, we confirmed equal GFP expression levels in both lines by immunoblotting (S9A Fig). These co-localization and PLA experiments confirm an enhanced association of a mutated farnesylation-incompetent cpx with Brp and suggest that lack of farnesylation renders cpx in close proximity to release sites of AZs.
In order to assess this possibility further, we used pharmacological and genetic tools to modulate cpx farnesylation and compared protein localization and synaptic release following farnesyl transferase (FTase) inhibition and NO exposure. Reduced expression of the Drosophila ortholog of FTase or inhibition of FTase by L-744,832 and GGTI-298 have strong effects on fly lethality [55], implicating a crucial function of this signaling in fly.
First, we tested how FTase inhibition (20 μM L-744,832 + 10 μM GGTI-298) and NO exposure affect cpx co-localization with the SNARE complex proteins syntaxin and synaptotagmin or Brp, using the PLA. We measured total PLA volume of NMJ z-stacks and normalized PLA signals to NMJ volume. We found that both treatments (depicted as “farnesyl inh” and “NO,” Fig 7A and 7B and S7 Data) led to enhanced co-localization of cpx with syntaxin and Brp (syntaxin-cpx: Ctrl: 0.04 ± 0.007, NO: 0.12 ± 0.02, farnesyl inh: 0.11 ± 0.02, Brp-cpx: 0.02 ± 0.007, NO: 0.08 ± 0.03, farnesyl inh: 0.09 ± 0.05, Fig 7A and 7B; p < 0.01, p < 0.001 versus Ctrl), suggesting that NO PTMs and farnesylation inhibition enrich cpx at the AZ. When analyzing the interactions between the Ca2+ sensor synaptotagmin and cpx, we found that this interaction was completely suppressed following treatments (Ctrl: 0.2 ± 0.06, NO: 0.03 ± 0.006, farnesyl inh: 0.04 ± 0.007, Fig 7A and 7B; p < 0.01 versus Ctrl). The PLA data were further supported by STED imaging studies showing identical changes in protein co-localization, as determined by Pearson’s coefficient analysis (S7 Fig and S9 Data). One possibility to allow for greater amounts of cpx to be available for binding to SNAREs is by enhancing free and soluble cytosolic levels as a consequence of reduced farnesylation. Farnesylation of cpx results in its membrane tethering, and thus protein fractions, which are membrane bound, are less mobile than soluble cytosolic proteins. To assess the mobility of potentially farnesylated versus soluble (non-farnesylated) cpx and thus distinguish between these two pools of cpx, we performed fluorescence recovery after photobleaching (FRAP) analysis of GFP-tagged WT and farnesylation-incompetent cpx (CpxΔX). Although a previous study did not detect differences between farnesylated versus non-farnesylated cpx isoform using this method with a photo-bleaching area of half a bouton [15], we found that accurate FRAP analysis of cpx-GFP mobility can only be performed by using substantially smaller bleaching areas, as reported previously [56] (S8 Fig and S9 Data). Using this approach, we found that bleaching an area of 2.5 μm2 (instead of >10 μm2) generally leads to faster recovery rates (S8 Fig and S9 Data). Our data confirmed that lack of farnesylation (CpxΔX) allows for greater movement of cpx and faster recovery (tau: WT cpx: 18.1 ± 1.7 ms, CpxΔX: 11.9 ± 1.2 ms [p < 0.05], WT Cpx + NO: 8.8 ± 0.8 ms [p < 0.0001], n = 18–20, Fig 7C), as expected for a soluble protein. Our data further show that NO treatment caused the same increase in recovery rates (Fig 7C), suggesting that NO also prevented farnesylation.
These data suggest that due to enriched local levels, cpx outcompetes synaptotagmin for SNARE binding at the AZ, thereby displacing synaptotagmin, as reported previously in biochemical studies [53]. Our data show that pharmacological and genetic inhibition of farnesylation promotes cpx co-localization with the AZ and supports the notion that this negatively impacts on synaptotagmin-SNARE complex binding, subsequently reducing release. The specificity of the PLA was corroborated by lack of Brp-cpx PLA signals in cpx-/- larvae (S9B–S9D Fig). Next, we explored the possibility of whether specific inhibition of FTase activity by L-744,832 and GGTI-298 and FTase RNAi mimics the effects of NO on synaptic transmission. We found that, in both conditions, the frequency of mEJCs was reduced to similar values seen following NO exposure (fmEJC: L-744,832 + GGTI-298: 0.7 ± 0.1 s−1 [n = 8], p = 0.0051 versus Ctrl, FTase RNAi: 0.9 ± 0.2 s−1 [n = 9], p = 0.0136 versus Ctrl, Student t test, Fig 7D). Importantly, both L-744,832 + GGTI-298 and FTase RNAi expression reduced evoked transmission and available vesicle pool size to levels similar to those following NO incubation (L-744,832 + GGTI-298: eEJC: 56 ± 5 nA, QC: 80 ± 13 [n = 9], pool size: 180 ± 27 [n = 9], p < 0.0001 versus each w1118 Ctrl; FTase RNAi: eEJC: 75 ± 5 nA, QC: 82 ± 6 [n = 9], pool size: 120 ± 17 [n = 9], p < 0.0001 versus each w1118 Ctrl, Student t test, Fig 7E and 7F). These data suggest that the farnesylation status of cpx mediates nitrergic effects, resulting in changed SNARE protein interactions, which determines the physiological outcome of cpx.
To further investigate the effects of NO directly on the prenylation process, we employed the well-characterized GFP-CAAX transfection model [57]. Here, human embryonic kidney (HEK) cells were transfected with GFP-CAAX (K-Ras motif) and the membrane association was assessed in response to prenylation inhibition and NO treatment. In control conditions, GFP exhibited a strong fluorescence signal at the membrane, which disappeared and redistributed into the cytosol following pharmacological inhibition of prenylation (L-744,832 + GGTI-298, p < 0.0001), confirming the prenylation-mediated localization of GFP-CAAX to the membrane (Fig 8A and S8 Data). Importantly, we showed that NO treatment (propylamine propylamine NONOate [PAPA-NONOate], p < 0.0001) induced a similar phenotype, with GFP being localized predominantly in a cytosolic manner—suggesting that NO prevents farnesylation through the same pathway (Fig 8A). To confirm that the Cys within the CAAX motif can undergo S-nitrosylation, we performed the Biotin Switch Assay on cpx-3 from isolated mouse retinas. NO donor incubation induced a >2-fold increase in SNO-cpx (Fig 8B), confirming this PTM on cpx and suggesting that this PTM is responsible for NO-induced changes in localization and function of cpx.
To specifically confirm the effects of S-nitrosylation and SNO interaction with farnesylation of cpx in Drosophila, we generated and expressed a nitroso-mimetic cpx mutant (Dmcpx 7AC140W) in a cpx null background (cpxSH1) and assessed synaptic responses. The Cys140 of Dmcpx is located within a hydrophobic region, as predicted in the Kyle Doolittle plot, which favors S-nitrosylation [58]. This mutant exhibits reduced evoked responses, QC, and vesicle pool sizes (eEJC: 70 ± 7 nA, QC: 106 ± 8, pool size: 204 ± 23 [n = 15 each], p < 0.0001 versus each w1118 Ctrl, Fig 8C and 8D), indicating that the mimicking of S-nitrosylation and simultaneous lack of farnesylation of cpx caused the observed changes. Importantly, this mutation also induced a reduction in spontaneous activity (fmEJC: 1.3 ± 0.2 s−1 [n = 15], p < 0.05 versus w1118 Ctrl, Fig 8C and 8D), reinforcing the argument of enhanced clamping function due to SNO formation and lack of farnesylation. The expression of WT cpx in the null background did not affect QC, pool size, or mEJC frequency (QC: 167 ± 17 [n = 5]; pool size: 381 ± 76 [n = 5]; fmEJC: 2.4 ± 0.4 s−1 [n = 10 each], p > 0.05 versus each w1118 Ctrl). To confirm changes in localization of Dmcpx 7AC140W, we analyzed PLA signals and found that Dmcpx 7AC140W highly co-localizes with Brp, in strong contrast to WT cpx (WT: 0.025 ± 0.013, Dmcpx 7AC140W: 0.17 ± 0.03 [n = 6–7], p < 0.0001, both expressed in cpx-/- background, Fig 8E and 8F). The data from the PLA experiments were confirmed by STED confocal microscopy, showing significantly higher Pearson’s coefficients for the co-localization of the cpx mutant C140W with Brp relative to the interaction of WT cpx with Brp (WT cpx: 0.13 ± 0.03, Dmcpx 7AC140W: 0.34 ± 0.02 [n = 20–24], p < 0.0001; Fig 8E and 8G). These data demonstrate that independent approaches to block farnesylation (and mimic of cpx-SNO) recapitulate nitrergic modulation of release and protein localization and therefore link for the first time NO-induced PTM and farnesylation signaling of cpx. We propose that S-nitrosylation acts as a novel endogenous pathway to alter cpx farnesylation signaling and protein–protein interactions and thereby allows a fine-tuning of synaptic function.
NO regulates a multitude of physiological and pathological pathways in neuronal function via generation of cGMP, thiol-nitrosylation, and 3-nitrotyrosination in health and disease [59]. Here, we show by employing biochemical and genetic tools in Drosophila, mouse, and HEK cells that NO can S-nitrosylate cpx and modulate—in a cGMP-independent manner—neurotransmitter release at the NMJ by interfering with its prenylation status, thereby affecting the localization and function of this fusion-clamp protein. We found that these nitrergic effects are reversed by GSH application or overexpression of GSH-liberating and de-nitrosylating enzymes (GCLm/c, GSNOR). GSH is the major endogenous scavenger for the NO moiety by the formation of S-nitrosoglutathione (GSNO) and consequently reduces protein-SNO levels via trans- and de-nitrosylation. The suppression of NOS activity facilitates synaptic function and the data support the notion that endogenous or exogenous NO enhances S-nitrosylation, reduces cpx farnesylation, and diminishes release.
Of the numerous synaptic molecules involved in release, cpx in particular has been implicated in the regulation of both evoked and spontaneous release due to its fusion-clamp activity. Despite the seemingly simple structure of cpx, its physiological function is highly controversial, as this small SNARE-complex binding protein can both facilitate but also diminish fast Ca2+-dependent and spontaneous release, depending on the system studied [22, 25, 53, 60]. In addition, there are different mammalian isoforms of cpx (1–4), which differ in their C-terminal region, with only cpx 3/4 containing the CAAX prenylation motif. Farnesylation in general determines protein membrane association and protein–protein interactions [61], and some cpx isoforms, such as muscpx 3/4 and Dmcpx 7A, are regulated in this manner [13, 23, 62]. However, muscpx 1/2 does not possess a CAAX motif, suggesting differential regulatory pathways to modulate cpx function. In Drosophila, there are alternative splice variants resulting from a single cpx gene, but the predominant isoform contains the CAAX motif (Dmcpx 7A), implicating the importance of this signaling molecule [15, 16]. The other splice isoform (Dmcpx 7B) lacks the CAAX motif and is expressed at about 1,000-fold lower levels at the larval stage [15], thus making Dmcpx 7A the dominant isoform to be regulated by farnesylation. However, the lack of Dmcpx 7B phosphorylation by PKA induces similar phenotypes as seen in our experiments when assessed following an induction of activity-dependent potentiation of mEJC frequency [7], which also may involve cpx–synaptotagmin 1 interactions.
Interestingly, both depletion and excessive levels of cpx suppress Ca2+-dependent and -independent exocytosis [63]. Cpx may promote SNARE complex assembly and simultaneously block completion of fusion by retaining it in a highly fusogenic state. Ca2+-dependent fusion is promoted below a concentration of 100 nM of cpx, whereas above 200 nM, it exhibits a clamping function resulting in a bell-shaped response curve [64]. Previous work suggests that synaptotagmin 1, once bound to Ca2+, relieves the cpx block and allows fusion. Another study reported that selective competition between cpx and synaptotagmin 1 for SNARE binding allows regulation of release [53]. Our data are in agreement with the latter findings, as we observed reduced synaptotagmin 1–cpx interactions following the block of farnesylation (Fig 7), indicating fewer synaptotagmin molecules binding to the SNARE complex to displace cpx. This limited replacement of cpx by synaptotagmin has been implicated in biochemical studies showing that local excess of cpx inhibits release, presumably by outcompeting synaptotagmin binding [53, 60]. Thus, synaptotagmin-SNARE binding is strongly dependent upon the local concentration of cpx [53]. Alternatively, and we cannot exclude this possibility, the modulation of cpx may simply alter its binding to the SNARE complex without directly displacing synaptotagmin, but interpretation of the data from our assays (PLA, co-localization) would not allow us to distinguish between these possibilities.
Our data are compatible with the idea that cpx binds to the SNARE complex, facilitates assembly, and then exerts its clamping function by preventing full fusion due to SNARE complex stabilization and subsequent increased energy barrier to allow fusion. Our model could provide an explanation of how cpx can be regulated to signal downstream to modulate transmitter release. So far, there are no data available, apart from mutation studies, as to how cpx function can be altered. We provide data indicating a physiologically relevant mechanism to adjust cpx function, possibly to the requirements of the neuron to adjust synaptic transmission. This likely occurs due to Cys S-nitrosylation and suppression of farnesylation, allowing greater amounts of hydrophilic cpx, not bound to endomembranes, to be available for binding with the SNARE complex in an altered configuration. This cross signaling between nitrosylation/farnesylation has been proposed to act as a molecular switch to modulate Ras activity [65]. Our data show that enhanced nitrergic activity and blocking farnesylation, either genetically (CpxΔX) or pharmacologically, alters the localization of cpx at the Drosophila NMJ and that of GFP-CAAX in HEK cells (Figs 6–8). Furthermore, by using a nitroso-mimetic cpx mutant, we found enhanced co-localization of cpx with the AZ protein Brp, implying a localization-function relationship (Fig 8). This consequently increases the net-clamping function because of elevated local concentrations of cpx. Dmcpx specifically exhibits a strong clamping function, as shown following overexpression in hippocampal neurons, which causes suppression of evoked and spontaneous release accompanied by a reduction of the release probability [23] or reduced vesicle fusion efficiency in in vitro assays [64].
Two independent studies eliminating the CAAX motif in Dmcpx (cpx572 and cpx1257) investigated localization-function interactions and showed disagreeing effects on both release and cpx localization [15, 16]. In particular, it has also been shown that the truncated cpx (cpx572, lacking the last 25 amino acids) does not co-localize with Brp [15]. Interestingly, this mutant causes a strong decrease in C-terminal hydrophobicity and a modest physiological response (increased mini frequency, decreased evoked amplitudes equivalent to a loss of clamping and loss of fusion function) relative to the total knock-out (KO). In contrast, the cpx mutant with single amino acid deletion (cpx1275) causes no effect on evoked but identical effects on the frequency of spontaneous release, suggesting a lack of clamping but no lack of fusion function. In addition, this mutant now co-localizes with the AZ at the NMJ [16]. These two studies indicate that the different mutations cause contrasting electrophysiological and morphological phenotypes, indicating that it is due to the nature of the mutation (lack of the last 25 amino acids versus 1 amino acid), which highlights the importance of a functional C-terminus. More recent studies have shown that deletions of the final amino acids (6 or 12 residues) completely abolished the membrane binding of cpx-1, impairing its inhibitory function and confirming the requirement of an intact C-terminus for inhibition of release [66, 67]. Here, we use an endogenous cpx with intact hydrophobic C-terminus, allowing physiological membrane binding. This is essential for inhibitory function, as the C-terminus is required for selective binding to highly curved membranes, such as those of vesicles [68]. Thus, as we used different approaches to alter farnesylation and generated a single amino acid mutant cpx (Dmcpx 7AC140W), leaving the C-terminus intact, our studies were performed under conditions of endogenous regulation of cpx function and thus provide new functional data on cpx signaling. Importantly, our data show that this regulation alters cpx function, and this is the first study to provide an explanation for the differential effects observed using cpx mutants or even cpx protein fragments in mammals, worm, and fly in various cross-species rescue experiments [20, 23].
Our data are in agreement with a model that non-farnesylated hydrophilic and soluble cytosolic cpx binds to the vesicular membrane via its C-terminal interactions, thereby exerting its inhibitory effect. When proteins are farnesylated, they are likely tethered to endomembranes, other than vesicle membranes [12]. It has to be distinguished between cpx interaction with the vesicle membrane as a result of the hydrophobic C-terminus, allowing cpx to become in close proximity to the AZ, and cpx endomembrane binding following farnesylation, which prevents cpx interactions with the AZ. However, in our case, SNO modification may enhance the binding to other proteins (e.g., SNAREs), thereby augmenting the effects. These additional interactions with unknown binding partners may affect proper cpx function and explain some of the discrepancies seen in studies using other genetically altered cpxs.
In summary, our study provides new data to illustrate a potential mechanism to regulate cpx function in a physiological environment, and we showed that NO acts as an endogenous signaling molecule that regulates synaptic membrane targeting of cpx, a pathway that may reconcile some of the controversial findings regarding cpx function. We suggest that increased S-nitrosylation and consequent lack of farnesylation leads to enhanced cytosolic levels of a soluble hydrophilic cpx and less endomembrane-bound fractions (Fig 9), because farnesylation-incompetent proteins remain in the cytosol [12]. These novel observations advance our understanding of similar nitrergic regulation of farnesylation that may be relevant for mammalian cpx-dependent synaptic transmission at the retina ribbon synapse and other brain regions [13]. Finally, this work has broader implications for physiological or pathological regulation of the prenylation pathway not only during neurodegeneration and aging, when enhanced S-nitrosylation might contribute to abnormal farnesylation signaling [69, 70], but also in other biological systems in which nitrergic activity and prenylation have important regulatory functions such as in cardio-vasculature or cancer signaling [71].
Flies were raised on standard maize media at 25 °C at a 12-h LD cycle. The elav-Gal4 [C155] driver was obtained from the Bloomington Stock Center (Indiana, US). The UAS-RNAi lines (GCLm [CG4919], GCLc [CG2259], and Fnta [CG2976]) were purchased form the Vienna Drosophila Resource Centre (VDRC). The use of the UAS-Gal4 bipartite expression system to drive pan-neuronal expression excludes potential postsynaptic effects. The elav-Gal4 driver (female flies) and the UAS responder lines (male flies) were crossed to obtain offspring expressing the genes of interest and w1118 were used as Ctrls.
The fluorescent Ca2+ sensor GCaMP5 was tethered to the plasma membrane with an N-terminal myristoylation (myr) sequence as described previously [72]. The UAS-myrGCaMP5 and cpxSH1 null mutant lines were provided by Troy Littleton (MIT, Cambridge, MA) [73]. GCaMP5 was expressed in glutamatergic neurons (OK371-Gal4; UAS-GCaMP5). cpx expression levels are shown for w1118 and cpxSH1 larvae in S9B and S9C Fig.
UAS-fdh31 (expression of fdh homologue of mammalian GSNOR/ADH-5) and UAS-fdhri34/25 (expression of fdh RNAi) mutant transgenic lines were kindly provided by Li Liu Institute of Biophysics, Chinese Academy of Sciences, Beijing, China) [51].
NOSΔ15/NOSC lines were provided by Patrick O’Farrell (UCSF, San Francisco, CA). NOSΔ15 deletion removes sequences encoding residues 1–757, encompassing the entire oxygenase domain and including regions that bind the catalytic heme and the substrate rendering the lines NOS “null” [33, 34].
The syn-null mutant transgenic line (Syn97) was generously provided by Erich Buchner (Universitätsklinikum Würzburg, Germany) [44]. UAS-EGFP-cpx and UAS-EGFP-cpx1257 transgenic lines were kindly provided by Fumiko Kawasaki (Penn State University, PA) [16]. UAS-GCLm and UAS-GCLc transgenic lines were provided by William C. Orr (Southern Methodist University Dallas, TX).
cDNAs encoding for cpx 7A was a gift from Troy Littleton and used as a template for downstream PCRs. Cys 140 of cpx7A isoform was mutated to tryptophan to generate S-nitrosylation mimic mutant. PCR products, which include XhoI and XbaI restriction sites, were cloned into the pJFRC2 vector [74]—a gift from Gerald Rubin (Addgene plasmid no. 26214)—by standard methods. The resulting constructs were injected into attP40 Drosophila strains. The resulting transgenic lines (Dmcpx7AC140W and WT Dmcpx) were crossed into a cpxSH1 background [7] using standard balancing techniques. The FlincG3 ORF was amplified from pTriEx4-H6-FGAm (FlincG3) (Addgene plasmid no. 49202) and the resultant PCR product cloned into pUASTattB by the Protein Expression Laboratory (PROTEX), University of Leicester. Microinjection of the pUASTattB plasmid was performed by the University of Cambridge, Department of Genetics Fly Facility.
TEVC recordings were performed as described previously [75]. Sharp-electrode recordings were made from ventral longitudinal m6 in abdominal segments 2 and 3 of third instar larvae using pClamp 10, an Axoclamp 900A amplifier and Digidata 1440A (Molecular Devices, US) in hemolymph-like solution 3 (HL-3) [76]. Recording electrodes (20–50 MΩ) were filled with 3 M KCl. mEJCs were recorded in the presence of 0.5 μM tetrodotoxin (Tocris, UK). All synaptic responses were recorded from muscles with input resistances ≥4 MΩ, holding currents <4 nA at −60 mV and resting potentials more negative than −60 mV at 25 °C, as differences in recording temperature cause changes in glutamate receptor kinetics and amplitudes [77]. Holding potentials were −60 mV. The extracellular HL-3 contained (in mM): 70 NaCl, 5 KCl, 20 MgCl2, 10 NaHCO3, 115 sucrose, 5 trehalose, 5 HEPES, and 1.5 CaCl2 (0.5–3.0 mM in Fig 3 and S3 Data, as specified). Average single eEJC amplitudes (stimulus: 0.1 ms, 1–5 V) are based on the mean peak eEJC amplitude in response to 10 presynaptic stimuli (recorded at 0.2 Hz). Nerve stimulation was performed with an isolated stimulator (DS2A, Digitimer). Paired-pulse experiments were performed by applying 5 repetitive stimuli (0.2 Hz) at different intervals (20, 40, 100, 200 ms) for each cell at each ISI. All data were digitized at 10 kHz and for miniature recordings, 200-s recordings, we analyzed to obtain mean mEJC amplitudes, decay, and frequency (f) values. QC was estimated for each recording by calculating the ratio of eEJC amplitude/average mEJC amplitude, followed by averaging recordings across all NMJs for a given genotype. mEJC and eEJC recordings were off-line low-pass filtered at 500 Hz and 1 kHz, respectively. Materials were purchased from Sigma-Aldrich (UK) unless otherwise stated.
Approximately 40 eEJCs were elicited at different [Ca2+]e, ranging from 0.5 to 3 mM to give mean eEJC amplitudes (I). The mean eEJC is given by I = Npvrq [45], with N being the number of independent release-ready vesicles, pvr the vesicular release probability, and q the quantal size at each given [Ca2+]e. The eEJC variance was calculated as previously described [45]. The plots of the variance-mean were obtained for each cell and fitted with the parabolic function Var(I) = I2/N + qI that was a constraint to pass through the origin. Upon fitting the parabola, pvr and q were calculated using the equations: q = A/(1+CV2) and pvr = I(B/A)(1+CV2) where CV2 is the coefficient of variation of the eEJC amplitudes at a given [Ca2+]e concentration calculated as CV2 = (eEJCs standard deviation/mean amplitude)2; A and B were obtained from the fitting parameters. Estimated values were not corrected for variability in mEJC amplitude distributions or latency fluctuations.
Ca2+ cooperativity was assessed by plotting eEJC amplitudes over [Ca2+]e and fitted with the Hill equation (mean eEJC amplitude plotted versus different [Ca2+]e: eEJC([Ca2+]) = eEJCmax[1+(EC50/[Ca2+])slope]−1), yielding the Hill slope as a measure of Ca2+ cooperativity.
The apparent size of the RRP was probed by the method of cumulative eEJC amplitudes [78]. Muscles were clamped to −60 mV and eEJC amplitudes during a stimulus train (50 Hz, 500 ms [of a 1-s train]) were calculated as the difference between peak and baseline before stimulus onset of a given eEJC. Receptor desensitization was not blocked as it did not affect eEJC amplitudes, because a comparison of the decay of the first and the last eEJC within a train did not reveal any significant difference in decay kinetics. The number of release-ready vesicles (N) was obtained by back extrapolating a line fit to the linear phase of the 500-ms cumulative eEJC plot (the last 200 ms of the train) to time zero. N was then obtained by dividing the cumulative eEJC amplitude at time zero by the mean mEJC amplitude recorded in the same cell. To calculate the QC in the train, we used mean mEJC amplitudes measured before the train.
Third instar larvae were dissected in ice-cold PBS then fixed in 4% paraformaldehyde. After permeabilization with PBS-0.1% Triton (PBS-T) and blocking with PBS-T containing 0.2% bovine serum albumin (BSA) and 2% normal goat serum, larval fillets were incubated at 4 °C overnight in solutions of primary antibody. The following antibody dilutions were used: NC82 (supernatant) anti-Brp (Bruchpilot) 1:200, cpx (1:500), syntaxin (1:200), synaptotagmin (1:200), and GFP (1:200). After 3 × 10 min washes in PBS-T, larvae were incubated with AlexaFluor 488 goat anti-HRP (Jackson Immuno Research) and AlexaFluor 546 goat anti-mouse 1:500 dilution for 90 min at room temperature. Larvae were mounted using Vectashield mounting medium (Vector Labs) and NMJ 6/7 (segments A2 and A3) images were acquired with a Zeiss laser-scanning confocal microscope (LSM 510, Zeiss). Image analysis was performed with ZEN (Zeiss) and Volocity 6.3 software.
Images were acquired on a Leica TCS SP8 system attached to a Leica DMi8 inverted microscope (Leica Microsystems). Excitation light (488 nm for AlexaFluor488 or 561 nm for AlexaFluor568) was provided by a white light laser with a repetition rate of 80 MHz. Images were acquired using a 100× 1.4 NA oil immersion objective and fluorescence was detected through a bandpass of 495–550 nm (AlexaFluor488 detection) or 570–650 nm (AlexaFluor 561 detection). Gated STED imaging of samples was achieved through use of 592-nm and 660-nm depletion lasers with a time gate set to 1.8–8 ns using the Leica STED 3X system. All images were acquired with 32-line averages and 22 × 22 nm pixel size.
Images were taken using an LSM 510 confocal microscope (Zeiss). The size of the bleaching area was optimized as shown previously [56]. Bleaching areas were selected within each bouton (about 2.5 μm2) and images acquired every 10 s. Data were fitted with a single exponential to reveal tau values of fluorescence recoveries.
The assay was performed as described [79]. Briefly, dissected third instar larvae were fixed in Bouin’s solution for 15 mins on ice, washed in PBT (PBS with 0.1% Triton) 3 times for 10 min each and blocked in PBT/1% BSA for 1 h. Larvae were incubated overnight at 4 °C in mouse and rabbit antibodies against the 2 proteins of interest, diluted in PBT/1% BSA. Primary antibodies used were anti-rabbit cpx (Littleton), anti-rabbit GFP (Abcam), anti-mouse Brp (Developmental Studies Hybridoma Bank [DSHB]), anti-mouse syntaxin (DSHB), and anti-mouse Synaptotagmin (DSHB). All antibodies were used at 1:200 dilution. The next day, PLA probe binding, ligation, and amplification steps were performed as described [79]. Before mounting, larvae were counterstained with AlexaFluor 488 goat anti-HRP (Jackson Immuno Research) at 1:500 dilution for 40 mins. PLA signals were only measured within the HRP signals. PLA signal and NMJ volumes of z-stack images were analyzed in Volocity 6.3. PLA signals were only measured within the HRP signals. All PLA signals were expressed relative to total NMJ volume (S10 Fig and S9 Data).
A plasma membrane targeted eYFP CAAX protein was constructed by fusing the last 15 amino acids of Human K-Ras isoform b with the C-terminus of eYFP. A short linker sequence GTMASNNTASG was inserted between the last amino acid of eYFP and the membrane targeting CAAX sequence. The resulting construct was subcloned into expression vector pcDNA5 frt and verified by DNA sequencing.
HEK293 FT cells were plated on poly-d-lysine coated glass coverslips in 6 well plates and transfected with 0.5 g eYFP CAAX per well using polyethylenimine (PEI) at a ratio of 1 g DNA to 6 g PEI. Prior to imaging, cells were treated for 12 h with the NO donor DETA-NONOate or a combination of the farnesyl transferase inhibitor L-744,832 (20 μM) and the geranylgeranyltransferase I inhibitor GGTI-298 (20 μM). Cells were then washed 3 times with PBS and fixed for 15 min with 4% paraformaldehyde. Coverslips were mounted on glass microscope slides with VectaShield H1500 and observed using a Zeiss laser scanning confocal microscope.
Animals were kept in the dark 3 h before removing the retinas in order to decrease basal levels of protein nitrosylation. Retinas were kept in DMEM (Gibco 31053–028) with protease inhibitors (Complete) and treated with NO donors (GSNO and PAPA-NONOate, 20 μM) for 40 min at room temperature and protected from light. The biotin-switch assay was performed with the S-nitrosylated Protein Detection kit (Cayman Chemical, 10006518) in the dark. Bradford assay was performed and equal amount of proteins were incubated with Streptavidin beads (Sigma) overnight. Western blot was performed with cpx 3 antibody (Synaptic Systems), 1:1,000.
Third instar larvae were “filleted” in phosphate-buffered saline at room temperature and then fixed in 2% (wt/vol) glutaraldehyde in 0.1 M sodium cacodylate buffer (pH 7.4) at 4 °C overnight. They were postfixed with 1% (wt/vol) osmium tetroxide/1% (wt/vol) potassium ferrocyanide for 1 h at room temperature and then stained en bloc, overnight, with 5% (wt/vol) aqueous uranyl acetate at 4 °C, dehydrated and embedded in Taab epoxy resin (Taab Laboratories Equipment Ltd, Aldermaston, UK). Semi-thin sections, stained with toluidine blue, were used to identify areas containing synaptic regions (m6/7 in regions A2/A3). Ultra-thin sections were cut from these areas, counterstained with lead citrate, and examined in an FEI Talos transmission electron microscope (FEI Company [Thermo Fisher Scientific Inc.], Hillsboro, OR). Images were recorded using an FEI Ceta-16M CCD camera with 4k × 4k pixels. SV measurements were made using ImageJ software. A total of about 500–900 SVs were measured in 5–10 boutons from 3 animals per genotype.
Wandering third instar larvae expressing presynaptic UAS-myrGCaMP5 or UAS-GCaMP5 using the pan-neuronal C155 or glutamatergic neuronal OK371 driver, respectively, were dissected in low Ca2+ HL-3 saline (0.2 mM CaCl2) at room temperature. The motor nerves were carefully snipped below the ventral nerve cord, and the CNS was removed. The preparation was washed several times with HL-3 containing 1.5 mM Ca2+. Nerve stimulation was performed with an isolated stimulator (DS2A, Digitimer) and images were recorded before, during (2–6 s in a train at 60 Hz) and after the stimulation period (8 s) in HL-3 containing 3 mM Ca2+ or during 15 s in a 20-s train at 20 Hz at indicated Ca2+ levels in the presence of 5 mM L-glutamic acid. We acquired images at a rate of 1 image per 4 s using a Zeiss laser-scanning confocal microscope (LSM 510 Meta; Zeiss) with a 63× 1.0 NA water immersion objective (Zeiss). Excitation was set at 488 nm (Argon laser) using a dichroic mirror 490 nm and a bandpass filter 500–550 nm. Low sampling rates were sufficient to investigate Ca2+ plateau levels during the 8-s stimulation periods [80]. A single confocal plane of muscle pair 6/7 NMJ in segments A2 or A3 was imaged to establish a baseline. Small z-drifts were manually corrected during the imaging session. Imaging sessions in which significant movement of the muscle occurred were discarded. Images were analyzed using Volocity 6.3 Image Analysis software (PerkinElmer). Single bouton fluorescence intensities were measured (average within a bouton) and bouton ΔF/F0 values were averaged for each NMJ.
NMJs of larvae expressing UAS-FlincG3 presynaptically were imaged as described above to measure GCaMP fluorescence. To prevent cGMP breakdown by PDE activity, preparations were incubated with 10μM Zap prior to imaging.
High resolution respirometry was performed with an Oroboros O2K oxygraph (Oroboros Instruments Ltd.). For each measurement, 3 third instar larvae were homogenized in 100 μL of respiration buffer MiR05 [81]. Leak state respiration was measured after adding 5 mM of pyruvate, 2 mM of malate, and 10 mM of glutamate. Oxphos capacity supported by Complex I was measured after addition of 1.25 mM ADP. After addition of 10 mM succinate, Oxphos capacity supported by both Complex I and Complex II were measured. Free Oxphos capacity was calculated as the difference Oxphos–Leak. Respiratory Ctrl ratios (RCRs) were calculated as the ratio Oxphos/Leak.
Larval brains (30 per condition) were isolated and assessed for cGMP production. Briefly, brain extracts were diluted 5-fold in 100 mM sodium acetate, pH 6.2, and acetylated by consecutive addition of triethylamine (10 μL) and acetic anhydride (5 μL) and used in the radioimmunoassay [82] within 60 min. Cyclic GMP standards (100 μL; 0–4 nM) were treated identically. Acetylated samples (100 μL) were mixed with 2′-O-succinyl 3-[125I]-iodotyrosine methyl ester cyclic GMP (GE Healthcare, IM107) (50 μL, about 3,000 d.p.m. made up in 50 mM sodium acetate, 0.2% BSA, pH 6.2), and 100 μL of anti-cyclic GMP antibody (GE Healthcare, TRK500; diluted in 50 mM sodium acetate, 0.2% BSA, pH 6.2). Samples were intermittently vortex mixed during a 4-h incubation at 4 °C. Free and bound cyclic GMP was separated by charcoal precipitation with 500 μL of a charcoal suspension (1% [w/v] activated charcoal in 100 mM potassium phosphate, 0.2% BSA, pH 6.2). After vortex mixing for 5 min, samples were centrifuged (13,000 × g, 4 min, 4 °C) and radioactivity determined in an aliquot of supernatant (600 μL). Unknown values were determined from the cyclic GMP standard curve using GraphPad Prism 7 (GraphPad Software Inc., San Diego, CA). Data points represent 2 measurements of 30 brains for each condition.
NO donor solutions were made freshly from stock solutions on the day and working solutions (200 μM sodium nitroprusside [SNP] and 5 μM PAPA-NONOate, each releasing about 200 nM NO) [29] were kept on ice for up to 6 h. All experiments to assess NO signaling were made between 40 and 60 min of NO exposure (NO: 200 μM SNP, 20 μM PAPA-NONOate; presented data comprise responses following incubation with either donor as they are not different from each other [Student t test, p > 0.05]; 500 μM SNP was used in Figs 7A, 7B, 8E and 8F [S7 and S8 Data]). Incubations with drugs: Zap (PDE inhibitor), ODQ (sGC inhibitor), L-744,832, and GGTI-298 (FTase inhibitors) incubation for 1 h; both block FTase with an IC50: 1.8 nM and IC50: 203 nM, respectively [83]. Drugs were purchased from Tocris or Sigma.
Statistical analysis was performed with Prism 6.3 and 7 and InStat 3 (Graphpad Software Inc., San Diego, CA). Statistical tests were carried out using an ANOVA test when applicable with a posteriori test (1-way ANOVA with Tukey’s multiple comparisons test) or unpaired Student t test, as indicated. Data are expressed as mean ± SEM where n is the number of boutons, NMJs, or larvae as indicated and significance is shown as *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001.
|
10.1371/journal.pntd.0003210 | Cutaneous Leishmaniasis and Sand Fly Fluctuations Are Associated with El Niño in Panamá | Cutaneous Leishmaniasis (CL) is a neglected tropical vector-borne disease. Sand fly vectors (SF) and Leishmania spp parasites are sensitive to changes in weather conditions, rendering disease transmission susceptible to changes in local and global scale climatic patterns. Nevertheless, it is unclear how SF abundance is impacted by El Niño Southern Oscillation (ENSO) and how these changes might relate to changes in CL transmission.
We studied association patterns between monthly time series, from January 2000 to December 2010, of: CL cases, rainfall and temperature from Panamá, and an ENSO index. We employed autoregressive models and cross wavelet coherence, to quantify the seasonal and interannual impact of local climate and ENSO on CL dynamics. We employed Poisson Rate Generalized Linear Mixed Models to study SF abundance patterns across ENSO phases, seasons and eco-epidemiological settings, employing records from 640 night-trap sampling collections spanning 2000–2011. We found that ENSO, rainfall and temperature were associated with CL cycles at interannual scales, while seasonal patterns were mainly associated with rainfall and temperature. Sand fly (SF) vector abundance, on average, decreased during the hot and cold ENSO phases, when compared with the normal ENSO phase, yet variability in vector abundance was largest during the cold ENSO phase. Our results showed a three month lagged association between SF vector abundance and CL cases.
Association patterns of CL with ENSO and local climatic factors in Panamá indicate that interannual CL cycles might be driven by ENSO, while the CL seasonality was mainly associated with temperature and rainfall variability. CL cases and SF abundance were associated in a fashion suggesting that sudden extraordinary changes in vector abundance might increase the potential for CL epidemic outbreaks, given that CL epidemics occur during the cold ENSO phase, a time when SF abundance shows its highest fluctuations.
| In this study we analyze data on sand fly (SF) abundance and cutaneous leishmaniasis (CL) cases from Panamá. We asked whether weather patterns and climatic variability could have an impact on vector abundance that is ultimately reflected in CL transmission. We found that large epidemics of CL occur during the cold phase of El Niño Southern Oscillation (ENSO), with regular cycles coinciding with the oscillation patterns of ENSO. By contrast, and counterintuitively, we found that during the hot and cold phase of ENSO the average number of SF was reduced when compared with the normal phase of ENSO. Nevertheless, the cold ENSO phase also shows the largest variability in vector abundance, which is a likely indicative of sudden and extraordinary increases in SF abundance. We, therefore, propose that marked SF abundance changes, triggered by anomalous weather patterns associated with ENSO, likely play a major role in shaping CL interannual transmission cycles in Panamá.
| Cutaneous leishmaniasis (CL) is a major neglected tropical disease [1] with a complex ecology [2], whose transmission, in the New World, requires the co-existence of vectors, reservoirs and humans [3], [4]. In Panamá, detailed studies on the eco-epidemiology of the disease [2], [5] described parasitological aspects of reservoirs [6] and vectors [7] and the environmental context of parasite-reservoir-vector interactions [8]. These studies set several landmarks for understanding New World CL epidemiology, including the demonstration of two toed sloths, Choloepus hoffmanni [9], [10] and other mammals [11], [12] as reservoirs of Leishmania spp parasites. Insights on sand fly vector ecology included: catholic bloodfeeding patterns in dominant vector species [13], [14], very limited dispersal of flying adults [15], high sensitivity of larval biology to environmental factors [16], [17], [18], [19], [20], differential adult resting behavior with species segregating along tree height [21], large diversity of co-occurring sand fly species across CL transmission foci [22], [23], [24] and heterogeneities in species composition across landscape gradients [23]. Currently, the most common parasite causing CL in Panamá is Leishmania panamensis [25], [26] and the resurgence and exacerbation of disease transmission has led to renewed efforts aimed at improving vector control [27] as a measure to reduce transmission at emerging transmission hotspots [25]. However, larger questions about what is driving the resurgence of the disease, and how to best predict epidemic outbreaks remain unanswered [28], [29].
The longitudinal nature of eco-epidemiological studies on CL in Panamá [2] revealed interannual patterns of variability in reservoir infection and abundance [10] and the sensitivity of sand fly density to weather fluctuations [30]. Nevertheless, no study has addressed if large scale meteorological phenomena, such as El Niño Southern Oscillation (ENSO), are associated with interannual CL epidemic cycles, as observed in neighboring Costa Rica [28], nor the impacts of ENSO on Sand Fly populations. These issues are of special interest, given the increased reports of direct and indirect impacts of climatic variability patterns associated with global warming on the disease transmission, both in the New World [32], [33], [34], [35], [36], [37], [38] and the Old World [39], [40]. Here, we employ an 11 years long (2000–2010) monthly time series recording CL cases in the República de Panamá to investigate seasonal and interannual cycles on this disease. During this time at the coarse spatial scale forest cover has marginally increased, yet locally some areas have seen deforestation [41], an activity usually linked with CL outbreaks [42]. We specifically ask which climatic factors are associated with seasonal and interannual disease cycles, considering local temperature and rainfall records and sea surface temperature 4 (SST4), an index associated with ENSO activity in the region. We also ask whether dominant sand fly (SF) vector species undergo marked abundance changes associated with ENSO, that are ultimately reflected in CL transmission. We employed seasonal autoregressive models and cross wavelet coherence analysis to depict the association of CL with climatic factors, and found that while seasonal CL patterns were mainly associated with temperature and rainfall, interannual cycles of the disease were associated with SST4. Moreover, SST4 was also associated with interannual cycles in temperature and rainfall. SF vectors showed marked abundance changes associated with ENSO, where abundance in general decreased during the hot and cold phases of ENSO. Our results highlight both the general association of ENSO and weather patterns with CL dynamics in Central America, and how changes in SF vector abundance associated with ENSO might play a role in the emergence of CL epidemics.
Monthly CL cases were compiled by the Epidemiology Department of Panamá's Ministry of Health for the period January 2000 – December 2010. Briefly data consisted of cases clinically diagnosed [43], and often confirmed by the microscopic examination of skin lesson scrappings/biopsies, Montenegro skin tests (MST) [44] or Indirect Immuno-Fluorescent Agglutination Tests (IFAT) [26]. Data (Figure 1A) were collected from all the health facilities administered by Panamá's Ministry of Health and all data came from passive case detection. Reports were then compiled at the health area level (the operational geographical units of Panamá's Ministry of Health which are slightly different from Panamá's provinces and autonomous indigenous comarcas). Slightly over 80% of the cases came from West Atlantic Panamá, the area facing the Caribbean Sea, west of the Panamá Canal up to the border with Costa Rica [25]. Representative samples from all over Panamá [25] indicate that over 95% (at least 90% for each reporting health area) of the CL cases in the time series are due to Leishmania panamensis. Cases due to Leishmania mexicana, Le amazonensis and Le colombiensis continue to be rare and sporadic, as observed in earlier epidemiological studies in Panamá [5]. Moreover, all CL cases observed in migrants that likely acquired the infection in Panamá have been typified as Le panamensis [45], [46]. Temperature data were obtained from the US National Oceanic and Atmospheric Administration, NOAA (ftp://ftp.ncdc.noaa.gov/pub/data/ghcn/v2/) for Tocumen (Station 787920), Albrook (Station 783842 and 788060), Bocas del Toro (Station 787935), David (Station 787930) and Santiago (Station 787950). These daily time series were averaged per month, considering all values and the monthly maxima and minima (Figure 1B). Rainfall data, an average considering all weather stations, were obtained from ETESA, Panamá's electrical company (Figure 1C). Monthly SST 4, often referred as El Niño 4 (Figure 1D), was obtained from (http://www.cpc.ncep.noaa.gov/data/indices/ersst3b.nino.mth.81-10.ascii). The NOAA data for SST4 were collected from the area delimited by 5°North-5°South and 160°East-150°West of the Pacific Ocean. We also classified each month from the time series into the ENSO phases following the Oceanic Niño Index (ONI) from NOAA (http://www.cpc.ncep.noaa.gov/products/analysis_monitoring/ensostuff/ensoyears.shtml). Briefly, the ONI is estimated by detrending the monthly Sea Surface Temperature 3.4 (collected in the area defined by 5°N-5°S, 120°-170°W) over periods of 30 years, with residues with a value over 0.5 corresponding to the warm (or hot) ENSO phase, residues below −0.5 corresponding to the cold ENSO phase and residues in the interval [−0.5,0.5] being normal conditions [47].
SF abundance data came from six studies performed either at Universidad de Panamá or Instituto Conmemorativo Gorgas de Estudios de la Salud (ICGES). All these studies were performed within the República de Panamá (Figure S1A). All the studies used a common sand fly sampling method, which consisted in the use of unbaited light traps placed at 1.5–2.0 m height above ground, with traps operating from 6 pm to 6 am (a sampling effort referred as trap-night), and placed in three well defined eco-epidemiological environments: (i) domiciliary for samples from inside houses; (ii) peridomiciliary for samples collected in a radius of 100 m from a house; and (iii) forested areas for samples collected in areas with primary/secondary vegetation and outside a 100 m radius from a house. This standardized sampling, in principle, renders the comparison of the different datasets plausible. In most of the studies sand fly sampling was done with the purpose of describing the fauna at endemic locations [48], [49], [50], as part of the evaluation of SF control trials [27] and we also report unpublished data from entomological surveillance of endemic CL locations in 2007, 2009, 2010 and 2011. We focused our analysis in the abundance of Lutzomyia trapidoi, Lu gomezi and Lu panamensis (Figure S1B), the dominant vector species for CL in the República de Panamá [2], [7]. This was done given the proven vectorial role of these species [2], [7], and the lack of abundant records for other sand fly species. Figure S1C shows the eco-epidemiological environments sampled in each location and Figure S1D the year when samples were collected at each location. A total of 12580 SF were collected over 640 trap-nights. In all the studies, the spermathecae or genitalia were inspected for species identification using the key by Young and Duncan [51].
Figure 2 shows CL case seasonality. Figure 2A shows how CL cases peak at the beginning of the rainy season in Panamá, in April and May [41]. From January 2000 to December 2010 a total of 26140 CL cases were recorded in Panamá. Figure 2B shows how ENSO phases have been uniformly distributed across the year, and how peaks in CL cases tend to occur during the cold ENSO phase, a pattern also shown in Figure S2 when pooling all months. Given the monthly nature of our data and the focus on the total number of reported CL cases, with observations spanning 132 months, we employed a battery of tools for time series analysis in the time domain, including temporal correlation functions and seasonal autoregressive models, and time-frequency domain, i.e., wavelets [65]. Figure S3 show the different correlation functions employed to fit seasonal autoregressive models to the CL time series. The PACF (Figure S3A) suggested a seasonal autoregressive structure in the CL time series, where up to the first three lags and the seasonal lag (12 months) were associated with CL number at any time. The autoregressive pattern was also suggested by the ACF (Figure S4). Indeed a 3rd order seasonal autoregressive model was selected as the best null model (Table S1) and employed for pre-whitening climatic covariates and for subsequent estimation of Cross Correlation Functions (CCFs) between CL and climatic covariates (Figure S3B, S3C, S3D, S3E, S3F). CL was negatively correlated with SST4 at lags 4 and 5 and positively correlated with temperature at lag 12 (Figure S3B). CL was negatively correlated with rainfall (Rain) at lag 15 (Figure S3C). CL was positively correlated with average (Temp) and maximum temperature at lag 13 (Figure S3D & S3E) autonomous from minimum temperature (Figure S3F). Thus, based on the significant correlations observed in Figure S3 the following null model was fitted:(1)
Where μ is the mean value of the time series, φ's indicate autoregressive parameters, β, γ and α are parameters for climatic covariates and ε indicates an error normally distributed and with variance . After several rounds of backward elimination (Table S1) the following model was selected as best:(2)
Whose parameter estimates are presented in Table 1. The impacts of Temp(t-13) and SST4(t-12) on CL were positive, in contrast with SST4(t-5) which was negative, and assumptions about the error were met, ensuring a sound inference. The process of model selection (Table S1) left out variables that were strongly correlated (Figure 3) with the ones present in equation (2). For example, SST4(t-5), SST4(t-4) and Rain (t-15) were positively associated between them (Figure 3) and negatively with CL(t). Thus, SST4(t-5) was able to capture a common impact of ENSO on both rainfall and CL. Similarly, Temp(t-13) and CL(t-12) were positively associated between them and with CL(t) (Figure 3), rendering the inclusion of the former, in addition to SST4(t-5), enough to account for seasonality in CL(t).
The cross wavelet coherence analysis (Figure 4) confirms the outcome of autoregressive models, showing that CL was associated with: SST4 (Figure 4A), rainfall (Figure 4B) and temperature (Figure 4C) during the study period, all the three variables associated with the inter-annual variability in CL, specifically cycles with period between 2–4 years, the latter two also associated with the seasonal cycles (period of 1 year). Similarly, rainfall (Figure 4D) was associated with SST4 at inter-annual scales; and both rainfall (Figure 4D) and temperature (Figure 4E) were seasonally associated with SST4. In synthesis, ENSO, measured through SST4, shows an imprint on CL transmission robustly revealed by both the autoregressive model and the cross wavelet coherence analysis, where SST4 also impacts Rainfall and Temperature, which are, as well, associated with CL.
Data on SF vectors showed a common pattern where Lu gomezi (Figure 5A), Lu trapidoi (Figure 5B) and Lu panamensis (Figure 5C) reduced their average abundance during the hot and cold ENSO phases. Nevertheless, these three dominant SF vector species showed an increased variability in abundance during the cold ENSO phase, i.e., the boxes in the boxplots were longer, and large outliers were more frequent during the cold ENSO phase. After controlling for differences related to heterogeneity in eco-epidemiological sampling environment, and for seasonality associated with different sampling months, the GLMPMs for Lu trapidoi and Lu panamensis showed that abundance reductions were significant during both the hot and cold ENSO phases (P<0.05), but not for Lu gomezi, (Table 2). The abundance across eco-epidemiological environments showed a similar pattern for Lu gomezi (Figure 5D) and Lu trapidoi (Figure 5E), where abundance was slightly larger in domiciliary than in peridomiciliary or forest environments, a pattern statistically significant (Table 2). By contrast, Lu panamensis (Figure 5F) was, respectively, about 9 and 2.5 times more abundant in forests and peridomiciles than inside the houses (P<0.05, Table 2). Table 2 shows that, in general, the unexplained variance in the SF abundance GLMPMs associated with spatial heterogeneity (Location Var) was about one order of magnitude higher than the unexplained temporal variance (Year Var). Similarly, model selection with AIC and BIC showed that it was not necessary to nest the spatial random effects within each study (Table S2).
Monthly CL case records were positively associated with the abundance of Lu gomezi (Figure 5SA) and Lu trapidoi (Figure 5SB) with the CCF peaking at three months lag, i.e., after a peak in SF vector abundance there was a peak in CL cases three months later. For Lu gomezi (Figure 6A) the 3 month-lagged pattern of association with CL cases was mainly linear, but for Lu trapidoi (Figure 6B) the number of cases seemed to flatten out at high SF abundances. By contrast, monthly CL cases and Lu panamensis abundance were negatively and significantly associated with a one and a two month lag (Figure 5SC). As expected from the CCF, the association between CL cases and the 3 month-lagged abundance of Lu panamensis had no clear pattern (Figure 6C). When the abundance of Lu gomezi and Lu trapidoi were added together the maximum positive association occurred at three months of lag, but the positive association was also significant at 2 and 4 months of lag (Figure 5SD). The 3 month-lagged pattern of association between CL cases and the combined abundance of Lu gomezi and Lu trapidoi (Figure 6D) resembled the one observed for Lu trapidoi alone (Figure 6B) with the number of cases flattening out at high SF abundance.
A general criticism for studies addressing the impact of climate change on vector-borne disease transmission is that little to no attention has been given to what changes, if any, occur in entomological risk patterns [66], e.g., what happens to the vectors during ENSO?. Here, we tried to address that knowledge gap for CL by going beyond the description of the association between ENSO and weather patterns and CL epidemics in Panamá, we inquired whether SF vectors change their abundance during ENSO. Our data showed that interannual cycles of CL transmission, as inferred from a CL case time series from Panamá, were associated with ENSO, a pattern observed in neighboring Costa Rica [29], and also observed for malaria in the República de Panamá [67], highlighting the impacts of ENSO on vector-borne diseases in Central America. Large CL Epidemics were observed during the cold ENSO phase or shortly after it, where the delay might reflect the delay between transmission and clinical symptoms in American CL [29], [43], [68], a possibility further re-inforced by the 3 month delayed association between vector abundance and CL incidence.
Seasonal (intrannual time scale) changes in CL transmission were associated with temperature, a weather component with low variability, i.e., low amplitude fluctuations, in the Panamá isthmus. This pattern may make sense in light of Schmalhausen's law, the notion that biological systems are more sensitive to small changes in low variability factors when stressed by other environmental components [65]. SF population dynamics may become more sensitive to changes in temperature given their need to cope with more marked changes in other weather factors, e.g., rainfall which has a more marked seasonal imprint than temperature in Panamá [41]. Indeed, the pattern of higher sensitivity to changes in temperature in places with marked seasonality in rainfall has been observed for other disease vectors [69]. Here, we want to also note that large CL epidemics occurred during or shortly after the cold ENSO phase, a time when, on average, SF vector abundance is the smallest across ENSO phases. Nevertheless, as observed in the raw data, the cold ENSO phase is a time when SF vectors are also prone to show extremely large abundance records per trap-night, which might reflect insect population outbreaks [70], [71], i.e., sudden extraordinary increases in vector abundance [72]. Thus, the occurrence of large CL epidemics during or shortly after the cold ENSO phase might indicate a role for SF vector outbreaks on CL epidemics. Indeed, a detailed study in Venezuelan village showed that CL cases in an endemic village were associated with vector abundance [63], [64]. Nevertheless, the abundance of SF vectors in those studies didn't show potential “outbreaks” [63], [64] in SF abundance, thus not allowing to assess whether CL case incidence flattens out with large vector abundance. This information is necessary to properly understand the role of climate on the entomological risk for CL transmission [29], [62]. This goal will require new longitudinal studies on vector abundance in Panamá [30] where Leishmania spp infection in the vectors is also tracked [40], in order to better understand the relationship between vector abundance and vector infection, since constant or nearly constant infections rates in vectors have different implications to understand the role of vectors on transmission patterns. For example, if infection rate decreases with vector abundance, such a density-dependent pattern might partially explain the flattening relationship between vector abundance and cases, such as observed with bloodfeeding success by SF vectors, which decreases with density [73]. Nevertheless, the flattening can also emerge by, or in synergy with, the regulation in the recruitment of susceptible hosts [29] and/or the zoonotic reservoirs, some of which might also experience population outbreaks with ENSO [74], [75] which can ultimately be linked to ENSO mediated changes in the resources sustaining wildlife reservoir populations in the neotropics [76], [77]. Indeed, the most studied wildlife spp reservoir in Panamá, the two toed sloth, has shown relatively large interannual fluctuations in Leishmania (Viannia) sp. infection [10].
Our results also support the major role of Lu gomezi and Lu trapidoi as dominant SF vectors of CL in Panamá [2], [43], given their ubiquity across domestic, peridomestic and forest environments. This ubiquity has implications for the role of these species in both the transmission to humans and as bridge of pathogens across vertebrate Leishmania spp hosts and eco-epidemiological environments [78]. By contrast, Lu panamensis was mainly present in forest environments, which suggests that it might not be heavily involved in domiciliary/peridomiciliary CL transmission, a possibility also put forward by recent studies on spatial patterns of human infections [43] and dogs (unpublished data), which were mainly associated, respectively, with Lu gomezi and Lu trapidoi abundance, but not Lu panamensis. Nevertheless, our inferences are limited given our focus on acknowledged dominant vectors species, which is an approach that potentially biases the identification of other vectors present at CL transmission foci, a problem that can only be solved when studying the whole SF community [79].
Finally, our study has some limitations related to the nature of the data and its countrywide geographical scale. Although our CL and SF data have a consistent quality, there is ample room to improve CL and SF surveillance. For CL surveillance, an urgent need is to standardize diagnostics across the country using sensitive and specific methods [1], [26]. Even if it is impossible to standardize diagnostics at the health post level, an effort should be made to estimate the error in diagnostics, as done for malaria, where all clinically diagnosed cases are confirmed at the ICGES, and quality controls on the specificity of diagnosis are equally performed [67]. The non-spatial nature of our analysis precludes the identification of transmission hotspots requiring attention [80] or zones were biases in case report might be occurring [81]. Similarly, the role that patterns of socio-economic inequity might have in the impacts of climate change and weather variability on CL transmission [32] cannot be estimated. Nevertheless, the relative low variability in rainfall patterns across human inhabited zones in Panamá [41] suggests that a country-wise analysis is a sound method to make inferences about the relationship between CL transmission, ENSO and weather patterns. The CL cases time series showed no trend which allowed us to ignore a denominator for the cases, nevertheless we cannot assert whether the lack of trends is due to stationary population patterns in the population at risk, or if they reflect other unknown changes in the populations at risk and/or transmission. This point could be further clarified by the establishment of health demographic surveillance systems that could both improve the understanding of disease transmission patterns and the demography of populations living in CL endemic areas. Similarly, SF monitoring can be improved and more systematically done at endemic areas. This is an issue of major importance, since, given the delay between transmission and clinical CL, an early prediction of CL epidemics will be more robust if based on the monitoring of SF abundance and Leishmania sp. infection [29]. Although each SF abundance estimate came, on average, from ten trap-nights, locations were variable and in some instances SF estimates came from as few as three tree-trap nights and one location, yet these scarce records are abundant in the context of entomological surveillance for CL, and for most neglected tropical diseases. In that sense, an effort could be made to establish sentinel posts in highly endemic counties, thus rendering feasible a highly standardized estimation of SF abundance across endemic areas, where ideally vector infection is also tracked and this information used for prediction and pro-active vector control [31].
Our data clearly supports that changes in SF abundance and CL cases reported at health facilities in Panamá are associated with ENSO. Interannual variability in CL cases is associated with ENSO, where large epidemics follow the cold ENSO phase, while seasonal patterns are associated with temperature and rainfall variability. CL cases were positively associated with 3-month lagged Lu gomezi and Lu trapidoi abundance estimates from light traps. SF vector abundance, on average, decreased during the hot and cold ENSO phases, when compared with the normal ENSO phase, yet variability in SF was largest during the cold ENSO phase suggesting that SF population outbreaks might play a role in CL epidemics, a subject deserving further research.
|
10.1371/journal.pgen.1003356 | Crif1 Deficiency Reduces Adipose OXPHOS Capacity and Triggers Inflammation and Insulin Resistance in Mice | Impaired mitochondrial oxidative phosphorylation (OXPHOS) has been proposed as an etiological mechanism underlying insulin resistance. However, the initiating organ of OXPHOS dysfunction during the development of systemic insulin resistance has yet to be identified. To determine whether adipose OXPHOS deficiency plays an etiological role in systemic insulin resistance, the metabolic phenotype of mice with OXPHOS–deficient adipose tissue was examined. Crif1 is a protein required for the intramitochondrial production of mtDNA–encoded OXPHOS subunits; therefore, Crif1 haploinsufficient deficiency in mice results in a mild, but specific, failure of OXPHOS capacity in vivo. Although adipose-specific Crif1-haploinsufficient mice showed normal growth and development, they became insulin-resistant. Crif1-silenced adipocytes showed higher expression of chemokines, the expression of which is dependent upon stress kinases and antioxidant. Accordingly, examination of adipose tissue from Crif1-haploinsufficient mice revealed increased secretion of MCP1 and TNFα, as well as marked infiltration by macrophages. These findings indicate that the OXPHOS status of adipose tissue determines its metabolic and inflammatory responses, and may cause systemic inflammation and insulin resistance.
| Type 2 diabetes is one of the most challenging health problems in the 21st century. Although insulin resistance is regarded as a fundamental defect that precedes the development of type 2 diabetes, the nature and cause of insulin resistance remain unknown. Adipose tissue is an important organ that determines whole-body energy metabolism, and its dysfunction is a critical element in the development of systemic insulin resistance. Adipose mitochondrial function is suppressed in the insulin-resistant state, and increased adipose mitochondrial biogenesis is associated with the reversal of insulin resistance by a PPARγ agonist. However, despite these important observations, little is known about how mitochondrial respiratory dysfunction in white adipose tissue (WAT) causes insulin resistance. To determine whether adipose deficiency of mitochondrial respiratory capacity plays an etiological role in systemic insulin resistance, the metabolic phenotype of mice with mitochondrial OXPHOS (oxidative phosphorylation)–deficient adipose tissue was examined. Crif1 is a protein required for the translation of mtDNA–encoded OXPHOS subunits. Interestingly, mice haploinsufficient for Crif1 in adipose tissue showed reduced OXPHOS capacity and developed marked insulin resistance.
| White adipose tissue (WAT) determines whole-body energy metabolism by controlling lipid storage and by releasing adipokines, which may directly or indirectly affect the physiological functions of almost all cell types (for a review, see [1], [2]). These adipocyte functions are perturbed by genetic and environmental factors, which lead to adipocyte dysfunction characterized by hypertrophy, hypoxia and inflammatory process within adipose tissue [3]. Adipocyte dysfunction is further characterized by impaired insulin sensitivity, which is associated with changes in cellular composition or organelle dysfunction, particularly of the endoplasmic reticulum (ER) and mitochondria. An emerging concept to explain insulin resistance in obese individuals is maladaptive responses within the ER, which are prominent in adipose tissue (for a review, see [4]–[6]).
Besides the ER, the mitochondria in white adipocytes are linked with adipocyte differentiation and with the function of mature adipocytes. Recent studies show that drastic increases in mitochondrial biogenesis and reactive oxygen species (ROS) production via the OXPHOS complex play a crucial role in adipocyte differentiation. In addition, the mitochondria in differentiating adipocytes support high energy-consuming lipogenic processes to maintain mature adipocyte function [5], [7]. Therefore, it is suggested that the contribution of adipocyte mitochondria to whole-body energy metabolism or adipocyte plasticity may depend on the mitochondrial OXPHOS capacity of the adipose tissue [6]. Consistent with this, decreased mitochondrial capacity in adipocytes may also alter their insulin sensitivity and/or function due to the high energy requirements of fatty acid storage, adipokine secretion, insulin signaling, and glucose uptake [8], [9].
It is interesting that a marked decrease in the level of transcripts for nuclear-encoded mitochondrial genes in cells derived from the epididymal fat pads of ob/ob mice accompanies the onset of obesity [10]. In db/db and diet-induced obese mice, the expression of OXPHOS genes was markedly reduced compared with that in db/+ mice and control mice fed a standard-fat diet, respectively [11]. In humans, the mtDNA copy number is enriched in adipocytes in adipose tissue, but it decreases slightly with age and increasing BMI, and shows a strong positive correlation with basal and insulin-stimulated lipogenesis in fat cells [12]. More interestingly, suppression of OXPHOS genes is prominent in the visceral adipose tissue of humans with type 2 diabetes independent of obesity [13]. Agonists of peroxisome proliferator-activated receptor-gamma (PPARγ) increase the number of mitochondria and induce mitochondrial remodeling in adipocytes [10], [11], [14], and significantly increase the mitochondrial copy number and expression of factors involved in mitochondrial biogenesis, including PPARγ coactivator-1alpha (PGC1α) and mitochondrial transcription factor A (TFAM), which are required for mitochondrial transcription of OXPHOS genes in humans [15]. These observations in rodent models and human subjects suggest that the OXPHOS capacity of adipose tissue may affect the changes in adipocyte plasticity, which controls insulin sensitivity and may determine the therapeutic responsiveness to antidiabetic agents such as thiazolidinediones and CB1 receptor blockers that affect the mitochondrial content of adipocytes [10], [16].
Here, we demonstrate that primary OXPHOS dysfunction in adipose tissue causes insulin resistance and a diabetic phenotype in mice with a Crif1 loss-of-function mutation. Crif1 is a mitochondrial protein that associates with large mitoribosomal subunits, which are located close to the polypeptide exit tunnel, and the elimination of Crif1 led to both aberrant synthesis and defective insertion of mtDNA-encoded nascent OXPHOS polypeptides into the inner membrane [17]. Targeted elimination of the Crif1 gene resulted in a phenotype characterized by organ-specific failure of OXPHOS function; therefore, we attempted to identify the adipose tissue phenotypes of adipose-specific Crif1-knockout mice using Fabp4-Cre and Adiponectin-Cre mice models. Reduced OXPHOS capacity in the WAT of Crif1-deficient mice triggered spontaneous adipose inflammation, which was characterized by macrophage infiltration and systemic insulin resistance. Therefore, the OXPHOS reserve may be the critical determinant controlling the metabolic and inflammatory responses of adipose tissue, which are closely related to systemic changes in insulin sensitivity.
Crif1 is a mitochondrial protein that specifically interacts with the protein components of the large subunit of the mitochondrial ribosome [17]. It specifically regulates the translation and insertion of the 13 polypeptide subunits that comprise mitochondrial OXPHOS complexes I, III, IV and V. Homozygous Crif1-null mouse embryonic fibroblasts (MEFs) showed a profound failure in translation and expression of these subunits, along with markedly low levels of basal and stimulated (CCCP-treated) mitochondrial oxygen consumption [17]. Disruption of the mouse Crif1 gene consistently resulted in a profound OXPHOS deficiency characterized by the loss of OXPHOS complex subunits and respiratory complexes in vivo.
Crif1 mRNA is ubiquitously expressed, and it is highly expressed in brain, heart, liver kidney and skeletal muscle (Figure S1A). Two types of adipose tissues, brown (BAT) and white (WAT), contained substantial amounts of Crif1 mRNA (Figure S1A). Crif1 mRNA levels were decreased in the WAT, BAT and liver of db/db and ob/ob mice compared to db/+ and ob/+ mice, respectively (Figure S1B). Interestingly, Crif1 mRNA expression in WAT of C57BL/6 mice was downregulated when they were fed a high fat diet (HFD) for 8 weeks (Figure S1C). These findings indicate that Crif1 expression correlates with the nutritional status in adipose tissue.
To identify the roles of Crif1 and mitochondrial OXPHOS in adipose tissue, we tried to induce primary OXPHOS deficiency in adipose tissue in vivo using conditional Crif1 knockout mice. We crossed conditional Crif1 mice (Crif1flox/flox) [18] with mice expressing a Cre recombinase gene under the control of the fatty acid binding protein-4 (Fabp4) promoter (Fabp4-Cre) and the adiponectin promoter (Adipoq-Cre). The resulting pups were born healthy and viable, and showed a normal Mendelian ratio. However, these homozygous Crif1f/f,Fabp4 mice showed delayed weight gain and poor development of adipose tissue (Figure 1A–1C). Unlike the control (Crif1+/+,Fabp4) and Crif1 heterozygous (Crif1f/+Fabp4) mice, Crif1f/f,Fabp4 mice showed uniform lethality within 24 days of birth (median survival = 19.4 days) (Figure 1D).
The perirenal, subcutaneous and epididymal fat pads of Crif1f/f,Fabp4 mice comprised small adipocytes with dystrophic changes (Figure 1E). To verify any mitochondrial abnormalities, the adipose tissues of Crif1f/f,Fabp4 mice were examined by transmission electron microscopy (TEM). The adipocytes of these mice contained mitochondria with ultrastructural abnormalities, such as swollen and distorted cristae, but mitochondrial number was unaffected (Figure 1F and 1G). In heterozygous Crif1f/+,Fabp4 mice, hematoxylin and eosin (H&E) staining of adipose tissue showed no evidence of histological abnormalities compared with the controls (Figure 1E). Consistent with the results of H&E staining, the mitochondria of Crif1f/+,Fabp4 mice showed no morphological or numerical abnormalities of mitochondria in TEM (Figure 1F and 1G). Collectively, this comprehensive analysis of the adipose tissues in Crif1f/f,Fabp4 mice indicated that loss of Crif1 results in a marked failure of WAT and BAT development.
The Fabp4-Cre transgene is expressed and localized within the dorsal root ganglion, centrum of the vertebra and the carpals of the embryo from the mid-gestation stage [19]. Neonatal Crif1f/+,Fabp4 or Crif1f/f,Fabp4 mice did not show developmental abnormalities when compared with control mice. Therefore, embryonic expression of the Fabp4-Cre transgene may not affect the development of Crif1f/+,Fabp4 and Crif1f/f,Fabp4 mice, and may not be a plausible reason for observed lethality at around post-natal Week 3. Mice are normally weaned at post-natal Week 3, at which point the rate of lipogenesis and UCP1 expression in the BAT rises sharply and reaches maximal levels to enhance thermogenesis [20]. The Fabp4-Cre transgene was uniformly detected in BAT from the early post-natal period (Day 7), the Crif1 protein and OXPHOS complex subunits are downregulated in the BAT of 3-week-old mice (Table S1). As shown in Figure 1C, Crif1f/f,Fabp4 mice had less BAT at Day 21, but histological examination of inter-scapular BAT showed normal histological findings (Figure S2A). Crif1f/f,Fabp4 mice had fewer mitochondria than control mice, but these were larger in size and were characteristically disorganized and swollen, suggesting OXPHOS defects (Figure S2B and S2C). Consistent with these findings, Crif1f/f,Fabp4 mice showed a low body temperature under ambient conditions (23°C) and rapidly reached a fatally low rectal temperature within 5 minutes of immersion in cold water (4°C) (Figure S2D). However, although the mass of BAT was reduced, the level of UCP1 expression was not altered in Crif1f/f,Fabp4 mice (data not shown). When Crif1f/f,Fabp4 mice were housed at thermoneutrality (30°C), the median survival rate was increased and mortality was reduced (Figure S2E). This indicates that thermal stress caused by mitochondrial OXPHOS dysfunction in BAT following Crif1 ablation may be a critical factor in the premature death of Crif1f/f,Fabp4 mice.
By contrast, the BAT of Crif1f/+,Fabp4 mice showed normal development and histological and ultrastructural findings (Figure S2A–S2C). Furthermore, the response of Crif1f/+,Fabp4 mice (in terms of core temperature) to a cold environment were identical to those of control mice (Figure S2D). These results showed that Fabp4-Cre driven haploinsufficiency of Crif1 may not affect the physiological function of BAT.
A previous study revealed that Crif1-deficient (−/Δ) MEFs prepared from Crif1−/flox mice showed marked OXPHOS defects due to a profound failure of translation and insertion of the newly-synthesized OXPHOS polypeptides encoded by the mtDNA. Also, Crif1 −/Δ MEFs showed increased anaerobic glycolysis, which eventually led to accelerated cell death [17]. Similar to Crif1 −/Δ MEFs, loss of Crif1 in adipose-derived stem cells (ADSCs) (Crif1M−/−) resulted in marked impairment of differentiation and accelerated cell death, which prevented functional analysis of the mitochondria (data not shown). However, control (Crif1+/+) and Crif1-haploinsufficient ADSCs (Crif1+/−) prepared from Crif1+/+,Fabp4 and Crif1f/+,Fabp4 mice showed identical levels of cell viability and differentiation to those of control cells (Figure 2A and 2B). Interestingly, Crif1+/− ADSCs showed lower expression of OXPHOS subunits (ND1, NDUFA9, UQCRC2 and COX4) and assembled OXPHOS complex I on Western blot and Blue Native PAGE (BN-PAGE) analysis, respectively (Figure 2C and 2D). Crif1+/− ADSCs consumed less oxygen under basal conditions and showed reduced maximal OXPHOS capacity (Figure 2E). Taken together, Crif1 haploinsufficiency in adipocytes resulted in normal differentiation but reduced genetically-determined OXPHOS capacity.
Several experimental criteria have been proposed to test whether a primary in vivo OXPHOS deficiency plays a causal role in insulin resistance [21]. One of these criteria is that perturbations in OXPHOS gene expression and function must be as modest as possible [21]. Thus, we analyzed Crif1 and OXPHOS gene expression to test whether Crif1f/+,Fabp4 mice were suitable for our proposed experiments. Compared with Crif1+/+,Fabp4 mice, Crif1f/+,Fabp4 mice showed about ∼50% of the Crif1 mRNA and protein expression in epididymal WAT (eWAT) (Figure 2F and 2H). Although basal ATP levels in eWAT were not affected by Crif1 haploinsufficiency (Figure 2G), the expression levels of OXPHOS complex I, III and IV subunits were reduced in the epididymal fat pads of Crif1f/+,Fabp4 mice (Figure 2H). BN-PAGE analysis showed that the levels of Complex I and IV and supercomplex in WAT were approximately 20%, 40% and 50% lower, respectively, in Crif1f/+,Fabp4 mice compared to control mice (Figure 2I and 2J). However, normal levels of Crif1 and OXPHOS complexes were expressed in the liver and heart of Crif1f/+,Fabp4 mice (Figure S3A–S3C). In contrast to homozygous Crif1f/f,Fabp4 mice, heterozygous Crif1f/+,Fabp4 mice exhibited normal levels of OXPHOS subunits in BAT and mitochondrial morphology was normal (Figure 2H–2J and Figure S2B).
Food intake and weight gain were comparable in Crif1f/+,Fabp4 and Crif1+/+,Fabp4 mice when fed a normal chow diet (NCD) (Figure S4A and S4B). MR images of control and Crif1f/+,Fabp4 mice fed a NCD or a HFD showed a similar pattern of adipose distribution (Figure S4C). Triglyceride levels in the liver and plasma of Crif1f/+,Fabp4 mice were the same as those in control mice, regardless of whether they were fed a NCD or a HFD. Serum free fatty acid (FFA) levels tended to be higher in Crif1f/+,Fabp4 mice, but were not significantly different from those in control mice (Figure S4D–S4F). Taken together, these results show that Crif1f/+,Fabp4 mice have mildly reduced primary OXPHOS deficiency in adipose tissue but, unlike the lipodystrophic model, they show no defects in adipose tissue development, and no hyperlipidemia or ectopic lipid accumulation.
To identify the relationship between insulin resistance and reduced OXPHOS capacity in adipocytes in vivo, control and Crif1f/+,Fabp4 mice were subjected to glucose tolerance tests after 8 weeks or 14 weeks on a NCD or HFD. Neither control nor Crif1f/+,Fabp4 mice fed a NCD for 8 weeks showed any differences in glucose tolerance following an intraperitoneal injection of glucose (IPGTT, 2 g/kg body weight) (Figure 3A). However, Crif1f/+,Fabp4 mice fed a NCD for 14 weeks developed glucose intolerance (Figure 3B). More impressively, Crif1f/+,Fabp4 mice fed a HFD for 8 weeks showed an earlier onset of glucose intolerance, which was characterized by higher peak glucose levels than those measured in control mice in the intraperitoneal glucose tolerance tests (Figure 3C). Crif1f/+,Fabp4 mice fed a HFD for 14 weeks showed more advanced glucose intolerance, with higher basal (168.8±13.2 mg/dL vs 131.3±8 mg/dL) and peak (516.8±34.8 mg/dL vs 420.4±52.3 mg/dL) plasma glucose levels (Figure 3D). Therefore, regardless of the caloric state, mice with Crif1 haploinsufficiency showed reduced glucose tolerance. Crif1f/+,Fabp4 mice fed a HFD for 14 weeks showed decreased Akt phosphorylation in the liver and muscle and a reduced glucose disposal rate after an intraperitoneal insulin challenge (Figure 3E and 3F). Furthermore, suppression of hepatic glucose production (HGP) by insulin was not different between the two groups, but the glucose infusion rate (GIR) and glucose uptake rate decreased by approximately 18.6% and 14.7%, respectively, during hyperinsulinemic euglycemic clamping after 14 weeks on a HFD (Figure 3G); these data supported the insulin tolerance tests (ITT) results. These findings indicate that Crif1f/+,Fabp4 mice, which have limited OXPHOS capacity in their adipose tissue, may show exacerbated diabetic mechanisms, which are characterized by insulin resistance.
The levels of saturated fatty acids and ceramides in WAT, muscle and liver were not significantly altered in Crif1f/+,Fabp4 mice (Figure S4G and S4H). Thus, abnormal accumulation of ceramides and saturated fatty acids in insulin sensitive tissues does not appear to underlie the insulin resistance of Crif1f/+,Fabp4 mice (Figure S4G and S4H).
To determine the molecular pathways that are dysregulated by mitochondrial OXPHOS dysfunction following Crif1 knockdown by siRNA in adipocytes, we introduced Crif1 siRNA into fully-differentiated 3T3-L1 cells. Crif1 knockdown in differentiated 3T3-L1 cells resulted in decreased expression of the OXPHOS subunits, ND1, NDUFA9, UQCRC2 and ATP5A1, but did not affect the expression of Ppar-gamma, adiponectin, and Cd36 (Figure 4A). A complementary DNA (cDNA) microarray analysis showed prominent increases in the expression levels of inflammatory cytokine and chemokine genes in adipocytes following knockdown of Crif1 (Figure S5). In particular, the chemokines monocyte chemotactic protein 1 (Mcp1/Ccl2), IFN-γ-inducible protein (Ip10/Cxcl10), Regulated upon Activation, Normal T cell Expressed and Secreted (Rantes/Ccl5) and Mig/Cxcl9, which are important for the recruitment of macrophages and T cells to WAT, were elevated in Crif1-silenced 3T3-L1 adipocytes [22]. The elevation of Mcp1, Ip10, and Rantes expression observed in cDNA microarrays was confirmed by real-time PCR experiments with Crif1-silenced 3T3-L1 adipocytes (Figure 4B). In parallel experiments, levels of mitochondrial and cytoplasmic superoxide anions were increased in Crif1-silenced 3T3-L1 adipocytes compared to control cells (Figure 4C). Treatment with the antioxidant N-acetylcysteine (NAC) suppressed the expression of Mcp1, Ip10 and Rantes in Crif1-silenced 3T3-L1 adipocytes (Figure 4D).
Adipose inflammation links adipocyte dysfunction to insulin resistance, which are frequently observed in excessive adiposity (for a review, see). The inflammatory process in adipose tissue is provoked by the activation of stress kinases, e.g., c-Jun N-terminal kinase (JNK), which inhibit insulin signaling and activate transcription factors that mediate the expression of chemokine genes [24], [25]. Intracellular stress signals including mitochondrial ROS, FFA, ceramide and ER stress activates the stress kinases, JNK, p38 MAPK and NF-κB in adipocytes [4], [26], [27]. JNK mediates macrophage activation and expression of proinflammatory cytokines and inhibits insulin receptor substrate 1 (IRS-1)-mediated insulin signaling pathways (for a review, see [28], [29]). To identify the roles of stress kinases in the expression of chemokines in Crif1-silenced 3T3-L1 adipocytes, p38 MAPK and JNK phosphorylation were observed by Western blot analysis. Levels of phosphorylated p38 MAPK and JNK were elevated in Crif1-silenced 3T3-L1 adipocytes compared to control cells; however, this activation was suppressed by NAC treatment (Figure 4E). These results indicate that chemokine dysregulation is associated with increased ROS generation and inappropriate activation of p38 MAPK and JNK. To confirm these results, 3T3-L1 cells were treated with inhibitors of p38 MAPK and JNK (SB203580 and SP60125, respectively). Two inhibitors effectively inhibited the expression of Mcp1 in Crif1 silenced 3T3-L1 cells (Figure 4F). Crif1 deficiency in MEFs results in increased ROS production [17] and induces phosphorylation of p38 (Figure S6A). However, Crif1 -/Δ MEFs did not show increased Mcp1 and Ip10 expression (Figure S6B). Taken together, these results suggest that limited mitochondrial OXPHOS function in fully-differentiated adipocytes triggers the expression of chemokines (Mcp1, Ip10 and Rantes) in a cell-specific manner.
The chemokines (Mcp1, Ip10 and Rantes) upregulated in Crif1 siRNA-treated adipocytes are thought to be critical for attracting macrophages and T lymphocytes into adipose tissue in obese subjects [30]. Therefore, we wondered whether Crif1-silenced 3T3-L1 cells would trigger the migration of macrophages. As shown in Figure 4G, Crif1-silenced 3T3-L1 cells enhanced the migration of RAW 264.7 cells and NAC treatment inhibited the migration of RAW 264.7 cells. Thus, our in vitro studies show that OXPHOS deficiency induced in differentiated cultured 3T3-L1 adipocytes by Crif1 silencing results in the upregulated expression of chemokines, which then recruit or activate macrophages, ROS and stress kinase dependently.
To observe ROS stress associated with abnormal chemokine responses in adipose tissues in vivo, we measured lipid peroxidation (TBAR assays), stress kinase activation and cytokine expression in WAT of Crif1f/+,Fabp4 mice fed a NCD or a HFD for 8 weeks. Consistent with the in vitro studies, lipid peroxidation in WAT and plasma was increased in Crif1f/+,Fabp4 mice fed a HFD for 8 weeks compared to control mice (Figure S7A). Levels of p38 MAPK and JNK phosphorylation were higher in WAT of Crif1f/+,Fabp4 mice fed a HFD for 8 weeks than in control mice (Figure 5A). Furthermore, the expression of Mcp1, Ip10 and Rantes was higher in adipose tissue from Crif1f/+,Fabp4 mice than in control mice (Figure 5B). In addition, the level of secreted MCP1, but not IP10, were higher in the serum of Crif1f/+,Fabp4 mice than in control mice (Figure S7B). The results showing dysregulation of chemokines in the absence of Crif1 suggest that mitochondrial OXPHOS dysfunction may trigger immune cell recruitment in adipose tissue.
To observe inflammation in the adipose tissue of Crif1f/+,Fabp4 mice directly, the eWAT were stained with anti-F4/80, an antibody that detects macrophages. Increased F4/80 reactivity was observed in the eWAT of Crif1f/+,Fabp4 mice fed a NCD for 8 weeks. Aging and a HFD had an even more pronounced effect (Figure 5C). Based on the quantitative real-time PCR results, the relative expression of proinflammatory M1 macrophage markers (Cd11c, Cd11b and Tnfα) increased significantly; however, the relative gene expression of an anti-inflammatory M2 macrophage marker (arginase 1) did not change (Figure 5D). To quantify the number of macrophages in the adipose tissue, multi-parameter flow cytometry was performed with anti-F4/80, anti-CD11c and anti-CD206 antibodies using isolated stromal vascular fractions (SVF). F4/80+/CD11c+/CD206- M1 macrophages were predominant in Crif1f/+,Fabp4 mice compared with control mice. The proportion of F4/80+/CD11c-/CD206+ M2 macrophages tended to be higher in Crif1f/+,Fabp4 mice, but this did not reach statistical significance (Figure 5E). Taken together, the results suggested that the infiltrating macrophages were skewed towards the M1 phenotype in Crif1f/+,Fabp4 mice. Recent studies show that B cell-mediated CD4+ and CD8+ T cell activation is required to induce inflammation and insulin resistance [31], [32]. The present study found no difference between the numbers of CD4+ and CD8+ T cells in Crif1+/+,Fabp4 and Crif1f/+,Fabp4 mice (data not shown).
Adipocytes in adipose tissue secretes adipokines, such as adiponectin, leptin, IL-6 and TNFα, which are involved in the control of whole-body insulin sensitivity. However, proinflammatory TNFα is released by dysfunctional adipocytes and amplifies local immune responses by recruiting macrophages to WAT [33], [34]. Serum levels of TNFα were consistently higher in Crif1f/+,Fabp4 mice fed a HFD than in control mice (Figure 5F). This indicates that TNFα may be a crucial mediator of inflammation in WAT and whole-body insulin resistance of Crif1f/+,Fabp4 mice.
The Fabp4 gene is expressed in macrophages [35], but no Cre expression or activity was detected in macrophages isolated from Crif1f/+,Fabp4 mice. As shown in Figure S8A, the expression levels of Crif1 and macrophage markers (Cd11c, Tnfα, Cd11b, and Arg1) were not reduced in peritoneal macrophages obtained from Crif1f/+,Fabp4 mice (Figure S8A). Homologous recombination using PCR [36] identified Cre recombinase activity in WAT and BAT, but not in peritoneal macrophages in Crif1f/+,Fabp4 mice at 20 weeks-of-age (Figure S8B and S8C).
To verify the adipose inflammation characterized by macrophage infiltration in Crif1-null mice, we generated another adipose tissue-specific Crif1 knockout mouse by crossing floxed Crif1 mice with Adipoq-Cre transgenic mice. Adipoq-Cre transgenic mice expressed Cre recombinase in WAT and BAT, but not in macrophages (including adipose tissue resident macrophages, alveolar macrophages, or thioglycolate-stimulated peritoneal macrophages) [37].
The homozygous Crif1 knockout mice (Crif1f/f,Adipoq) showed about ∼30% of the Crif1 expression observed in the eWAT of controls (Figure S9A). They showed decreased expression of OXPHOS subunits (ND1, NDUFA9, UQCRC2 and COX4) in eWAT and BAT, not in liver and heart (Figure S9B). We compared adipocyte development in the adipocyte-specific Crif1 knockout mouse with that of Adipoq-Cre mice. H&E staining of adipose tissues indicated that the adipocytes of Crif1f/f,Adipoq mice were relatively smaller and irregularly shaped in comparison to those of Crif1+/+,Adipoq mice (data not shown).
Consistent with the Crif1f/+,Fabp4 mouse model, Crif1f/f,Adipoq mice showed higher plasma levels of MCP1 (1.9-fold higher), IP10 (2.5-fold higher) and marked F4/80 immuno-reactivities in eWAT, suggesting inflammation in WAT (Figure 6A and 6B). The nature of the macrophage phenotypes was further identified by flow cytometry using fluorescently-labeled anti-F4/80, anti-CD11c, and anti-CD206 antibodies. In addition, the T cell population was also analyzed using anti-CD3, anti-CD8, and anti-CD4 antibodies. Crif1f/f,Adipoq mice had a higher level of M1 macrophages and a lower level of M2 macrophages in eWAT compared to Crif1+/+,Adipoq mice (Figure 6C). The level of cytotoxic CD8-positive T cells was increased 5.5-fold and the level of CD4-positive helper T cells was decreased 0.5-fold in Crif1f/f,Adipoq in comparison to control mice; however, the levels of these cells in Crif1f/+,Fabp4 mice were not significantly different from control mice (Figure 6D). This reflect differences in the severity of the defect in the mitochondrial OXPHOS complex in Crif1f/f,Adipoq and Crif1f/+,Fabp4mice. Similar to Crif1f/+,Fabp4 mice, Crif1f/f,Adipoq mice developed glucose intolerance even in being fed a NCD, at 8 weeks-of-age (Figure 6E).
Unlike Crif1f/f,Fabp4 mice, Crif1f/f,Adipoq mice were viable. This discrepancy could be due to the inherent differences in the activities of Cre-recombinase driven by Fabp4 and adiponectin promoter (Table S1). Loss of Crif1 was consistently observed in WAT in both mouse lines; however, the degree of Crif1 loss was more severe in Crif1f/f,Fabp4 mice than in Crif1f/f,Adipoq mice. Also, Crif1f/f,Fabp4 mice exhibited a severe loss of BAT and WAT mass, whereas the mass of these tissues was only mildly reduced in Crif1f/f,Adipoq mice. Consistent with these findings, homozygous Crif1f/f,Fabp4 mice rapidly reached a fatal low rectal temperature of 22.6+1.9°C (Figure S2D), whereas homozygous Crif1f/f,Adipoq mice reached a milder rectal temperature of 29.5+0.7°C within 5 minutes of immersion in cold water (4°C) (Figure S9C). These results indicated that thermal stress caused by mitochondrial OXPHOS dysfunction in BAT following Crif1 ablation may be a causative factor of the premature death of adipocyte specific Crif1 knockout mice, and that BAT dysfunction may be partially involved in systemic glucose intolerance.
To determine whether macrophages in the adipose tissue of Crif1f/+,Fabp4 mice play a role in insulin resistance, we depleted macrophages from adipose tissue by intraperitoneal treatment with clodronate liposomes [38]. Clodronate is an apoptosis-inducing drug; therefore, injection of liposome-encapsulated clodronate into the intraperitoneal cavity can deplete phagocytic cells, such as macrophages. Control and Crif1f/+,Fabp4 mice fed a HFD for 8 weeks were intraperitoneally administered two rounds of clodronate liposomes with an interval of 3 days. The accumulation of macrophages positively stained with an anti-F4/80 antibody was decreased in the eWAT following injection of clodronate liposomes (Figure 7A). Furthermore, the level of Cd68 mRNA was significantly lower in the eWAT of mice injected with clodronate than in untreated mice (Figure 7B). Administration of clodronate liposomes dramatically improved the insulin and glucose tolerance of Crif1f/+,Fabp4 mice fed a HFD for 10 weeks (Figure 7C and 7D). These findings indicate that suboptimal reserves of mitochondrial OXPHOS in the adipose tissue of Crif1-deficient mice induce macrophage recruitment, which may trigger systemic insulin resistance (Figure 8).
Mitochondrial dysfunction, characterized by reduced OXPHOS function in liver and skeletal muscle, is thought to be one of the underlying causes of insulin resistance and type 2 diabetes (for a review, see [39], [40]). In addition, reduced hepatic OXPHOS function is closely related to hepatic lipid accumulation and insulin resistance [41]. Collectively, these studies provide evidence of a role for mitochondrial OXPHOS dysfunction in the development of human insulin resistance and type 2 diabetes. However, animal models of OXPHOS dysfunction in skeletal muscle and liver do not exhibit the human insulin resistance and type 2 diabetes phenotype [21], [42], [43]. The absence of insulin resistance in mice with homozygous or heterozygous Crif1 deletion in the liver (Crif1f/f,Alb, Albumin-Cre) or skeletal muscle (Crif1f/+,MLC, MLC-Cre) is in agreement with previous findings that hepatic and skeletal mitochondrial dysfunction does not cause insulin resistance (Figure S10). Therefore, whether or how mitochondrial OXPHOS contributes to the pathogenesis of insulin resistance remains to be resolved. It is reported that adipose OXPHOS capacity is controlled by both genetic and diet-induced obesity [10], [44], [45], which potentially contribute to adipose tissue dysfunction and exacerbation of insulin resistance. However, whether changes in adipose OXPHOS capacity are a cause or a consequence of complications associated with insulin resistance has not been clarified in vivo.
In this study, we have shown an association between limited mitochondrial OXPHOS capacity and adipose tissue inflammation and insulin resistance in a Crif1 haploinsufficiency animal model. Mitochondria play a key role in the differentiation and maturation of adipocytes. It is reported that marked mitochondrial biogenesis is observed during the adipocyte differentiation process in vitro. In fact, the concentration of mitochondrial proteins in differentiated 3T3-L1 adipocytes showed a 20- or 30-fold increase compared with that in pre-adipocytes [14], [46]. Notably, chemical inhibition of respiratory chain function, for example by rotenone treatment, suppresses adipogenesis and induces changes in the expression levels of the key transcription factors, C/EBPα, PPARγ, and SREBP-1c [47]. However, the role played by mitochondria during adipogenesis has mostly been investigated in vitro by inhibiting or knocking down the genes encoding the OXPHOS complex. This study showed that homozygous Crif1 null mice generated by both Fabp4-Cre and Adipoq-Cre recombinase have defects in WAT and BAT development. These observations support that notion that intact OXPHOS function is critical for adipogenesis in vivo. By contrast, our own observations show that heterozygous Crif1 knockout mice do not have defects in adipogenesis and maturation under NCD and HFD conditions. This finding indicates that Crif1 haploinsufficiency and mildly reduced OXPHOS capacity do not cause the apparent failure of adipogenesis in WAT and BAT. Consistently, plasma and liver lipid levels were not increased in heterozygous Crif1 knockout mice, suggesting that the mice do not have the lipodystrophy phenotype. Furthermore, it is reported that insulin resistance in a mouse model of lipodystrophy was not relieved by controlling inflammation [48]. Therefore, insulin resistance in Crif1 haploinsufficient knockout mice may not be related to lipodystrophic changes.
Similar to WAT, the development of BAT was severely perturbed in homozygous Crif1-null mice (Crif1f/f,Fabp4), which may be a critical factor in the early mortality of these mice. By contrast, BAT development was normal in heterozygous Crif1f/+,Fabp4 mice, and the histology and ultrastructure of mitochondria were normal. Furthermore, core temperature responses to a cold environment suggest that Fabp4-Cre-driven haploinsufficiency of Crif1 may not affect the physiological function of BAT. Therefore, decreased BAT function and impaired energy expenditure may not be principal cause of the development of insulin resistance in Crif1-haploinsufficient mice.
The OXPHOS capacity in adipose tissue may be controlled by tissue-specific, genetic and environmental factors. In fact, it is well known that each cell type develops and maintains a specific OXPHOS capacity to satisfy its metabolic and energetic demands. In addition, individual OXPHOS capacity is genetically determined by specific tissues [49]. The cellular and genetic factors that control adipose-specific OXPHOS capacity are not fully understood. Therefore, white adipocyte responses to marginal or limited OXPHOS capacity in vitro and in vivo remain to be elucidated. In the present study, we characterized the enhanced secretory chemokine responses in Crif1-silenced mature adipocytes. Chemokine production in WAT is physiological, but enhanced production is linked to adipose inflammation, which is usually observed in cases of excessive adiposity (for a review, see [50]). Therefore, the earliest events that trigger the process of enhanced chemokine secretion are of great interest. Studies on signals that initiate adipose inflammation are mainly based on the model of ER homeostasis, lipolysis, and fatty acid signals in obese individuals (for a review, see [4]). It is not known how mitochondria or OXPHOS dysfunction modify the chemokine responses in WAT under physiological and pathological conditions. We found that abnormal increases in ROS production and activation of p38 and JNK were associated with increased expression of Mcp1, Ip10 and Rantes. The increase in ROS and the activation of p38 and JNK in adipose tissue are common denominators that respond to cellular stresses. Unexpectedly, haploinsufficient heterozygous Crif1f/+,Fabp4 mice and control mice exhibited similar levels of adipose Akt phosphorylation in response to insulin injection. This indicates that insulin signaling in adipose tissue may not be the principal cause of the systemic glucose intolerance of Crif1f/+,Fabp4 mice. Therefore, the relative importance of these factors (increased ROS, p38 and JNK activation) need to be addressed by suppressing or eliminating these events while studying abnormal chemokine responses and systemic insulin resistance.
We showed that the WAT in Crif1-deficient mice is predominantly infiltrated by macrophages, regardless of excessive adiposity. An increase in the number of adipose tissue macrophages (ATM) is a prominent feature associated with excessive adiposity [51], [52]. Increased expression of chemokines, especially MCP1, is responsible for recruiting macrophages into the WAT [30]. The ATM infiltration of epididymal fat pads in Crif1-deficient mice showed several characteristic features. First, it was present in fat pads with normal adiposity. This finding suggests that macrophage recruitment to adipose tissue caused by impaired OXPHOS capacity may also develop independently of excessive adiposity, but is accentuated in cases of increased adiposity. Mitochondrial OXPHOS dysfunction in the adipose tissue of Crif1f/+,Fabp4 mice fed a NCD for 8 weeks resulted in macrophage recruitment; however, the mice showed normal glucose tolerance. This suggests that a threshold level of macrophage recruitment or activation is required for the development of insulin resistance. Phenotypic analysis of ATM in Crif1-deficient mice demonstrated that the proportion of both M1 and M2 macrophages tended to be increased under NCD and HFD conditions. However, a phenotypic shift toward M1 macrophages was observed in the adipose tissue of Crif1-deficient mice. Thus, these features of macrophage recruitment in WAT were similar to those observed in a mouse model of diet-induced obesity [53].
Our data provide novel insights into the relationship between adipose inflammation and insulin resistance. This study supports the idea that adiposity overwhelms the genetically-determined OXPHOS capacity in adipose tissue, provoking an inflammatory response and insulin resistance. Therefore, it is possible that adipose mitochondrial OXPHOS capacity is an independent factor determining the risk of adipose inflammation and systemic insulin resistance in obese and even in non-obese subjects.
3T3-L1 cells were maintained in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% bovine calf serum (Gibco BRL). Forty-eight hours post-confluence, the cells were differentiated with IBMX (0.5 mM), dexamethasone (1 µM), insulin (10 µg/ml) and 10% fetal bovine serum (Gibco BRL) [54]. Crif1 siRNA (GGA GUG CUC GCU UCC AGG AAC UAU U) was transfected by using Lipofectamine RNAiMAX reagent (Invitrogen) into 3T3-L1 adipocytes on day 4 of differentiation. Migration of Raw264.7 cells was examined in 8.0 µm Transwell filters (Corning Corp). Raw 264.7 cells were maintained on the top well, with the media from 3T3-L1 adipocytes in the bottom well. After twenty-four hours, the Raw 264.7 cells that had not migrated to the filter were removed, and the cells that had migrated through the filter were collected and stained with trypan-blue. ADSCs were cultured as previously described [55]. ADSCs were differentiated into adipocytes using IBMX (0.5 mM), dexamethasone (1 µM), insulin (10 µg/ml) and rosiglitazone (0.5 µM) in M199 medium (Gibco BRL) supplemented with 10% fetal bovine serum (Gibco BRL). After induction of differentiation, lipid accumulation was detected with Oil red O staining. ADSCs were fixed with 10% neutralized formalin, washed with water, and then stained with freshly prepared 0.2% Oil red O solution.
Primary antibodies against OXPHOS complex subunits (NDUFA9, SDHA, UQCRC2, and ATP5A1) were purchased from Invitrogen. Anti-COX4 (#4844) antibody was purchased from Cell Signaling. Anti-ND1 antibody (sc-65237) was purchased from Santa Cruz Biotechnology. Secondary antibodies (goat anti-mouse and goat anti-rabbit) were obtained from Cell Signaling. Anti-p38 antibody, anti-phospho-p38 antibody, anti-JNK-antibody, anti-phospho-JNK antibody, anti-phospho-Akt and total-Akt antibodies were obtained from Cell Signaling and anti-β-actin, α-tubulin antibody was obtained from Sigma-Aldrich. Anti-UCP1 antibody was obtained from Abcam.
Total RNA was isolated using Trizol (Invitrogen). For Northern blot analysis, 10–20 µg of total RNA was loaded onto a 1.5% agarose-formaldehyde gel. A Crif1 probe was constructed using the mouse Crif1 gene digested with KpnI enzyme. The relative intensity of the Crif1/β-actin bands was normalized against that in the brain. Complementary DNA (cDNA) was prepared from total RNA using M-MLV Reverse Transcriptase and oligo-dT primers (Invitrogen). Real-time PCR was performed using cDNA, QuantiTect SYBR Green PCR Master Mix (QIAGEN), and specific primers. The primers used are described in Table S2. Relative expressions were calculated normalized with 18s ribosomal RNA, using Rotor-Gene 6000 real-time rotary analyzer Software (Version 1.7, Corbett Life Science).
Total RNA was prepared from fully-differentiated 3T3-L1 adipocytes transfected with control or Crif1 siRNA. RNA amplification and labeling were performed with the Low RNA Input Linear Amplification kit PLUS (Agilent Technologies). Array hybridization and scanning were performed with a DNA microarray Chip and scanner (Agilent Technologies). Array data was analyzed using the Feature Extraction and GeneSpring Software (Agilent Technologies).
Dihydroethidium (DHE) or MitoSOX were used to detect intracellular superoxide. Fully-differentiated 3T3-L1 cells were incubated with 10 µM DHE or 5 µM MitoSOX at 37°C for 15 min. Fully-differentiated 3T3-L1 cells were washed with Krebs-HEPES buffer (pH 7.4) or HBSS. Images of cells stained with DHE or MitoSOX were obtained by fluorescence microscopy (Olympus, Japan). Cells were trypsinized and analyzed using a FACScan flow cytometer (BD Bioscience) and data analysis was performed using BD FACSDiva software (BD Bioscience).
Before BN-PAGE, mitochondrial isolation was performed as previously described [56] with modifications. Pellets of ADSCs or tissues from mice were resuspended in buffer B (210 mM mannitol, 70 mM sucrose, 1 mM EGTA, and 5 mM HEPES, pH 7.2) and incubated for 5 min at 4°C. After centrifugation at 600× g for 10 min, the supernatant was re-centrifuged at 17,000× g for 10 min. The pellet containing the mitochondrial fraction was used in the Native PAGE Novex Bis-Tris Gel system (Invitrogen) to determine the content of the OXPHOS complex. A total of 20 µg of the mitochondrial fraction in Native PAGE sample buffer supplemented with 0.5% n-dodecyl-β-D-maltoside was loaded onto a Native PAGE Novex 3–12% Bis-Tris gel. The mitochondrial fraction was mixed with Native PAGE sample buffer containing 1% of digitonin to detect the supercomplexes. After electrophoresis, the separated proteins in the gel were transferred to a PVDF membrane, which was then incubated with an anti-OXPHOS antibody mixture (Invitrogen).
OCR was measured using a Seahorse XF-24 analyzer (Seahorse Bioscience). Control Crif1+/+ and Crif1+/− ADSCs were prepared from the eWAT of Crif1+/+,Fabp4 and Crif1f/+,Fabp4 mice. After seeding ADSCs on an XF-24 plate, cells were incubated in differentiation M199 media contained with FBS, IBMX, dexamethasone, insulin and rosiglitazone. After 2 days later, Crif1+/+ and Crif1+/− ADSCs maintained M199 media with insulin for 8 days. The day before OCR measurement, the sensor cartridge was calibrated with calibration buffer (Seahorse Bioscience) at 37°C. Fully-differentiated ADSCs were washed and incubated with M199 (Gibco BRL) without sodium bicarbonate at 37°C in an incubator. Three readings were taken after each addition of mitochondrial inhibitor before injection of the subsequent inhibitors. The mitochondrial inhibitors used were oligomycin (2 µg/ml), carbonyl cyanide m-chloro phenyl hydrazine (CCCP, 10 µM), and rotenone (1 µM). OCR was automatically calculated and recorded by the sensor cartridge and Seahorse XF-24 software. The plates were saved and the protein concentration was calculated to confirm that there were an approximately equal number of cells in each well.
Floxed Crif1 (Crif1flox/flox) mice were generated as previously described [18]. Fabp4-Cre, Albumin-Cre transgenic mice (C57BL/6J) were purchased from the Jackson Laboratory. Adiponectin-Cre transgenic mice were kindly provided by Dr. Evan Rosen. Dr. Steven J Burden provided the MLC-Cre mouse strain. The HFD, which contained 60% fat, was purchased from Research Diets Inc. (D12492). Mice were maintained in a controlled environment (12 h light/12 h dark cycle; humidity 50–60%; ambient temperature 23°C±1°C) and fed ad libitum. For the cold challenge experiments, mice were individually housed in cages pre-chilled to 4°C. Body temperature was monitored using a rectal probe attached to a digital thermometer (TD-300, Shibaura Denshi. Japan) with/without cold stress. For the thermoneutrality experiments, 2-week-old mice were housed with their mothers at a temperature of 30°C±1°C. All mouse experiments were performed in the animal facility according to institutional guidelines, and the experimental protocols were approved by the institutional review board of Korean Research Institute of Biotechnology and Bioscience, and Chungnam National University.
To measure the activity of Cre recombinase, PCR was performed as previously reported [36]. Briefly, after isolation of genomic DNA from WAT, BAT, and thioglycolate-induced peritoneal macrophages, PCR was performed with a combination of three primers: forward primer 1, GGGCTGGTGA AATGTGTTG; reverse primer 2, TCAGCTAGGG TGGGACAGA; and reverse primer 3, TATCAGTCCG AGAAGACCTG. To ensure product specificity from PCR, the extension time was limited to 30 sec.
WAT was fixed in 10% neutralized formalin for 16 h, washed, and then embedded in paraffin. Tissue sections of 5 µm thickness were deparaffinized, rehydrated, and heated in a microwave for 10 min in citrate buffer. The tissue sections were then incubated with primary antibodies (anti-F4/80 (diluted 1∶100; Abcam)) for 16 h at 4°C. Immunohistochemistry was performed using a Polink-1 HRP Rat-NM DAB Detection System (GBI Inc).
WAT and BAT were fixed in 1% glutaraldehyde at 4°C and then washed with 0.1 M cacodylate buffer at 4°C. After washing five times, the tissue was post-fixed with 1% OsO4 in an 0.1 M cacodylate buffer (pH 7.2) containing 0.1% CaCl2 for 1 h at 4°C. Samples were dehydrated by serial ethanol and propylene oxide treatment and embedded in Embed-812 (EMS). The resin was then polymerized at 60°C for 36 h. Tissue was sectioned using an EM UC6 ultramicrotome (LEICA) and stained with 4% uranyl acetate and citrate. Observation was performed using a Tecnai G2 Spirit Twin transmission electron microscope (FEI Company, USA) and a JEM ARM 1300S high-voltage electron microscope (JEOL, Japan).
For IPGTT, mice were fasted for 16 h and then 2 g/kg or 1 g/kg glucose was injected into the intraperitoneal cavity of each mouse. Blood glucose levels were measured at 0, 15, 30, 60, and 90 min using a glucometer (Bayer breeze). ITT was performed by measuring blood glucose after 6 h of fasting followed by intraperitoneal insulin injection (0.75 U/kg; Humalog).
Hyperinsulinemic euglycemic clamping was performed as previously described [57]. Briefly, after an overnight fast, a 2 h hyperinsulinemic euglycemic clamping was performed in Crif1f/+,Fabp4 and control littermates (n = 8). The insulin clamp began with a primed-continuous infusion of insulin (0.3 U/kg bolus followed by 2.5 mU/kg/min). Blood samples (20 µl) were collected at 10 to 20 min intervals for immediate measurement of plasma glucose concentrations, and 20% glucose was infused at variable rates to maintain glucose at basal concentrations (∼120 mg/dL). Basal and insulin-stimulated whole-body glucose uptake was estimated with a continuous infusion of 3H glucose (Perkin Elmer Life and Analytical Sciences) for 2 h before clamping (0.05 µCi/min) and throughout the clamping (0.1 µCi/min), respectively. At 75 min after the start of the clamp, 2-deoxy-d-1-14C glucose (PerkinElmer Life and Analytical Sciences) was injected with a Hamilton syringe to measure insulin-stimulated glucose transport activity and metabolism in skeletal muscle. Blood samples were taken before, during, and at the end of the clamps for measurement of plasma 3H glucose and 2-deoxy-d-1-14C glucose concentrations, and/or insulin concentrations. At the end of the clamps, tissue samples (gastrocnemius, eWAT, and liver) were rapidly taken and stored at −70°C prior to biochemical and molecular analysis.
To quantified M1 macrophages, M2 macrophages, and CD4+ and CD8+ T cell populations, the stromal vascular fractions (SVF) was isolated from mouse eWAT. The SVF was prepared by the lysis of eWAT with type 1 collagenase (Gibco BRL) in collagenase buffer at 37°C in a shaking water bath for 40 min, followed by centrifuging at 2000 rpm for 5 min. The suspended solid matter comprised adipocytes and the cell pellet comprised T cells, B cells and macrophages. The cell pellet was then incubated with RBC lysis buffer and the remaining cells were stained with specific antibodies. Anti-CD3 (BD bioscience), anti-CD4 (BD Bioscience) and anti-CD8 (eBioscience) were used to stain the T cell population [58], and F4/80 (eBioscience), CD206 (eBioscience) and CD11c (eBioscience) were used to stain M1/M2 macrophages. The stained SVF cells were analyzed using a FACScan flow cytometer (BD Bioscience) and data analysis was performed using BD FACSDiva software (BD Bioscience).
The TBAR assay kit (Cayman Chemicals) was used to measure lipid peroxidation in the WAT and plasma of mice. WAT (25 mg) suspended in RIPA buffer was sonicated, centrifuged at 1,600 g for 10 min at 4°C, and the supernatant was collected. The SDS solution was added to the supernatant, which was then mixed with the Color reagent according to the manufacturer's instructions. The sample was boiled for 1 h, centrifuged, and the supernatant was collected. Fluorescence at the excitation wavelength of 530 nm and emission wavelength of 550 nm was measured.
The generation of liposome-encapsulated clodronate was performed as previously described [38]. Cholesterol (10 mg/ml; Sigma-Aldrich) was dissolved in 100% ethanol, and 100 mg/ml phosphatidylcholine in 100% ethanol (Sigma-Aldrich) was made into a phospho-lipid film by drying with a low-vacuum rotary. Clodronate (0.6 M) (Sigma-Aldrich) was dissolved in purified water and incubated with the phospho-lipid film by gentle rotation at room temperature and sonication in a water bath for 3 min at 55 kHz. After removing the non-encapsulated clodronate, liposome-encapsulated clodronate was resuspended in 1X PBS. Two intraperitoneal injections (3 days apart) of clodronate were administered to mice fed a HFD for 8 weeks. IPGTT and ITT were performed 6 days after the first injection.
Measurement of hepatic triglycerides: Liver triglycerides were extracted with chloroform and methanol, dissolved in 1× PBS, and measured in a Hitachi 7150 chemistry analyzer (Hitachi, Japan).
Measurement of ceramides in WAT, liver and muscle: Prior to extraction of total lipids, C17 ceramide was added as an internal standard. Ceramides were measured as previously described [59]. All liquid chromatography-mass spectrometry (LC-MS/MS) experiments were performed using an Agilent 1200 HPLC system (Agilent Technologies, Santa Clara, CA, USA) coupled to a Thermo LTQ linear ion trap mass spectrometer (Thermo Scientific, San Jose, CA) equipped with an electro spray ionization (ESI) source. Briefly, LC separation was achieved using a LunaC18 RP column (150 mm×2 mm I.D., 5 µm 100 Å particles; Phenomenex, Torrance, CA) with gradient elution. Lipid molecules separated by LC were detected by the mass spectrometer in Positive ESI mode using selected reaction monitoring (SRM). The SRM channels were arranged as follows: 538→264 for C16, 552→264 for C17, 566→264 for C18, 594→264 for C20, 648→264 for C24:1, and 650→264 for C24. The peak area was normalized according to the internal standard and tissue weight. All values are presented as relative differences in the ratio of the extracted lipids to the internal standard.
To measure the level of saturated fatty acids, tissues were homogenized in ice-cold methanol, and 1 µg of pentadecanoic acid (C15:0) was added as an internal standard. Samples were incubated at 45°C overnight, then cooled to room temperature. Hexane and 1 mL of H2O were added, samples were vortexed and centrifuged, and fatty acid methyl esters were collected from the upper hexane layer. Samples were analyzed by gas chromatography–mass spectrometry (GC-MS) using an Agilent HP6890 GC interfaced with an HP5973N MSD. A DB-5 column was used. The GC oven temperature was initially 150°C and then increased to 280°C for 52 min. Full MS scans over a m/z range of 60 to 800 were obtained and the peaks of the characteristic ion chromatogram for each fatty acid methyl ester were used for quantification. All samples were normalized against the internal standard.
Whole cardiac blood from the mice was incubated at room temperature for 2 h. The blood was centrifuged at 2,500 rpm for 5 min, and the supernatant was collected. TNFα and IL-4 were measured using a mouse cytokine/chemokine multiplex panel (Millipore). MCP1 and IP10 were measured using an ELISA kit (R&D Systems). Serum triglycerides and FFAs were measured with a Hitachi 7150 chemistry analyzer (Hitachi, Japan).
Data are presented as means ± or + standard deviation (SD). Statistical significance for comparisons was determined using the Student's two-tailed T-test. A p value less than 0.05 was considered statistically significant.
|
10.1371/journal.pcbi.1004632 | Remodeling and Tenacity of Inhibitory Synapses: Relationships with Network Activity and Neighboring Excitatory Synapses | Glutamatergic synapse size remodeling is governed not only by specific activity forms but also by apparently stochastic processes with well-defined statistics. These spontaneous remodeling processes can give rise to skewed and stable synaptic size distributions, underlie scaling of these distributions and drive changes in glutamatergic synapse size “configurations”. Where inhibitory synapses are concerned, however, little is known on spontaneous remodeling dynamics, their statistics, their activity dependence or their long-term consequences. Here we followed individual inhibitory synapses for days, and analyzed their size remodeling dynamics within the statistical framework previously developed for glutamatergic synapses. Similar to glutamatergic synapses, size distributions of inhibitory synapses were skewed and stable; at the same time, however, sizes of individual synapses changed considerably, leading to gradual changes in synaptic size configurations. The suppression of network activity only transiently affected spontaneous remodeling dynamics, did not affect synaptic size configuration change rates and was not followed by the scaling of inhibitory synapse size distributions. Comparisons with glutamatergic synapses within the same dendrites revealed a degree of coupling between nearby inhibitory and excitatory synapse remodeling, but also revealed that inhibitory synapse size configurations changed at considerably slower rates than those of their glutamatergic neighbors. These findings point to quantitative differences in spontaneous remodeling dynamics of inhibitory and excitatory synapses but also reveal deep qualitative similarities in the processes that control their sizes and govern their remodeling dynamics.
| Synaptic plasticity is widely believed to constitute a fundamental mechanism for altering network function. An (implicit) extension of this belief is an assumption that spontaneous changes in synaptic function should not occur to any significant degree. Where excitatory synapses are concerned, recent studies have questioned the validity of this assumption. Where inhibitory synapses are concerned, however, much less is known. Here we followed the spontaneous remodeling dynamics of inhibitory synapses for days, and analyzed these dynamics within a statistical framework previously developed for glutamatergic synapses. Like their excitatory counterparts, sizes of individual synapses fluctuated considerably. Similarly, these spontaneous fluctuations were governed by a well-defined statistical process which assures that synaptic size distributions remain constant. Contrary to the aforementioned assumption, these spontaneous fluctuations drove changes in synaptic size configurations; interestingly, however, change rates were slower for inhibitory synapses. Unlike excitatory synapses, suppressing network activity barely affected inhibitory synapse remodeling dynamics, synaptic configuration change rates or synaptic size distributions. Our findings thus point to quantitative differences in spontaneous remodeling dynamics of inhibitory and excitatory synapses, but also indicate that the processes that control their sizes and govern their remodeling dynamics are fundamentally similar.
| Activity-driven changes in synaptic properties are widely believed to constitute a fundamental mechanism for altering network function. This belief also (implicitly) implies that synapses, when not driven to change their properties by physiologically relevant stimuli, should retain these properties over time. Otherwise, physiologically relevant modifications would be gradually lost, or drowned in a sea of spurious changes.
The capacity of individual synapses to maintain their properties over behaviorally relevant time scales is by no means obvious: Imaging studies, carried over the last decade, have led to the realization that synapses are not structures in a strict sense but are better thought of as complex assemblies of dynamic components (receptors, scaffolding molecules, synaptic vesicles and organelles) which move in, out and between synaptic junctions on time-scales of seconds to hours [1,2]. Conceivably, these dynamics might challenge the capacity of synapses to maintain their individual properties over long time-scales (a capacity we refer to as synaptic tenacity). Indeed, when molecular contents of individual synapses are followed over many hours and days, these exhibit very considerable fluctuations, even in the absence of specific activity patterns, or, for that matter, any activity at all [3–15] (reviewed in [2]). Collectively these studies indicate that synapses exhibit significant spontaneous remodeling in addition to changes directed by specific activity patterns.
So far, synaptic tenacity has been studied mainly in the context of excitatory glutamatergic synapses. Where such synapses are concerned, these studies have pointed to several general principles: First, they suggest that whereas individual synapses exhibit significant spontaneous remodeling over long time scales, distributions of synaptic sizes are skewed and remarkably stable [5,7,14]. Second, it has been shown that spontaneous remodeling dynamics are altered by pharmacological manipulations of network activity, and that such changes can drive the scaling of synaptic size distributions [5,14]. Third, it has been shown that the spontaneous remodeling of individual synapses drives gradual changes in synaptic size configurations, such that over time scales of several days, relations with original synapse sizes are gradually lost [5,14] (a phenomenon we refer to here as a deterioration of synaptic size configurations). Finally, it has been shown that synaptic size dynamics and their effects on synaptic size distributions, scaling and synaptic configurations are described exceptionally well by a statistical process known in probability theory as the Kesten process [14].
At present, it remains unknown if these principles apply to inhibitory synapses as well. Like their excitatory counterparts, inhibitory synapses are dynamic assemblies, in which receptors—both glycine and γ-amino butyric acid (GABA) receptors—continuously diffuse in and out of synaptic sites (e.g. [16–19]; reviewed in [20,21]). These receptors are temporarily retained at synaptic sites through interactions with scaffolding molecules, mainly gephyrin (e.g. [16,22–25]) which, in turn, also exhibit substantial dynamics (e.g. [21,23,26,28–29]; reviewed in [1,20,21]). It thus might be asked: How tenacious are individual inhibitory synapses? How stable are inhibitory synapse size distributions? Are these rightward skewed as well? How are individual and population measures of synaptic tenacity affected by network activity levels? Is the suppression of network activity also associated with the scaling of inhibitory synapse size distributions? Do configurations of inhibitory neurons deteriorate over time? Are inhibitory synapse size dynamics also well-described by a Kesten process? And finally, how do spontaneous changes in inhibitory synapse sizes compare to spontaneous changes in excitatory synapse sizes? Do their size configurations deteriorate at similar rates?
To address these questions we used long-term imaging, multielectrode array recordings and fluorescently tagged gephyrin and PSD-95 to evaluate the spontaneous remodeling dynamics of inhibitory synapses and neighboring glutamatergic synapses, respectively. Our findings are described next.
Gephyrin is a highly conserved, widely expressed protein that is considered to be the core scaffolding protein at postsynaptic densities (PSDs) of inhibitory synapses [30,31]. In neurons, gephyrin plays key roles in the confinement of glycine and GABA-gated chloride channels to the postsynaptic membrane (e.g. 16, 22–24], and its synaptic content is considered to be a reliable indicator of inhibitory synapse size [31]. As gephyrin localizes very specifically to GABAergic (and glycinergic) synapses, fluorescently tagged variants of this molecule have been used in vivo and in vitro to visualize inhibitory synapses in living neurons. This approach has been used for studying synaptic targeting of gephyrin [33,34], inhibitory synapse formation [35] inhibitory synapse turnover following manipulations of network activity and sensory input [36–38], synaptic dynamism [26, 27], and activity-induced inhibitory synapse remodeling [28,29,38] (reviewed in [39]). In some of these studies [36–38], correlated light-electron microscopy of fluorescent gephyrin clusters established a tight correspondence between these fluorescent objects and inhibitory synapses identifiable at the ultrastructural level, validating the use of fluorescently tagged gephyrin as a highly reliable marker of inhibitory synapses.
Given its central role at inhibitory synapses, changes in gephyrin contents are likely to reflect changes in functional properties of the same synapses [21,39]. More conservatively, changes in gephyrin contents are very likely to reflect changes in the sizes of postsynaptic scaffolds at these synapses [32], and thus fluorescently tagged gephyrin can be used as a reporter of inhibitory synapse size. We thus created a fusion protein of a full length variant of gephyrin [23] and the bright cyan fluorescent protein variant mTurquoise2 [40], and used lentiviral vectors to sparsely express this fusion protein (mTurq2:Geph) in large networks of rat cortical neurons in primary culture. To follow these neurons over many hours and days, the neurons were grown on thin-glass multielectrode array (MEA) dishes, that in addition to allowing for long-term, high-resolution fluorescence imaging, allowed us to continuously record network activity in the same networks from the 59 extracellular electrodes embedded in these dishes [5,41].
As shown in Fig 1, mTurq2:Geph assumed a punctate appearance, in which fluorescent puncta were mainly distributed along weakly fluorescent dendritic shafts. As mentioned above, an excellent correspondence between such puncta and inhibitory synapses was previously reported. We confirmed these observations in two manners: We first performed labeling of live neurons using two different antibodies raised against extracellular epitopes of the most common subunits of postsynaptic GABAA receptors, namely γ2 and β2,3 [30] and compared the resulting labeling pattern with those of mTurq2:Geph in the same neurons (see Materials and Methods for further details). Live labeling was performed at 14–18 DIV in sparse networks of cultured hippocampal neurons rather than cortical neurons because the lower cell and synapse density facilitated antibody access and colocalization analysis. As shown in Fig 2A and 2B an excellent degree of colocalization of mTurq2:Geph with clusters of labeled receptors was observed (89% of mTurq2:Geph puncta colocalized with GABAA γ2, 3 experiments, 18 Cells, 1284 puncta; 91% of mTurq2:Geph puncta colocalized with β2,3 receptor clusters, one experiment, 4 Cells, 252 puncta; colocalization with images rotated 180°: 14% and 12%, respectively). On the other hand, when mTurq2:Gephyrin and antibody labeling fluorescence were compared on a synapse to synapse basis, the correlation, although positive, was quite imperfect (r = 0.48, Fig 2D).
We then examined the colocalization of mTurq2:Geph with functional presynaptic boutons of GABAergic neurons by visualizing the uptake of antibodies directed against the lumenal domain of the vesicular GABA transporter (VGAT) as previously described [42]. We found that 71% of mTurq2:Gephyrin clusters colocalized with VGAT puncta (Fig 2C; 4 Experiments, 17 cells, 985 puncta; colocalization with images rotated 180°: 20%). Although this colocalization was less than that observed for GABAA receptors, it is important to keep in mind that presynaptic labeling in this assay depended on the spontaneous activity levels of GABAergic neurons whose boutons impinged on the mTurq2:Geph-expressing neurons, and it is unlikely that these were all equally active. Here too we noted that the correlation between VGAT uptake and mTurq2:Geph fluorescence at single synapses was positive but imperfect (r = 0.34; Fig 2E).
In summary, the good colocalization between mTurq2:Geph, GABAA receptor clusters and VGAT-positive presynaptic boutons further supports prior observations that mTurq2:Geph puncta correspond to bona-fide GABAergic synapses. On the other hand, we did note that the quantitative correlation between the three, when examined on a synapse to synapse basis, was not perfect. This limited correlation might reflect imperfect antibody labeling (due to poor penetration into the synaptic cleft, for example), substantial GABAA receptor composition heterogeneity [43], the heterogeneity of pre/post synaptic “stoichiometry” [12] or differences in spontaneous activity levels. Thus, while some uncertainty remains as to quantitative relationships between mTurq2:Geph fluorescence intensity at a particular synapse and functional measures of the same synapse, mTurq2:Geph can be used to measure the constancy of gephyrin contents at identified inhibitory synapses and by extension, estimate the constancy of their sizes.
We [5,14] and others (e.g. [7]) have previously shown that glutamatergic synaptic size distributions are rightward skewed (heavy tailed), stable over days, and exhibit scaling upon a suppression of spontaneous network activity. Are inhibitory synapse size distributions similar in these regards? To address this question, cortical networks maintained in culture for 18–21 DIV, were mounted on a combined MEA recording/imaging system [5,14,41], and followed for several days, during which neurons expressing mTurq2:Geph were imaged periodically by automated time lapse microscopy. During these experiments the preparations were maintained at optimal conditions provided by covering the MEA dish with a custom built cap equipped with inlet/outlet ports, streaming a sterile mixture of 5% CO2 and 95% air into the dish, perfusing the preparation slowly with fresh feeding medium and heating the MEA dish and oil-immersion objective to 37°C. These conditions were essential for maintaining the long term vitality of these preparations, allowing us to carry out experiments lasting one week or more with no signs of deterioration or cell death (Fig 3). Stacks of images (at 8 focal planes) of neurons expressing mTurq2:Geph were collected automatically from 6–10 fields of view (or sites), with each site representing a portion of the dendritic arbor of a different neuron. Images were collected at one hour intervals for several days concomitantly with recordings of network activity (action potentials, Fig 4A, inset) from the 59 electrodes of the MEA dish. In agreement with many prior studies from multiple groups ([5,41] and references therein), activity in these preparations tended to occur as spontaneous, network-wide bursts (S1 Fig). Although activity levels were generally quite stable, the first 24–36 hours of each experiment were invariably associated with a substantial increase in spontaneous activity levels (reflecting both increasing numbers of active neurons and increased firing rates), as shown for one representative experiment in Fig 4A and for data pooled from 4 experiments in Fig 4B. This initial increase in activity almost certainly reflects the introduction of the slow perfusion [5] and thus reflects an initial transient related to the experimental conditions of these long term experiments.
To quantify the stability of population measures of inhibitory synapses under baseline conditions, fluorescent mTurq2:Geph puncta in the image series were detected programmatically (Fig 3C, bottom panel) counted and their fluorescence was quantified (see Materials and Methods for further details). Such counts and fluorescence levels for one experiment are shown in S2A and S2B Fig (6 neurons, ~260 synapses) and for data pooled from all experiments in Fig 4C and 4D (4 experiments, 27 neurons, ~4,000 synapses). This analysis revealed that puncta counts were generally stable apart from a small initial increase in mTurq2:Geph puncta numbers which became apparent after the first 12 hours (Fig 4C; see also [44–47]). Over these same periods, average fluorescence intensities of mTurq2:Geph puncta were generally stable apart from a slight decrease over time (Fig 4D). The distributions of mTurq2:Geph puncta fluorescence (averaged over three consecutive, 24 hour windows) revealed a majority of small (dim) puncta with a long tail of increasingly brighter (larger) puncta (Fig 4E). These rightward skewed distributions were remarkably stable and showed only a very modest change over time that was consistent with the slight decrease in average mTurq2:Geph puncta fluorescence (Fig 4D). Importantly, this modest decrease in mTurq2:Geph fluorescence was not due to photobleaching, as the same trend was observed in networks that were imaged only once every 24 hours (S2C Fig; 2 experiments, 16 neurons).
We then set out to examine whether blocking all spontaneous activity in these networks would affect population measures of inhibitory synapse size. Will such a manipulation (homeostatically?) downscale inhibitory synapse size distributions, or perhaps drive an upward scaling, consistent with increases in inhibitory synaptic sizes observed under some conditions [28]? (see also [44]) To address this question, we suppressed network activity by adding tetrodotoxin (TTX; 1 μM) to the MEA dish and perfusion media. We then compared recordings carried out for 46–48 hours before TTX application (starting at least 24 hours after mounting the preparations) with those collected for 46–48 additional hours in the presence of TTX.
Pooling data from a total of 4 experiments showed that activity levels during the periods that preceded TTX application were relatively stable, and that, as expected, TTX led to the abrupt transition from very high activity levels to complete and sustained silence (S3A Fig). The cessation of network activity was associated with slight trends, observable over the first 12 hours from TTX treatment, towards increases in mTurq2:Geph puncta numbers (Fig 4F), and average mTurq2:Geph fluorescence values (Fig 4G), but these were not statistically significant. In agreement with these observations, distributions of mTurq2:Geph puncta fluorescence were barely affected (Fig 4H). Of note, practically identical results were obtained in an additional experiment in which Bicuculline (6μM) was added together with TTX (8 neurons, ~210 synapses; S4 Fig).
The lack of changes in gephyrin contents does not necessarily imply that synaptic GABA receptor contents did not change [48]. To test whether activity blockade influences GABAA receptor contents we compared the labeling against GABAA subunit γ2 in live neurons (as described in Fig 2A) either exposed or unexposed to TTX for 24–48 hours. In these experiments (4 experiments, 55 fields of view, >7000 synapses) we observed a trend toward an increase in anti γ2 labeling following a 24 hour exposure period to TTX (~9%, on average), but this increase was not statistically significant (P> 0.19; paired two-tailed t test).
Thus, like their excitatory counterparts, distributions of inhibitory synapses sizes were skewed and very stable over hours and even days. Unlike these, however, abrupt suppressions of network activity were not associated with notable scaling of inhibitory synapse size distributions in one direction or another.
As mentioned in the introduction, where glutamatergic synapses are concerned, the stability of synaptic size population coexists with significant fluctuations in the sizes of individual synapses within the same populations [5,7,14]. To examine if this holds true for inhibitory synapses as well, we tracked specific mTurq2:Geph puncta at one hour intervals and measured how their fluorescence changes over the time course of several days. For this analysis, only puncta present throughout the entire experiment were considered. This process is exemplified for 3 synapses in Figs 3C and 5A. To minimize effects of short term fluctuations and measurement noise, fluorescence intensities measured for each synapse were smoothed using a three-point (3 hour) low-pass-filter. As shown in Fig 5A, some puncta exhibited significant changes in their fluorescence over such periods, whereas the fluorescence of other puncta remained quite stable. Importantly, changes in puncta fluorescence continued to occur even when activity was blocked by TTX.
To quantify the extent of change exhibited by individual mTurq2:Geph puncta and compare these in spontaneously active and silenced networks, we calculated, for the filtered data of each tracked synapse, the normalized range of fluorescence change (“Range over Mean”; [8]) as illustrated in Fig 5B (inset)
RangeMean=100*Fmax−FminF¯
(1)
where Fmax, Fmin and F¯ are the maximal, minimal and average fluorescence intensities measured for a given synapse over a given period, respectively. Average range over mean values measured over 24 hour periods were 24±10% (before TTX) and 24±13% (after TTX). Distributions of range over mean values before and after TTX are shown in Fig 5B (4 Experiments, 27 neurons, 749 tracked synapses). These distributions were rightward skewed, with about 20% of the puncta exhibiting changes of 30% or more over this period, and not different for active and silenced networks (P > 0.97; paired two tailed t-test).
These findings show that in common with what was observed for glutamatergic synapses, the stability of inhibitory synaptic size populations coexists with significant fluctuations in individual synapse sizes within the same populations. Moreover, the finding that the extent of such fluctuations is not different in highly active and completely silent networks indicates that these fluctuations are largely activity independent.
What is the cumulative effect of fluctuations in synaptic sizes? One possibility is that the observed fluctuations have no cumulative effect, merely reflecting, for each synapse, fluctuations around some mean size, such that on average, synaptic sizes remain constant. Put differently, these fluctuations might have no lasting impact on configurations of inhibitory synapse sizes for particular neurons. To examine how such fluctuations might affect configurations of inhibitory synaptic sizes, we followed individual synapses in spontaneously active networks (same neurons as those used for obtaining the data in Fig 4) and plotted the mTurq2:Geph fluorescence of each synapse (749 synapses from 27 neurons in 4 experiments) at increasing times against its fluorescence at an initial time point. We selected this initial time point to be t = 30h to move away from the transient observed at the beginning of such experiments (Fig 4C). As shown in Fig 6A, over the course of 42 hours, the correlation between initial and subsequent mTurq2:Geph fluorescence gradually decreased (manifested as decreases in the coefficient of determination, or R2, serving as a goodness of fit measure). More importantly, however, the slopes of linear regression lines for these plots gradually decreased, whereas their offsets gradually increased (Fig 6B). These systematic and monotonic changes in regression line parameters are not consistent with random fluctuations around fixed points. In contrast, they are very much in line with a statistical framework we recently formulated for glutamatergic synapse remodeling [14] which is based on a statistical process known as the Kesten process [49–51].
The basic premise of the aforementioned framework is that synaptic remodeling dynamics are governed by both additive and multiplicative processes, both of which are inherently stochastic (but can be parametrically dependent on activity) and represent the aggregate of multiple microscopic molecular processes. According to this model, for a given synapse of size x, its new size after some discrete time period (i.e. at t+1) will be
xt+1=εtxt+ηt
(2)
where εt and ηt are not fixed values but random variables drawn independently at each time step from some distribution. In this process, which can give rise to rich and complex dynamics, if εt is, on average less than 1.0 (or more accurately, if 〈ln ε〉 < 0) and ηt is on average positive, size distributions will be both stable and rightward skewed as in Fig 4E and 4H, even though sizes of individual objects fluctuate continuously. Furthermore, it can be shown [14] that when the distributions are stable, the iteration of this process k times will give rise to objects whose size 〈xk〉 on average is such that
〈xk〉=〈ε〉kxt=0+(1−〈ε〉k)
(3)
where 〈ε〉 is the average value of ε for the discrete time steps used, and xt = 0 is the initial size of the object. Note that in this solution, the slope 〈ε〉k decreases as k increases (recall that 〈ε〉 < 1), whereas the offset (1−〈ε〉k) increases. Indeed, this is what we observed for the slopes and offsets of regression lines (Fig 6A and 6B) when the fluorescence values of mTurq2:Geph puncta xk were plotted against their initial fluorescence (xt = 0) at increasingly greater time intervals (k). Moreover, plotting theses slopes as a function of the time interval (k) allowed us to estimate 〈ε〉 (= 0.9942) as illustrated in Fig 6C and then extrapolate the expected slopes for increasingly longer periods (Fig 6D). The estimate of 〈ε〉 and its extrapolation can then be used to estimate the expected rates of synaptic configuration deterioration; In this case, this extrapolation indicated that within a week, the slope would be reduced to 0.38 and within two weeks to 0.14 (not shown). The expected offset at that time would be ~0.9 (eq 3). It should be noted, however, that as mean mTurq2:Geph fluorescence slightly declined over these periods (Figs 4D and S2C), a small component of reduced slope values might be attributable to this slight decline. Correcting for this resulted in extrapolated slopes of 0.59 and 0.35 after one and two weeks, respectively.
Coefficient of determination (R2) values were more difficult to extrapolate as they depend on accurate estimates of the particular distributions of ε and η which are harder to obtain [14] but the empirical data suggested that R2 values decayed quite rapidly (from 0.99 after one hour to 0.69 after 42 hours; Fig 6A), indicating that after 1–2 weeks R2 values would be very small. Taken together with the changes in slopes and offsets described above, the data indicate that within 1–2 weeks, relationships with original synaptic configurations would be largely lost.
To examine if the suppression of network activity affected these decay rates, we performed a similar analysis for the same synapses after exposure to TTX. Specifically, we compared 24 hour time windows before (from -25 to -1 hours) and after (from 22 to 46 hours) adding TTX to the dishes and media. As shown in Fig 6E, the rate at which the slope decayed in these two time periods was practically identical.
Although the findings described so far would seem to indicate that blocking network activity had little effect on the spontaneous remodeling dynamics of inhibitory synapses, we did observe transient effects that mostly disappeared after <24 hours. The first manifestation of this transient was in the evolution of the slopes and offsets in plots such as those shown in Fig 6. Whereas the slopes and offsets before exposure to TTX and 24 hours after application were quite similar, transient increases and decreases in these slopes and offsets, respectively, were observed during the first 12 hours following TTX application (Figs 7A and S5). This transient was also apparent when changes in the mTurq2:Geph fluorescence of individual puncta were plotted as a function of their initial fluorescence. Specifically, changes in the fluorescence intensity of each synapse at the end of 12 hour time windows were calculated by subtracting the fluorescence at the beginning of each time window (F0) from the fluorescence at the end of the time window, and these changes (ΔF) were then plotted as a function of F0 (Fig 7B–7E). As shown here, significant changes in puncta fluorescence were observed for most puncta, regardless of their initial fluorescence. In addition, and in common with previous reports for glutamatergic synapses [5,41] and as expected for changes governed by the Kesten process [14], consistent relationships were observed between the direction of change and initial fluorescence, with the largest synapses showing a tendency to grow smaller, the smallest synapses a tendency to grow larger, and regression lines exhibiting negative slopes and a tendency to cross the abscissa near the mean fluorescence value. Interestingly, however, in first 12 hour window following TTX application, these tendencies were transiently reversed (Fig 7C), only to recover during the next time windows (Figs 7D, 7E and S5).
Collectively, these data suggest that spontaneous inhibitory synapse remodeling leads to a gradual deterioration of inhibitory synapse size configurations, to the point that in these preparations, these are lost, for the most part, after 1–2 weeks; Moreover, apart from transient effects, the data indicate that the deterioration of inhibitory synapse size configurations occurs in highly active and silent networks to a similar extent.
The findings described above indicate that in common with what was observed for glutamatergic synapses [5,14] configurations of inhibitory synapses deteriorate over time. Are configuration deterioration rates identical for these two major types of synapses? A comparison of estimates of 〈ε〉 (for one-hour time intervals) for glutamatergic synapses (0.9848; [14]) and inhibitory synapses (0.9942; Fig 7C) indicates that configurations of inhibitory synapses deteriorate at a slower pace than configurations of glutamatergic synapses. This comparison is based, however, on data collected in different sets of neurons at different times, questioning the validity of this comparison.
To directly compare the spontaneous remodeling of GABAergic and glutamatergic synapses, we used lentiviral vectors to express, in the same neurons, two spectrally separable fusion proteins of postsynaptic scaffolding molecules, namely mTurq2:Geph for inhibitory synapses (as described so far) and PSD-95 fused to EGFP (PSD-95:EGFP) for glutamatergic synapses. PSD-95 is a core scaffold protein of glutamatergic synapses which has been used extensively to study the formation and remodeling of such synapses (e.g. [5,12,15,27,52–54]) and was shown to provide an excellent measure of excitatory postsynaptic density size [15,54]. One neuron in which both fusion proteins were expressed is shown in Fig 8A. As expected, the majority of PSD-95:EGFP puncta were located on dendritic spines, whereas inhibitory synapses were mainly located along dendritic shafts. The experiments were performed as described above and here too, TTX was introduced after 3–4 days of baseline recordings to examine the dependence on network activity. At the end of the experiments, nearby mTurq2:Geph and PSD-95:EGFP puncta were tracked separately and their fluorescence intensities obtained. To further minimize potential confounds that might arise from different locations along dendrites, we focused our attention on small groups of neighboring GABAergic and glutamatergic synapses as shown in Fig 8B for one mTurq2:Geph and three PSD-95:EGFP puncta.
We first examined whether the dual expression of both fluorescent reporters affected the gross dynamics observed for each of these when expressed alone. During 24 hour baseline recording periods, average fluorescence levels of mTurq2:Geph exhibited a modest decline whereas those of PSD-95:EGFP remained relatively stable. After TTX application, average mTurq2:Geph fluorescence intensities remained relatively stable, whereas those of PSD-95:EGFP gradually increased (Fig 8D). These observations are in good agreement with the data presented in Fig 4; (mTurq2:Geph) and prior studies ([5,41]; PSD-95:EGFP), indicating that the dual expression did not dramatically alter the gross dynamics of these fusion proteins.
We then compared the deterioration rates of synaptic configurations for GABAergic and glutamatergic synapses positioned in the same neurons, at the same locations and thus, presumably, experiencing very similar histories. For each neuron expressing both reporters, 49 groups of inhibitory synapses and their neighboring excitatory synapses (in 6 neurons from 4 separate experiments) were selected and followed for 24 hours before and after TTX application as illustrated in Fig 8A–8C. We then plotted the fluorescence values of mTurq2:Geph and PSD-95:EGFP puncta, at increasing time intervals, as a function of their fluorescence at the beginning of each 24 hour window, in a manner similar to that shown in Fig 6A. The slopes in these plots decayed at faster rates for PSD-95:EGFP than they did for mTurq2:Geph, resulting in estimates of 〈ε〉 of 0.9916 and 0.9947 for PSD-95:EGFP and mTurq2:Geph, respectively (note that the latter is practically identical to the estimate of Fig 6). When this analysis was carried out on a cell to cell basis, however, differences between slope decay rates for PSD-95:EGFP and mTurq2:Geph did not reach statistical difference, possibly due to the difficulty of obtaining accurate estimates of 〈ε〉 from very small numbers of synapses. Importantly, however, goodness of fits (R2) values in these plots decayed much faster for PSD-95:EGFP in comparison to mTurq2:Geph, (Fig 9A and 9B) and this difference was statistically significant when carried out on a cell to cell basis (P = 0.045, paired t-test, n = 6 cells). As R2 is equal to the square of Pearson’s correlation coefficient, this difference suggests that inhibitory synaptic size configurations change at slower rates as compared to those of their excitatory counterparts.
Several recent reports suggest that spontaneous remodeling of nearby inhibitory and excitatory synapses can be coupled (e.g. [36,27,55–57]). We wondered if this coupling, mainly observed at the level of synapse turnover, extends to spontaneous changes in synaptic sizes. We first examined whether a relationship could be found between PSD-95:EGFP fluorescence and the distance from the nearest inhibitory synapse. To that end, the fluorescence intensities of PSD-95:EGFP puncta were averaged over 24 hour windows, before and after TTX application. When these normalized intensities were plotted as a function of distance to the nearest mTurq2:Geph puncta (measured at the beginning of each time window), no discernible correlation was observed, neither before nor after TTX application (Fig 10A; 49 mTurq2:Geph puncta; 163 PSD95:EGFP puncta from 6 neurons in 4 experiments). We then examined to what degree changes in PSD-95:EGFP fluorescence co-varied with changes in mTurq2:Geph fluorescence in nearby inhibitory synapses and how this co-variance depended on the distance between them. To that end, we calculated the linear correlation (Pearsons’ correlation) between the fluorescence traces of nearby mTurq2:Geph and PSD-95:EGFP puncta over 24 hour windows and plotted these as a function of distance between the puncta (Fig 10B). Here too, no statistically significant correlation was observed, neither before not after TTX addition. However, when comparing the average co-variance within groups (regardless of distance) before and after TTX application, we did observe that the average correlation before TTX application was positive, and that it was significantly reduced following exposure to TTX (Fig 10C, Matched). As a control, we calculated the correlations of fluorescence traces between all mTurq2:Geph and PSD-95:EGFP puncta in the entire data set. As might be expected, no significant correlations were observed (Fig 10C, Shuffled). Thus, while some activity-dependent co-variance was observed for the remodeling of nearby excitatory and inhibitory synapse remodeling, no strong relationships with distance were observed, at least on the length scales examined here.
In summary, these data indicate that even though some weak, activity dependent coupling might exist between the spontaneous remodeling of nearby excitatory and inhibitory synapses, configurations of inhibitory synapses and excitatory synapses seem to deteriorate at different rates, with inhibitory synapses appearing to be less labile in this respect.
Here we set out to study the tenacity and spontaneous remodeling of inhibitory synapses and examine if the same principles previously shown to govern glutamatergic synapse remodeling dynamics also apply to inhibitory synapses. By following mTurq2:Geph expressed in cortical neurons for several days we found that the numbers and sizes of inhibitory synapses were generally stable as were the rightward skewed distributions of GABAergic synapse sizes. Blocking spontaneous network activity barely affected synapse numbers and sizes, had no lasting effect on either of these measures, and did not lead to the scaling of inhibitory synapse size distributions. At the same time, the sizes of inhibitory synapses within the same population changed significantly, with the extent of these remodeling dynamics unaffected by the complete suppression of spontaneous network activity. These spontaneous changes in synaptic sizes led to the gradual deterioration of inhibitory synapse configurations over time scales of several days in a manner that was consistent with a statistical framework previously developed to describe remodeling dynamics of glutamatergic synapses [14]. The rates at which inhibitory synapse configurations deteriorated did not seem to be affected by network activity as configurations deteriorated at similar rates even when network activity was completely suppressed. Finally, comparisons of remodeling dynamics for GABAergic and glutamatergic synapses in the same neurons and dendritic locations revealed that configurations of inhibitory synapses seem to be less labile than those of glutamatergic synapses, even though some degree of activity-dependent coupling was observed for the remodeling of nearby inhibitory and excitatory synapses. These findings thus point to certain differences in the spontaneous remodeling dynamics of GABAergic and glutamatergic synapses. Importantly, however, they also point to deep similarities in the processes that control synapse sizes and govern their spontaneous remodeling: The ongoing, apparently stochastic changes in individual synapse size, the resulting skewed shape and stability of synaptic size distributions, the gradual deterioration of synaptic configurations and the good fit to the same statistical process. The implications of these findings are described below.
The assessment of inhibitory synapse size in this study was almost entirely based on quantifying the fluorescence of a fusion protein of gephyrin and mTurquoise2, with the assumption in mind that changes in mTurq2:Geph fluorescence reflect changes in GABAergic PSD size and, by extension, inhibitory synapse strength. This assumption has several caveats: First, unlike the excellent quantitative relationships established for glutamatergic synapses between the fluorescence of PSD-95 fusion proteins and electron microscopy-based measurements of PSD area (e.g. [15,54]), so far, to the best of our knowledge, similar relationships have not been established at this quantitative level of detail for gephyrin (but see [32]). Second, the very modest changes in synaptic gephyrin contents induced by prolonged activity suppression (in agreement with [58]) would seem to be at odds with other studies suggesting that activity-deprivation results in decreases in mIPSC amplitude and reductions in GABAAR labeling (e.g. [58–61]). We note, however, that in our hands no reductions in synaptic contents of the GABA receptor subunit γ2 were observed (see also [58,62]) and that others have reported opposite effects [44]. Finally, comparisons between mTurq2:Geph, GABAA receptor subunits and presynaptic vesicular markers at individual synapses revealed that the correlations between these measures, while positive, were quite imperfect (Fig 2D and 2E). While good explanations for this imperfect correlation can be provided (see Results) it has to be acknowledged that relationships between gephyrin contents and inhibitory synapse strength are multilayered and not entirely obvious (see ref. [31] for example). Yet, given the critical role played by gephyrin in regulating the numbers of resident GABAA receptors at GABAergic synapses, it is highly unlikely that synaptic gephyrin contents and synaptic strengths are unrelated ([39]; see also [25,34]). Indeed, a recent study showed that gephyrin recruitment is required for recruitment of GABAA receptors during a form of inhibitory synapse long term potentiation [29] (see also [38]).
A more general concern relates to the experimental system used here, that is, networks of dissociated rat cortical neurons in primary culture. This system is advantageous in that it allows for excellent long-term optical and electrophysiological access, with the latter being extremely important to measure baseline network activity levels, verify the effects of pharmacological manipulations, and measure the “contrast” of activity features before and after such manipulations. Yet the question remains whether the limited tenacity of inhibitory synapses reported here is somehow related to our experimental system. It is worth noting, however, that large changes in the fluorescence of tagged gephyrin at inhibitory synapses have been recently reported in vivo as well [32]. Moreover, where excitatory synapses are concerned, spontaneous changes in synaptic contents of PSD-95:EGFP in-vivo ([15]; see also [63]) are at least as large as those observed in our ex-vivo networks ([5,14]; this study) indicating that our observations are not limited to cell culture settings, although particular rates measured in these two settings might differ.
The experiments described above suggest that sizes of inhibitory synapses change spontaneously over time scales of hours and days. These observations are in excellent agreement with a recent study that examined spontaneous changes in presynaptic boutons of GABAergic neurons in organotypic hippocampal cultures [11]. In this study, it was shown that over time scales of hours, the volumes of individual presynaptic boutons exhibit very substantial fluctuations, comparable to those reported here for postsynaptic gephyrin. Similarly, despite these fluctuations, average bouton volumes remained generally stable over several hours. Furthermore, and in agreement with our own observations, manipulations of spontaneous activity levels had very modest effects on bouton volumes. Finally, and perhaps most striking are the observations on the transient effects of network activity suppression on bouton dynamics: whereas a 4 hour exposure to TTX suppressed bouton volume fluctuations, these returned to control levels after longer exposures, in excellent agreement with the transient effects we observed for postsynaptic gephyrin dynamics (Fig 7). These studies thus complement each other and indicate that the limited tenacity of GABAergic synapses is a synapse-wide phenomenon, in agreement with a prior report for excitatory synapses [12].
It might be asked how average synaptic sizes and synaptic size distributions remain constant when sizes of individual synapses change in an apparently stochastic fashion. As mentioned above, for populations of objects whose sizes fluctuate according to a Kesten process in the appropriate regime, size distributions are guaranteed to remain stable (see [14] for a more detailed explanation). Some intuition can be gained by examining Figs 6A, 7B, 7C, 7E and S5. As these figures show, even though synaptic sizes fluctuate in an apparently stochastic fashion, the combination of a negative (on average) scaling/multiplicative factor and a positive (on average) additive component leads to predictable tendencies: Larger than average synapses tend to grow smaller, whereas smaller than average synapses tend to grow larger. This powerful rule thus acts to maintain synaptic sizes within particular ranges even when synaptic sizes change in an apparently stochastic fashion [5,41]. It is also worth noting that the Kesten process also guarantees that such steady state size distributions will be rightward skewed, in concordance with the skewed size distributions observed here for inhibitory synapses (Fig 4) and those previously observed for glutamatergic synapses [5,7]. Interestingly, it has been suggested that the widespread occurrence of skewed distributions in brain organizational features has important implications for network function (reviewed in [64]); our finding concerning inhibitory size distributions would thus seem to further support this notion.
It is widely believed that the types, patterns, and strengths of synaptic connections formed among neurons determine the functional properties of neuronal networks or the information they contain. Moreover, changes in their properties or in their information content are generally believed to be realized by directed changes in patterns and strengths of synaptic connections. Increasing evidence, however, obtained both in-vitro and in vivo is leading to a gradual realization that excitatory synapses are not “static devices” that change only when instructed to do so, but also change spontaneously—even in the absence of particular activity patterns or any activity at all (reviewed in [2]). How such observations can be reconciled with prevailing notions on neuronal network function is not clear. An interesting result in this regard might be found in a recent study in which in-vivo measurements of spontaneous synaptic “volatility” were “plugged in” to simulations of large neuronal networks (Gianluigi Mongillo, Simon Rumpel, Yonatan Loewenstein, COSYNE 2015, abstract I-88). It was shown that key functional properties of such networks are much more sensitive to spontaneous changes in inhibitory connections as compared to spontaneous remodeling of excitatory ones, and it was concluded that inhibitory plasticity has a far larger potential for changing network functionality than the extensively-studied excitatory plasticity. Our findings that inhibitory synapse configurations are less labile than those of excitatory ones would seem to be in line with this idea, as they would imply that spontaneous changes in network function might be slower that what might be expected from observations of glutamatergic synapse “volatility”. Conversely, this finding might be interpreted to suggest that inhibitory synapses represent an effectively stable foundation that minimizes the impact of excitatory synapse size fluctuations, an interpretation that would be consistent with the relative insensitivity of inhibitory synapse remodeling to dramatic changes in network activity levels (see also [27]).
Although configurations of inhibitory synapses were less labile that those of their glutamatergic counterparts, such configurations changed nonetheless. Thus, quantitative differences notwithstanding, the emerging realization that excitatory synapse properties can change spontaneously seems to apply to inhibitory synapses as well. It is of particular note in this respect that major features of excitatory and inhibitory remodeling dynamics—the volatility of individual synapse sizes, the stability of their size distributions, the skewed shape of these distributions and the evolution of synaptic configurations—are all captured remarkably well by a statistical process that is essentially stochastic [14]. It thus remains to be understood how invariant function can be retained in networks composed of such highly unreliable components, a question that is not unrelated to one of the most fundamental issues in the life-sciences, that is the emergence of macroscopic order from microscopic disorder.
Primary cultures of rat cortical neurons were prepared as described previously [5] using a protocol approved by the Technion committee for the supervision of animal experiments. Briefly, cortices of 1–2 days-old Wistar rats of either sex were dissected, dissociated by trypsin treatment followed by trituration using a siliconized Pasteur pipette. A total of 1–1.5x106 cells were then plated on thin-glass multielectrode array (MEA) dishes (MultiChannelSystems—MCS, Germany) whose surface had been pre-treated with polyethylenimine (Sigma) to facilitate cell adherence. The preparations were then transferred to a humidified tissue culture incubator and maintained at 37°C in a gas mixture of 5% CO2, 95% air, and grown in medium containing minimal essential medium (MEM, Sigma), 25 mg/l insulin (Sigma), 20 mM glucose (Sigma), 2 mM L-glutamine (Sigma), 5 mg/ml gentamycin sulfate (Sigma) and 10% NuSerum (Becton Dickinson Labware). Half the volume of the culture medium was replaced three times a week with feeding medium similar to the medium described above but devoid of NuSerum, containing a lower L-glutamine concentration (0.5 mM) and 2% B-27 supplement (Invitrogen).
Primary cultures of rat hippocampal neurons for antibody labeling experiments were prepared as described previously [65]. Briefly, hippocampal CA1–CA3 regions of 1–2 days-old Wistar rats of either sex were dissected, dissociated as described above and plated onto 22×22 mm coverslips coated with poly-D-lysine (Sigma) inside 8-mm-diameter glass cylinder microwells (Bellco Glass). Culture medium consisted of MEM, 20 mM glucose, 0.1 g/l bovine transferrin (Calbiochem), 25 mg/l insulin (Sigma), 2 mM L-glutamine (Sigma), 10% NuSerum (Becton Dickinson Labware), 0.5% fetal bovine serum (HyClone), 2% B-27 supplement (Gibco), and 8 μM cytosine β-D-arabinofuranoside (Sigma) which was added to the culture medium after 3 days. Culture medium was replaced once a week with feeding medium.
Exogenous DNA was introduced into neurons by means of a 3rd generation lentiviral expression system. These vectors code for the fluorescently tagged postsynaptic protein PSD-95:EGFP (described in [5]) and mTurq2:Gephyrin, constructed as described next.
Venus:Gephyrin in pEGFP-N1 [23] was provided as a generous gift by Antoine Triller (Ecole Normale Supérieure, Paris). The insert was sequenced and apart from one silent point mutation (Gly->Gly) and the absence of the N-terminal methionine was found to be identical to full length rat Gephyrin (NCBI Reference Sequence: NM_022865.3). A silent point mutation (His->His) was then made to eliminate an AgeI site, an XhoI site was added to 3' terminus of Gephyrin by PCR and the resulting construct was cut out and inserted between the BsrGI and XhoI sites of the FUGW lentiviral backbone [66] which we previously modified by moving the XhoI site from the 3’ to the 5’ side of the woodchuck hepatitis post-transcriptional regulatory element (WPRE). A construct coding for mTurquoise2 [40] flanked by AgeI and BsrGI was synthesized by large scale DNA synthesis. mTurquoise2 then was inserted into the modified FUGW backbone instead of EGFP between the AgeI and BsrGI sites. All cloning and gene synthesis was done by Genscript (Piscataway NJ, USA).
Lentiviral particles were produced by transfecting HEK293T cells with a mixture of three packaging plasmids: pLP1, pLP2, and pLP\VSVG (packaging vector mix of the ViraPower four-plasmid lentiviral expression system, Invitrogen) and the expression vector. HEK cell transfection was performed using Lipofectamine 2000 (Invitrogen) in 10cm plates when the cells had reached 80% confluence. Supernatant was collected after 48 hours, filtered through 0.45μm filters, aliquoted, and stored at -80°C.
Transduction of cortical cultures was performed at 5 days in-vitro (DIV) by adding predetermined amounts of the filtered supernatant to each MEA dish.
A MEA system was used to continuously monitor the electrical activity of the network through 59, 30 μm diameter recording electrodes, arranged in an 8x8 array, spaced 200 μm apart. The dishes contain 59 rather than 64 electrodes because the corner electrodes are missing, and one of the remaining leads is connected to a large substrate embedded electrode designed to be used as a reference (ground) electrode. The flat, round electrodes are made of titanium nitride, whereas the tracks and contact pads are made of transparent Indium Tin Oxide.
Network activity was recorded through a commercial 60-channel headstage/amplifier (Inverted MEA1060, MCS) with a gain of 1024x. The amplified signal was multiplexed into 16 channels, amplified by a factor of 10 by a 16 channel amplifier (Alligator technologies, USA) and then digitized by an A/D board (Microstar Laboratories, USA) at 12 KSamples/sec per channel. Data acquisition was performed using AlphaMap (Alpha-Omega, Israel). All data was stored as threshold crossing events with the threshold set to -40μV. Electrophysiological data were imported to Matlab (MathWorks, USA) and analyzed using custom written scripts.
All imaging was performed on neurons grown on thin glass MEA dishes as described above. These particular MEA dishes are fabricated of very thin glass (180 μm), which allows for the use of high numerical aperture, oil immersion objectives and are thus ideally suited for high-resolution imaging [5]. Scanning fluorescence and brightfield images were acquired using a custom designed confocal laser scanning microscope based on a Zeiss Axio Observer Z1 using a 40×, 1.3 N.A. Fluar objective. The system was controlled by custom written software and includes provisions for automated, multisite time-lapse microscopy. MEA dishes containing networks of cortical neurons were mounted on the aforementioned headstage/amplifier which was attached to the microscope’s motorized stage.
mTurq2:Gephryin and PSD-95:EGFP were excited using 457 nm (Cobolt) and 488 nm (Coherent) solid state lasers, respectively. Fluorescence emissions were read through 467–493 nm and 500–550 nm bandpass filter (Semrock, USA and Chroma Technology, USA). Time-lapse recordings were usually performed by averaging six frames collected at each of 8 focal planes spaced 0.8 μm apart. All data were collected at a resolution of 640 x 480 pixels, at 12 bits/pixel, with the confocal aperture fully open. Data was collected sequentially from up to 12 predefined sites, using the confocal microscope robotic XYZ stage to cycle automatically through these sites at 1 hour (or 24 hour) time intervals. Focal drift during the experiment was corrected automatically by using the microscopes' "autofocus" feature.
MEA dishes were covered with a custom designed cap containing inlet and outlet ports for perfusion media and air mixtures, a reference ground electrode and a removable transparent glass window. The MEA dish was continuously perfused with feeding media (described above) at a rate of 2.5 ml/day by means of a custom built perfusion system based on an ultra-slow flow peristaltic pump (Instech Laboratories Inc., USA) and silicone tubing. The tubes were connected to the dish through the appropriate ports in the custom designed cap. A 95% air / 5% CO2 sterile mixture was continuously streamed into the dish at very low rates through a third port with flow rates regulated by a high precision flow meter (Gilmont Instruments, USA). The base of the headstage/amplifier and the objective were heated to 37°C and 36°C respectively using resistive elements, separate temperature sensors and controllers, resulting in temperatures of 36–37°C in the culture media.
All imaging data analysis was performed using custom written software (“OpenView”). Special features of this software allow for automated / manual tracking of individual synaptic puncta and measurements of fluorescent intensities of these over time (described in detail in [41]). 9 × 9 pixel (~1.3 x 1.3μm) areas were placed on the centers of fluorescent puncta and mean pixel intensities within these areas were obtained from maximal intensity projections of Z section stacks.
For measuring distributions of puncta intensities, areas were placed programmatically on fluorescent puncta at each time step using identical parameters but no tracking of individual puncta was performed (see Fig 3C, bottom panel). For tracking identified puncta, areas were placed initially over all puncta and then a smaller subset (typically 30–50 per site) was thereafter tracked. As the reliability of automatic tracking was not absolutely perfect, all tracking was verified and, whenever necessary, corrected manually. Puncta for which tracking was ambiguous were excluded.
Because mTurq2:Geph (and PSD-95:EGFP) expression levels varied slightly from one neuron to another, mTurq2:Geph (and PSD-95:EGFP) puncta fluorescence data from each neuron were first normalized to mean mTurq2:Geph (or PSD-95:EGFP) puncta fluorescence of that neuron, allowing us to pool data from different neurons and experiments, correcting for differences in expression levels and potential variations in optical parameters between experiments. Consequently, all fluorescence values throughout the study are expressed as fractions of mean neuronal mTurq2:Geph (or PSD-95:EGFP) fluorescence at initial time points.
Tetrodotoxin (TTX; Alomone Labs) was diluted in 100 μl of medium drawn from the culture dish while on the microscope. The mixture was subsequently returned to the dish and mixed gently. Applications to the dish were complemented by simultaneous addition to the perfusion media. Final concentrations in the dish and perfusion media were 1 μM. In one experiment, bicuculline (Sigma) was added as well.
Primary antibodies against extracellular epitopes of the GABAAR γ2 or β2,3 subunits (rabbit: 1:500; Synaptic Systems, Gottingen, Germany; mouse: 1:100; Millipore) were prepared in phosphate buffered saline (PBS) at the predetermined concentration. The primary antibodies were labeled with fluorescent Fab’ fragments (Zenon labeling kit; Invitrogen) according to the manufactures instructions (Zenon 647nm—rabbit and Zenon 488—mouse, respectively). At 14–18 DIV, primary hippocampal cultures were incubated for 1 hour at 37°C with the primary/Fab’ mixtures (antibodies against GABAAR γ2 or β2,3 subunits) or a primary antibody against Vesicular GABA transporter (VGAT) lumenal domain fluorescence-labeled with Oyster 488 (rabbit: 1:200; Synaptic Systems [42]). After the incubation period, the cells were gently washed three times in physiological solution ("Tyrode's", 119mM NaCl, 2.5mM KCl, 2mM CaCl2, 25mM HEPES, 30mM glucose, buffered to pH 7.4) and imaged immediately.
|
10.1371/journal.pntd.0005969 | Preclinical antivenom-efficacy testing reveals potentially disturbing deficiencies of snakebite treatment capability in East Africa | Antivenom is the treatment of choice for snakebite, which annually kills an estimated 32,000 people in sub-Saharan Africa and leaves approximately 100,000 survivors with permanent physical disabilities that exert a considerable socioeconomic burden. Over the past two decades, the high costs of the most polyspecifically-effective antivenoms have sequentially reduced demand, commercial manufacturing incentives and production volumes that have combined to create a continent-wide vacuum of effective snakebite therapy. This was quickly filled with new, less expensive antivenoms, many of which are of untested efficacy. Some of these successfully marketed antivenoms for Africa are inappropriately manufactured with venoms from non-African snakes and are dangerously ineffective. The uncertain efficacy of available antivenoms exacerbates the complexity of designing intervention measures to reduce the burden of snakebite in sub-Saharan Africa. The objective of this study was to preclinically determine the ability of antivenoms available in Kenya to neutralise the lethal effects of venoms from the most medically important snakes in East Africa.
We collected venom samples from the most medically important snakes in East Africa and determined their toxicity in a mouse model. Using a ‘gold standard’ comparison protocol, we preclinically tested the comparative venom-neutralising efficacy of four antivenoms available in Kenya with two antivenoms of clinically-proven efficacy. To explain the variant efficacies of these antivenoms we tested the IgG-venom binding characteristics of each antivenom using in vitro IgG titre, avidity and venom-protein specificity assays. We also measured the IgG concentration of each antivenom.
None of the six antivenoms are preclinically effective, at the doses tested, against all of the most medically important snakes of the region. The very limited snake polyspecific efficacy of two locally available antivenoms is of concern. In vitro assays of the abilities of ‘test’ antivenom IgGs to bind venom proteins were not substantially different from that of the ‘gold standard’ antivenoms. The least effective antivenoms had the lowest IgG content/vial.
Manufacture-stated preclinical efficacy statements guide decision making by physicians and antivenom purchasers in sub-Saharan Africa. This is because of the lack of both clinical data on the efficacy of most of the many antivenoms used to treat patients and independent preclinical assessment. Our preclinical efficacy assessment of antivenoms available in Kenya identifies important limitations for two of the most commonly-used antivenoms, and that no antivenom is preclinically effective against all the regionally important snakes. The potential implication to snakebite treatment is of serious concern in Kenya and elsewhere in sub-Saharan Africa, and underscores the dilemma physicians face, the need for clinical data on antivenom efficacy and the medical and societal value of establishing independent preclinical antivenom-efficacy testing facilities throughout the continent.
| Snakebite is one of the most under-researched, under-resourced high morbidty/high mortality NTDs, as reflected by the fact that many of the antivenoms used to treat snakebite victims in sub-Saharan Africa are of uncertain and untested efficacy. This Kenya case study is the first examination of the preclinical efficacy of all available antivenoms to neutralize the venom toxic effects of the most medically important snakes in any region of sub-Saharan Africa. Our results identify serious preclinical efficacy limitations in two of the most commonly used antivenoms, that no single antivenom is effective against all regionally important snakes and that the least effective antivenoms had the lowest IgG concentrations.
It is our aim that Ministry of Health medicine-supply regulators can use this data as evidence to demand more detailed efficacy evidence from manufacturers, and to justify the establishment of national/regional preclinical testing facilities. We hope this publication will also alert physicians treating African snakebite victims to check the efficacy of antivenom in their pharmacies. We have carefully qualified the extent and limitation of the results and of our interpretation of the clinical implications thereof.
| Snakebite annually kills over 95,000 people [1] residing in some of the most disadvantaged rural communities [2], and leaves about 300,000 surviving victims with permanent physical disabilities and stigmatising disfigurements. Since it is the most economically-productive and educationally-vulnerable 10–30 year olds that suffer most, snakebite also poses a significant additional socioeconomic burden on these remote, already impoverished communities. Available mortality data clearly indicate that snakebite deaths are greatest in Asia, and particularly in India [1, 3] followed by sub-Saharan Africa (Table 1). The increasing concern over the plight of sub-Saharan African snakebite victims [4, 5, 6] focuses upon the higher case fatality in sub-Saharan Africa than elsewhere (Table 1) and upon the declining availability of effective antivenom to treat snakebite victims.
The crisis in supply of effective and affordable antivenom to treat snakebite victims in sub-Saharan Africa was first reported in 2000 [7], and has since deteriorated. Akin to the late 1990s market failure of the Behringwerke-manufactured antivenom, Sanofi Pasteur had also supplied Africa with one of the most polyspecifically-effective and widely-used antivenoms, FavAfrique, but ceased its manufacture in early 2016 after a more than a decade of commercial disincentives. This latest market failure of effective antivenom particularly affected snakebite-treatment capability in those state, private (mostly city-based) and charity hospitals that could afford this relatively expensive antivenom ($140/vial; [8]). The SAIMR polyvalent antivenom, manufactured by the South African Vaccine Producers Pty (SAVP), was also widely used and recognised to be highly effective–but outside of the Southern Africa Economic Community it has become more expensive ($315/vial, SAVP, personal communication) than FavAfrique was and, also because of low production volumes, become increasingly difficult to source.
Cognisant presumably of potential commercial opportunities and the public health needs engendered by the snakebite-therapy vacuum in sub-Saharan Africa, several non-Africa based antivenom manufacturers have in the past two decades produced polyspecific antivenoms marketed at costs considerably lower ($18–75) than the FavAfrique or SAIMR antivenoms, and supplied in vastly greater quantities [8]. Superficially, this influx of new, affordable antivenoms into sub-Saharan Africa would seem highly desirable. However, in too many cases and African countries, this has not been the case—because some of these antivenoms have proved dangerously ineffective. Thus, reports from Ghana, Chad and the Central African Republic [9, 10, 11] document an increased case fatality rate (from under 2% to over 12%) following discontinuation of effective antivenoms and introduction of replacement products. In at least one case this was because the antivenom had been manufactured from IgG purified from horses immunised with venoms from Indian snakes–instead of venoms from African snakes [12]. Antivenom efficacy is predominantly restricted to snakes whose venoms were used in manufacture [13]–because the highly snake species-specific protein composition of venom dictates an equally specific IgG response in the immunised horses/sheep. Thus, the greater the biogeographic difference between the venom/s used in antivenom manufacture and the venom injected into the snakebite patient, the weaker the efficacy of the antivenom. For this reason, antivenom manufacturers are required to preclinically test and state the snake species for which their product is effective. Fig 1 evidences another antivenom marketed specifically for Central Africa but clearly inappropriately manufactured with venoms from Asian vipers.
There are very few published reports on the clinical effectiveness of the several antivenoms in current use in sub-Saharan Africa. Preclinical efficacy data is therefore the only information available to physicians and government purchasers to decide which antivenom to use/purchase. However, many, perhaps the majority, of sub-Saharan African countries do not apparently subject newly-imported antivenoms to independent preclinical efficacy and safety testing, and clinicians and purchasers per-force base clinical use/purchase decision making upon manufacture-stated efficacy statements. The above reports of antivenom ineffectiveness and rising case fatalities, seemingly throughout much of sub-Saharan Africa, demonstrate this trust can be misplaced. There is therefore an urgent need to establish independent preclinical antivenom-efficacy testing facilities and expertise in sites throughout sub-Saharan Africa.
With substantive new funding, the Liverpool School of Tropical Medicine has partnered with colleagues in Kenya, Nigeria and Cameroon to form the African Snakebite Research Group and established ‘Snakebite Research and Intervention Centres’ (SRIC) in each of these countries. Our remit includes improving the (i) availability of effective snakebite treatment in rural remote hospitals in greatest need and (ii) access to treatment for snakebite victims. To ensure this new programme is equipped with effective antivenom, and to provide the host government with independent antivenom-efficacy information, we purchased a vial of as many different antivenom brands as available from local pharmacies and preclinically tested their efficacy against venoms of the most medically important snakes in the region.
This first report from the African Snakebite Research Group emanates from Kenya-SRIC activity and demonstrates, for the first time in East Africa, that there is substantial variation in the preclinical efficacy of the available antivenoms against the lethal effects of venoms from black mambas, spitting and non-spitting cobras, puff adders and saw-scaled vipers–and that no one antivenom is preclinically effective at the doses tested, against all these life-threatening snake venoms.
The antivenoms used in this study are described in detail in Table 2 and were acquired from a commercial pharmacy in Nairobi, except the SAIMR antivenoms that were donated to the first author from expired stocks held by Public Health England and had 2012 expiry dates. We were unsuccessful in purchasing one of the ASNA antivenoms manufactured by Bharat Serums and Vaccines Ltd that we had seen in a rural hospital in Kenya. All the antivenoms are manufactured as F(ab')2 fragments of IgG, and for clarity to non-specialist readers we have used the term IgG to describe these antivenoms.
The comparative preclinical efficacy of the ‘test’ antivenoms was conducted before their respective expiry dates. We were unable to purchase the SAIMR polyvalent and ECHIS CARINATUS monovalent ‘gold standard’ antivenom in Kenya and therefore used 2012-expired vials donated to us from Public Health England. SDS-PAGE profiling of these antivenoms (Supplementary S3 Fig) reveal that the IgG in these SAIMR antivenoms possess the same structural integrity as IgG from the ‘in date’ test antivenoms. Further validation of using these expired SAIMR antivenoms for this study is provided by the comprehensive binding of venom proteins by IgG in these antivenoms.
East Africa is resident to multiple medically important vipers, elapids and colubrids. For this analysis, we selected venoms from the most relevant representative species of each genus. Venom was extracted from over four specimens of wild-caught puff adders (Bitis arietans, Kenya); saw-scaled vipers (Echis pyramidum leakeyi, Kenya); black mambas (Dendroaspis polylepis, Tanzania); Egyptian cobras (Naja haje, Uganda); black-necked spitting cobras (N. nigricollis, Tanzania) and red spitting cobras (N. pallida, Kenya) maintained in the Liverpool School of Tropical Medicine herpetarium (a UK Home Office accredited and inspected animal research facility). Freshly collected venom was snap frozen, lyophilised and stored as a powder at 4°C prior to reconstitution in phosphate-buffered saline (PBS). The same batches of these venoms were used for each of the analyses below to provide cross-experiment continuity.
We employed routine protocols in our laboratory [14] to measure the IgG titre, avidity, venom protein-specificity and protein (IgG) concentration/ml antivenom to provide a detailed immunological profile of the ‘gold standard’ and ‘test’ antivenoms.
We used the WHO-recommended antivenom effective dose (ED50) assay, which measures the amount of antivenom required to prevent venom-induced lethality in 50% of mice (5/dose group) injected with venom/antivenom mixtures. To assess the efficacy of the SAIMR ‘gold standard’ polyvalent and monovalent antivenoms, we determined the ED50 dose of (i) the SAIMR ECHIS CARINATUS monovalent antivenom against only the saw-scaled viper venom and (ii) the SAIMR polyvalent antivenom against venoms of the black mamba, the Egyptian, red and black-necked spitting cobras and the puff adder (Table 4).
We assigned the SAIMR antivenoms as the ‘gold standard’ comparators because of their sub-Saharan African clinical effectiveness [17] and because, unlike FavAfrique, these antivenoms are likely to be available for the foreseeable future.
We elected a ‘gold standard comparison’ experimental design to achieve our objective instead of 24 conventional ED50 assays (4 ‘test’ antivenoms tested against 6 venoms), which would have required a minimum of 600 mice (25 mice/experiment). Instead, we tested the extent to which the ‘test’ antivenoms prevented the death of mice (5/dose group) injected with the venom/antivenom mixtures at volumes equivalent to half (0.5x), equal (1x) or, where appropriate, 2.5-fold more (2.5x) the volume of the SAIMR antivenoms that impart 100% survival of the mice (calculated by doubling the ED50 dose volume). Thus, 100% survival of mice injected with 0.5x volume of a ‘test’ antivenom indicates a higher venom-neutralising efficacy than the SAIMR ‘gold standard’ antivenoms, and a test antivenom providing less than 100% efficacy at 1x volumes would be deemed less dose-effective than the ‘gold standard’. Failure of an antivenom to neutralise the lethal effects of a venom at 2.5x volumes in 100% of mice would raise serious concerns as to the potential clinical efficacy of that product.
This protocol enabled us to comprehensively analyse the efficacy, compared to a gold standard, of several antivenoms against several venoms using only 40% of the number of mice had we used conventional ED50 testing (Fig 2). It is important to note that, for the sake of clarity for non-specialist readers, we have described the efficacy of the ‘test’ antivenoms relative to 0.5x, 1x or 2.5x volumes of the SAIMR antivenom that protect 100% of mice. This differs from the more conventional description as ED50 volumes (the volume of antivenom that protects 50% of the venom/antivenom injected mice), and antivenom preclinical testing practitioners are referred to Supplementary S1 Table that depicts the same data in Fig 2 in the context of antivenom volumes (μL) and amounts (mg).
To ensure valid cross-comparison of antivenom-venom reactivity, we first standardised the IgG concentration of each antivenom to 5 mg IgG/ml. We next incubated serial dilutions of each antivenom with the same concentration of each venom (Fig 3). For space reasons we have excluded the graph showing the baseline reactivity of the naive control horse IgG to all the venoms. The OD readings of the antivenoms at the 1:2,500 dilution, in the middle of the downward slope, provide the most immunologically meaningful comparison and, for clarity, are presented as tables in each panel of Fig 3. For example, at this dilution, the SAIMR ECHIS CARINATUS antivenom IgG shows (i) highest binding to the E. p. leakeyi venom, (ii) some cross-reactivity to other viper (puff adder) venom proteins, and (iii) near-zero binding to the four elapid venoms–results entirely consistent with an antivenom generated by immunisation with only Echis species venoms.
While detectable venom-binding differences between the antivenoms exist, none of the antivenoms exhibited sufficiently poor IgG binding to venoms that account for the very poor polyspecific ED50 results of, for example, the VINS and INOSAN antivenoms (Fig 2). Thus, at the same IgG dilution (1:2,500) the OD values (venom-binding) of VINS antivenom to all the venoms was greater or equivalent to that of the more preclinically effective Sanofi Pasteur antivenom. Furthermore, INOSAN’s inefficacy against D. polylepis venom contrasted with its higher OD to this venom than the considerably more efficacious Sanofi Pasteur antivenom. Finally, the OD values of the Premium Serums & Vaccines antivenom was consistently higher to all the venoms than the more efficacious SAIMR antivenoms.
The ELISA IgG titration assay demonstrated that all the ‘test’ antivenoms contained venom-binding IgG titres not dissimilar to the ‘gold standard’ SAIMR antivenoms. Thus, while we identified many discreet IgG-venom binding differences between the antivenoms, we were unable to confidently attribute any of these as being responsible for the very different venom-neutralisation efficacies of these antivenoms. We therefore next performed an assay to identify whether the antivenoms possessed IgGs of variable avidity (binding strength) to the six different venoms that matched their distinct venom-neutralising efficacies.
Ammonium thiocyanate (NH4SCN) is a potent disruptor of protein-protein binding (chaotrope) and, by measuring the ELISA OD readings of the same concentration of IgG and venom in the presence of increasing amounts of the chaotrope, is used to test antivenom IgG-venom protein binding strength.
We incubated 1:1,000 dilutions of each of the 5 mg/ml standardised antivenom solutions with the venoms (in the same concentration as for the IgG titre ELISA) and determined the OD readings after addition of 0, 1, 2, 4, 6 and 8 moles of NH4SCN (Fig 4; we have excluded the graph showing the baseline reactivity of the naive control horse IgG to all the venoms). The most immunologically-informative results were gained by comparing the percentage reduction in OD values of the antivenoms without NH4SCN to that in the middle of the downward slope at 4 M NH4SCN—as illustrated by the table inserted into each panel of Fig 4.
This assay revealed that the Premium Serums & Vaccines (panel 4F) antivenom possesses the most consistent, and highest, cross-snake species IgG-venom binding avidity of all the antivenoms, including the ‘gold standard’ antivenoms. The INOSAN antivenom (panel 3E) exhibits the least consistent cross-species venom binding avidity.
The chaotropic ELISA assay revealed closer links between antivenom IgG-venom binding avidity and antivenom efficacy than the IgG titre results. Thus, the notably higher IgG-binding avidity of the SAIMR polyvalent gold standard (panel 4A) to the B. arietans, N. haje and D. polylepis venoms, and lower avidity to the spitting cobra (N. nigricollis, N. pallida) venoms accurately reflect the venom-neutralising dose-efficacy of this antivenom (Table 4). Also, the substantially higher IgG avidity of the INOSAN antivenom to E. p. leakeyi venom also matched its venom-neutralising dose-efficacy profile.
However, this assay did not provide data accounting for the snake-species distinct venom-neutralising efficacies of the Premium Serums & Vaccines, VINS and Sanofi Pasteur antivenoms, particularly in comparison with the superior venom-neutralising dose-efficacy of the SAIMR polyvalent antivenom against N. haje, D. polylepis and B. arietans venoms.
Venoms consist of multiple distinct protein groups and the protein composition of venom is markedly snake genus/species specific–with obvious implications on the venom protein-specificity of IgGs from venom-immunised horses. Neither the IgG titre nor the IgG avidity ELISA assays are designed to examine IgG binding to specific venom proteins. We therefore next used an immunoblot assay to investigate whether the six antivenoms express differences in IgG venom protein specificities that match their distinct venom-neutralising efficacies.
Fractionation of the six snake venoms by 15% SDS-PAGE revealed the numerical and molecular size diversity of the venom proteins (Fig 5A), with the cobra and mamba venoms (N. nigricollis, N. pallida, N. haje and D. polylepis) possessing a greater abundance of the low molecular mass neurotoxins/cytotoxins than the more evenly distributed molecular mass of the enzyme-rich, haemostasis-disruptive viper venoms (E. p. leakeyi and B. arietans). To determine the extent to which this wide spectrum of East African snake venom proteins are bound by IgG of the six antivenoms, we electrophoretically transferred these venom proteins to a membrane and incubated those (under identical conditions) with the antivenoms at 1:5,000 dilutions (Fig 5). For this test, we did not adjust the antivenoms to a standard 5 mg/ml concentration.
This analysis demonstrated that the intensity of the IgG-venom protein binding of the ‘gold standard’ SAIMR polyvalent and SAIMR ECHIS CARINATUS antivenoms was notably greater than all the ‘test’ antivenoms (Fig 5). It is important to note that, with minor brand-specific differences, that the difference between the ‘test’ and ‘gold standard’ antivenoms in this assay related to the intensity of venom protein binding and not the protein specificity. Thus, although the intensity of the IgG-venom protein binding was lower with the ‘test’ antivenoms, the immunoblots revealed that each antivenom possessed IgGs with similar venom protein specificities as the SAIMR ‘gold standard’ antivenoms. The lack of venom-reactivity of the control, naïve horse IgG (Fig 4B), evidences the venom-specificity of the antivenom IgGs.
The three immune assays above identified detectable differences in the IgG titre, avidity and venom-protein specificities of the six antivenoms, but these differences were relatively minor and could not be consistently applied to the interaction of each antivenom with each venom. We were therefore unable to identify an IgG-venom binding deficiency with any of the ‘test’ antivenoms, and by inference a deficiency in their venom-immunisation protocols, that could account for the ineffectiveness/weak efficacy of some of the ‘test’ antivenoms in our preclinical assays. The substantially higher IgG-venom protein binding intensity of the SAIMR polyvalent antivenom in the immunoblot assay suggested to us that this antivenom may be formulated with a higher amount of IgG/vial than the others.
Our final test was therefore to determine the protein (IgG) concentration (all antivenoms are formulated as F(ab’)2 fragments of IgG) of each antivenom because this has an obvious bearing on dose-efficacy, and because it was not stated by any manufacturer, despite it being the active component of these therapies.
We used a spectrophotometric instrument (NanoDrop) to measure protein content of each antivenom (in triplicate) and the results are presented in Table 5. We included control horse IgG of different known concentrations, to confirm the accuracy of the NanoDrop results. For the sake of completeness, we also used SDS-PAGE analysis of each antivenom to demonstrate their consistent IgG purity (Supplementary S3 Fig).
This analysis demonstrated the substantial inconsistency in the total IgG content of the antivenoms, and, importantly, that the VINS and INOSAN antivenoms respectively contained 19% and 28% of the IgG concentration of the SAIMR polyvalent antivenom. The IgG content of antivenom therefore exhibited the closest association to their comparative efficacy in neutralising the venoms of N. nigricollis, N. pallida and E. p. leakeyi (Fig 2). It is not however a universally-applicable explanation, because, for example, IgG content alone does not explain the superior anti-B. arietans and anti-E. p. leakeyi efficacy of Premium Serums & Vaccines (63 mg/ml) antivenom over that of the Sanofi Pasteur (96.7 mg/ml) antivenom. Assessing the antivenom efficacy data by mg antivenom did reveal that the anti-E. p. leakeyi venom efficacy of the INOSAN antivenom was equivalent to that of the SAIMR ECHIS CARINATUS ‘gold standard’ antivenom. Nevertheless, as stated above and depicted in Fig 2, it required 2.5 times more the volume of the SAIMR antivenom for the INOSAN product to achieve this efficacy parity, which in the human-treatment context has cost and adverse effect implications.
For completeness, we have presented the amount (mg) and volume (μl) of antivenom used for each of the ‘test’ antivenoms for each of the doses examined into Supplementary S1 Table. To facilitate comparison, we have also added the amount (mg) and volume (μl) of the calculated 2xED50 doses of the ‘gold standard’ antivenoms that provided 100% protection to the envenomed mice. This table therefore provides all the numerical data related to the preclinical assays, and interpretations of efficacy from this table is no different from that presented in Fig 2. No matter whether the data is examined by antivenom volume in μl or amount in mg, the least polyspecifically-effective antivenoms simply do not compare well to the ‘gold standard’ and to some of the other ‘test’ antivenoms. It was important that we tested the antivenoms by volume because that is the formulation in which the antivenom is provided by the manufacturer and used by the clinician. It is near impossible to envisage a clinician calculating the dose volume of antivenom he/she is going to administer based upon the mg/ml antibody content (and perhaps impossible because, as here, manufacturers rarely provide this information). Therefore, to ensure that our preclinical efficacy testing of the antivenoms is of value to clinicians and medicine-purchasing agencies in East Africa, we undertook this testing, and report the results by antivenom volume.
The reality is that the efficacy of a monospecific antivenom to its homologous venom is dictated by multiple factors, including IgG concentration, titre, avidity and protein specificity, which are themselves affected by the quality of the immunising venoms, the quantitative ratio of the venoms used for immunisation, the selected adjuvant and other aspects of the immunisation and antivenom-manufacturing protocols. All these interlocking factors are made more complex, as here, by increasing the number of venoms used to manufacture polyspecific antivenom, making it very difficult/impossible to confidently assign any one factor as being primarily responsible for lack of efficacy.
The plight of snakebite victims, particularly those in sub-Saharan Africa has been the subject of considerable recent attention [see 6, 8, 19, 20, 21] and the focus of recent Wellcome Trust and the Kofi Annan Foundation sponsored international meetings to identify remedial interventions. Reports from both meetings [4, 5] identify the urgent need for the provision of effective antivenom and preclinical efficacy testing of existing and new antivenoms to ensure this. This recommendation aligns fully with the new WHO initiative to establish a prequalification programme for African antivenom, which includes establishing venom standards for sub-Saharan Africa and using those for preclinical antivenom-efficacy testing [22]. Over the past five years, we have generated a comprehensive inventory of sub-Saharan African snake venoms of defined gene and protein composition and murine toxicity [23, 24, 25]. This unique resource has enabled the Kenya Snakebite Research & Intervention Centre to conduct this first preclinical assessment of the efficacy of antivenoms available for clinical use in Kenya.
This is important because the market failure of the Sanofi Pasteur antivenom and the very high costs of the SAIMR antivenoms leaves many East African countries with a choice of polyspecific antivenoms restricted to brands for which there is very little/no published data on their human effectiveness. Our preclinical results illustrate that the SAIMR polyvalent antivenom is considerably more effective in neutralising the murine lethality of the Egyptian cobra (N. haje), black mamba (D. polylepis) and puff adder (B. arietans) venoms than any of the ‘test’ antivenoms. The ‘test’ antivenoms exhibited a superior or equal dose-efficacy as the SAIMR polyvalent antivenom against the spitting cobra venoms (N. nigricollis and N. pallida). Preclinical neutralisation of the saw-scaled viper (E. p. leakeyi) venom was achieved by the Premium Serums & Vaccines and INOSAN antivenoms but required 2.5-fold greater volumes than the SAIMR ECHIS monovalent antivenom. Perhaps the most important result of our study was that no single antivenom, at the doses tested (see below for a detailed consideration of this assay), was effective in neutralising the murine lethal effects of venoms of all six medically important snakes of East Africa—despite the pan-African efficacy claims inherent in the names of many of these products. The snake species-specific dose efficacy of each of the ‘test’ and ‘gold standard’ SAIMR polyvalent antivenoms suggests that the clinical management of envenoming by these snakes with any one of these antivenoms may require distinct, snake species-specific dose regimens. In the absence of a rapid snake species diagnostic test, this is clinically problematic and will likely result in the administration of too little or too much antivenom–both highly undesirable scenarios resulting in either inefficacy or increased risk of antivenom-induced adverse effects, respectively.
On a more positive note, the preclinical efficacy of the Premium Serums & Vaccines product matched that of the highly regarded, but now unavailable Sanofi Pasteur product and approached that of the expensive SAIMR polyvalent antivenom. The Premium Serums & Vaccines product, at 26% of the cost of the SAIMR antivenoms (see Table 6), was the most affordable and effective antivenom of those tested here, although we note that preclinical efficacies against the two neurotoxic snake venoms (N. haje and D. polylepis) were weaker than against the other four species. Our results suggest that the preclinical efficacy of the VINS and INOSAN products could possibly be substantially improved by simply increasing the amount of IgG in each vial. The INOSAN product was the most expensive of the ‘test’ antivenoms, the most preclinically effective at neutralising the saw-scaled viper venom and one of the less effective antivenoms at neutralising the lethality of other snake venoms. It is notable that the IgG avidity of this antivenom to venoms of the six snakes varied more than the other antivenoms (Fig 4), and this avidity profile matched its snake-specific venom-neutralising efficacy. This may suggest that changes to the venom-immunisation regimen or analysis of the quality of the venoms could improve the venom-binding avidity and perhaps the efficacy of this antivenom. In consideration of the clinical efficacy of the SAIMR polyvalent and ECHIS monovalent antivenoms, it would be interesting to know whether SAVP has plans to combine the venom-immunising mixtures of these products to produce a truly pan-African polyspecific antivenom.
We have carefully qualified our interpretation/extrapolation of the results of our preclinical ‘gold standard’ comparison assays to the efficacy of antivenom treatment of human snakebite patients, and urge readers to be similarly cautious. Our preclinical protocol enables a rapid efficacy comparison of a matrix of six venoms and four antivenoms using the minimum number of mice, but it differs from the recommended WHO antivenom-efficacy testing protocols in that it does not provide an ED50 value for each ‘test’ antivenom against each venom. Thus for example, we are unable to state whether, or not, the polyspecifically effectiveness of the VINS and INOSAN products could attain 100% efficacy against more of the venoms by using substantially greater volumes–because of the volume constraints inherent to this murine assay. From a human-treatment perspective, the necessity to administer multiple vials of an antivenoms carries important treatment costs (the INOSAN product was the most expensive ‘test’ antivenom) and adverse effects issues that risk poor uptake in rural tropical regions.
On a more general note, these murine preclinical antivenom-efficacy testing assays are not infallible predictors of human efficacy. A recent study in Sri Lanka has questioned the value of predicting efficacy outcomes in human patients from ED50 results [26]. Conversely, the undoubted ability of the ED50 test to discriminate between effective and ineffective antivenoms in the murine model [13, 27], and the effective use of ED50 data [15] to help design the dose regimen of a human antivenom clinical trial [28] suggests that while the ED50 preclinical test has inadequacies many practitioners share, it should be retained until more accurate assays are tested, validated and become available. Thus, while the results of this study identify a potentially serious therapeutic concern, one of the priorities of the Kenya, Nigerian and Cameroon Snakebite Research & Intervention Centres will be to undertake conventional ED50 tests so that the Ministries of Health can be provided with pharmacopoeia-compliant data that they can act upon to restrict human use of preclinically-ineffective antivenom.
The dialogue above underscores the urgent need for more published data on the efficacy of the different antivenom brands in use in sub-Saharan Africa to treat human patients. The paucity of this information reflects the prohibitive expense, problems of recruiting sufficient patients envenomed by the myriad of venomous snake species, the lack of accurate diagnostic tools to distinguish such species-distinct envenoming, and time required to conduct full clinical trials. Until the funding and required clinical/diagnostic tools become available, we remain reliant upon clinical observation studies that have importantly identified the wide-spread use of some dangerously ineffective antivenoms [9–11]. However, only the first of these reports provided the clinically-vital information on the antivenom brand. Another objective of the recently-established African Snakebite Research Group will be to conduct surveys in many rural hospitals in Nigeria, Kenya and Cameroon experiencing high snakebite admissions. The outcomes of these surveys will include reporting antivenom availability, and assessments of the clinical outcomes of treating patients with the various brands of locally-available antivenom.
In conclusion, this first report of the African Snakebite Research Group identifies a worrying differential in the preclinical efficacy of available antivenoms in Kenya and underscores the need for independent preclinical testing of antivenoms throughout sub-Saharan Africa, and the need for venom standards for all the most medically important snakes to resource this testing. There is also an urgent need for practitioners to publish data on the clinical outcomes of treating patients with brand-named antivenoms (despite the inherent problem of such transparency). The widespread availability of pan-African preclinical testing and clinical observation information will substantially help to improve the effectiveness of snakebite management and thereby reduce the high case fatality currently suffered by sub-Saharan African snakebite victims.
|
10.1371/journal.pgen.1006587 | Genetic variants alter T-bet binding and gene expression in mucosal inflammatory disease | The polarization of CD4+ T cells into distinct T helper cell lineages is essential for protective immunity against infection, but aberrant T cell polarization can cause autoimmunity. The transcription factor T-bet (TBX21) specifies the Th1 lineage and represses alternative T cell fates. Genome-wide association studies have identified single nucleotide polymorphisms (SNPs) that may be causative for autoimmune diseases. The majority of these polymorphisms are located within non-coding distal regulatory elements. It is considered that these genetic variants contribute to disease by altering the binding of regulatory proteins and thus gene expression, but whether these variants alter the binding of lineage-specifying transcription factors has not been determined. Here, we show that SNPs associated with the mucosal inflammatory diseases Crohn’s disease, ulcerative colitis (UC) and celiac disease, but not rheumatoid arthritis or psoriasis, are enriched at T-bet binding sites. Furthermore, we identify disease-associated variants that alter T-bet binding in vitro and in vivo. ChIP-seq for T-bet in individuals heterozygous for the celiac disease-associated SNPs rs1465321 and rs2058622 and the IBD-associated SNPs rs1551398 and rs1551399, reveals decreased binding to the minor disease-associated alleles. Furthermore, we show that rs1465321 is an expression quantitative trait locus (eQTL) for the neighboring gene IL18RAP, with decreased T-bet binding associated with decreased expression of this gene. These results suggest that genetic polymorphisms may predispose individuals to mucosal autoimmune disease through alterations in T-bet binding. Other disease-associated variants may similarly act by modulating the binding of lineage-specifying transcription factors in a tissue-selective and disease-specific manner.
| Research to date has identified many genetic variants that are more common in people with a particular disease. However, in conditions that reflect multiple genetic and environmental factors, it is difficult to know with certainty if and why any particular genetic variant is causative and the mechanism that may underlie this. Such variants are often outside of protein-coding exons, instead falling in regions that regulate gene expression. In these cases, the genetic variation may alter transcription factor binding and subsequent gene expression. In this study, we have examined how genetic variation affects T-bet binding to DNA, as a key transcriptional regulatory mechanism in the immune response. An inability to mount this response effectively can result in increased susceptibility to infections or cancer, while a response that is too strong, or wrongly targeted, can result in uncontrolled/chronic inflammatory and autoimmune conditions. We have found that T-bet binding sites are specifically enriched in genetic variants associated with the mucosal autoinflammatory diseases UC, Crohn’s disease and celiac disease. We also identify genetic variants that alter T-bet binding and gene expression. This discovery thus identifies a molecular mechanism through which genetic variants can be associated with increased risk of mucosal autoimmune disease.
| The differentiation of naïve CD4+ T cells into distinct T helper cell (Th) lineages is essential for adaptive immunity. The original paradigm of interferon-gamma (IFN-γ) producing T-helper 1 (Th1), and type-2 (Interleukin 4, 5, and 13) cytokine producing Th2 cells has expanded to include both Interleukin-17 (IL-17) producing Th17 and anti-inflammatory T-regulatory (Treg) cells. Th cell differentiation is controlled by a set of master regulatory or lineage-specifying transcription factors, with the T-box family member T-bet necessary and sufficient for Th1 cell differentiation. GATA3, RORγT and FOXP3 perform parallel roles in Th2, Th17 and Treg cells, respectively [1]. Importantly, T-bet inhibits alternative lineage fate specification, repressing both the Th17 and Th2 lineages [2–4].
Inappropriate Th cell activation and polarization can lead to autoimmunity. Worldwide, autoimmune and auto-inflammatory diseases are now estimated to affect nearly 10% of the population [5]. The incidence of inflammatory bowel diseases (IBD), including Crohn’s disease and UC, and celiac disease, is rising rapidly, with more than 1.4 million people affected in the USA alone [6]. A role for T-bet is particularly apparent in the mucosal immune system and has been linked to IBD and celiac disease [7]. The expression of T-bet is upregulated in lamina propria T cells of patients with Crohn’s and celiac disease and ex vivo culture of biopsies from untreated celiac patients with gliadin increases T-bet expression through STAT1 activation [8,9]. In addition to this, it is now apparent that mucosal inflammation is also driven by IL-17, which is enhanced by IL-23 receptor signals in effector T cells [10]. Loss of T-bet in the innate immune system leads to a transmissible form of ulcerative colitis in the TRUC (T-bet and Rag deficient Ulcerative Colitis) model, driven by transcriptional derepression of TNF in colonic mononuclear phagocytes [11–13]. This susceptibility has also been shown to be dependent on IL-17 and mediated via repression of IL-7 receptor expression by T-bet in innate lymphoid cells (ILCs) [11]. T-bet has subsequently been shown to play a role in the development of the NKp46+ CCR6- subset of IL-22 expressing ILCs that, in turn, are important for protecting the epithelial barrier during Salmonella enterica infection [14,15].
Autoimmune diseases cluster in families, suggesting a large genetic component [16]. Genome-wide association studies (GWAS) have identified hundreds of risk loci for autoimmune diseases, including for IBD and celiac disease [16–23]. The majority of autoimmune disease-associated SNPs lie outside of gene coding regions in intergenic or intronic regions [24]. This can make it challenging to understand the molecular basis of how a genetic variant predisposes to disease. Furthermore, the causal variant can be difficult to identify from the large clusters of SNPs in linkage disequilibrium that tend to be identified by GWAS. Thus, efforts have been made to identify SNPs located within regulatory elements marked by open chromatin, histone modifications associated with active enhancers or known/predicted transcription factor binding sites [21,24–32]. Some of these variants have been shown to modulate transcription factor binding or epigenetic regulation. Genetic variants that alter DNase I hypersensitivity [27,33,34], DNA methylation [35–38], histone modification [27,39–43], and the binding of transcriptional regulators to DNA [27,33,34,44–51], have been identified, suggesting potential causal mechanisms.
Although previous studies have demonstrated enrichment of transcription factor binding sites at disease-associated polymorphisms, whether specific disease causing variants act to alter the binding of T cell lineage-specifying factors has not been investigated. Having previously mapped T-bet binding across the genome in human Th1 cells [52–54] we used a systematic functional GWAS (fGWAS) approach to determine the degree to which disease-associated SNPs were enriched within T-bet binding sites. SNPs were then tested for effects on T-bet binding in vitro using a novel flow cytometric assay and in vivo by allele-specific ChIP-seq. We report here that SNPs associated with mucosal inflammatory diseases are selectively enriched at T-bet binding sites. Furthermore, we show that the celiac disease associated variants of rs1465321 and rs2058622, and the IBD-associated variants of rs1551398 and rs1551399, exhibit decreased T-bet binding in vivo. We further demonstrate that the genes associated with these SNPs, IL18RAP and TRIB1, respectively, are transcriptionally regulated by T-bet and that rs1465321 is an expression quantitative trait locus (eQTL) for IL18RAP. Taken together, these data mechanistically link alterations in T-bet binding to disease predisposition.
To identify disease-associated polymorphisms at T-bet binding sites, we compared the locations of GWAS hits listed in the National Human Genome Research Institute (NHGRI) catalogue [55] with binding sites for T-bet in primary human Th1 cells [52–54]. As the published trait-associated SNP may not be the most functionally relevant [28], SNPs in high linkage disequilibrium LD (r2 >0.8) were also examined. This returned a list of 926 unique SNPs located at T-bet binding sites (hereafter referred to as T-bet hit-SNPs; Fig 1A and 1B, S1A Fig and S1 Table). In line with previous reports, a minority (143) of the T-bet hit-SNPs were the putative causal SNP from GWAS data, with the others being in high LD (total of 621 independent LD blocks). Examination of the location of T-bet hit-SNPs in relation to protein-coding genes revealed that the majority (63%) were distal (>1 kb) to gene promoters. As expected, H3K27ac and DNaseI hypersensitivity were highly enriched in Th1 cells at T-bet hit-SNPs compared with all disease-associated SNPs, consistent with these being located within active regulatory elements (Fig 1C).
As T-bet is only expressed in cells of the immune system, we hypothesised that T-bet hit-SNPs would be primarily associated with autoimmune diseases. To test this, we used fGWAS [56], a hierarchical model that assesses relative enrichment of GWAS associations within various functional elements. This model splits the genome into large blocks (larger than regions of linkage disequilibrium), assesses whether each block contains a SNP associated with the trait of interest or not and then searches among supplied functional annotations for those that improve the likelihood of predicting the presence of a trait-associated SNP, finally predicting which SNP in the block is most likely causal.
To test whether disease-associated SNPs were enriched at T-bet binding sites, we gathered GWAS data for the Th1-associated auto-inflammatory conditions celiac disease, Crohn’s disease, UC, rheumatoid arthritis (RA), psoriasis and, as a negative non-immune control, coronary artery disease (Fig 2). We compared T-bet binding sites with a number of other relevant functional annotations, including Th1 and Th2 cell DHS [57], H3K27ac [58], and sites of histone modification and transcription factor binding in immune cell lines from the ENCODE project [26] and other sources (S2 Table). Notably, we found that SNPs associated with all of the mucosal immune-mediated diseases tested (Crohn’s disease, UC and celiac disease) were enriched at T-bet binding sites, with the effect in Crohn’s disease being particularly striking. Enrichment at T-bet binding sites was similar to, or stronger than, DHS and H3K27ac and, in the case of Crohn’s and celiac disease, stronger than any other sets of transcription factor binding sites. As expected, SNPs associated with coronary artery disease were not enriched at T-bet binding sites. Of interest, no enrichment for T-bet binding sites was observed for RA- or psoriasis-associated SNPs, suggesting a specific role for altered T-bet binding in mucosal inflammatory disease.
To confirm that T-bet binding is enriched at IBD-associated SNPs, we compared T-bet binding sites to a set of credible SNPs identified at 94 IBD-associated loci [21]. We found that T-bet binding sites were more highly associated with credible SNPs than other SNPs at the same loci (93 bound by T-bet, p = 1.4x10-5, Fisher exact test). Furthermore, within the set of credible SNPs, the higher the posterior probability for causality, the more likely that the SNP overlapped a T-bet binding site (p = 6.3x10-6, continuous binomial regression, S1B Fig). The association of T-bet binding with causal SNPs is highlighted by the finding that, of the 93 credible SNPs bound by T-bet, 11 are the lead variants for their loci. Three of these (rs74465132, rs1887428 and rs61839660) have a posterior probability for causality of greater than 95%. These data suggest that the strong association of these SNPs with IBD is related to T-bet binding at these sites.
Having identified a set of SNPs overlapping T-bet binding sites, we next asked whether these sequence variants altered T-bet binding. The traditional pull-down technique is time intensive and semi-quantitative. Therefore, we explored whether transcription factor binding could be assayed using a flow cytometric readout. In this technique, which we call OligoFlow, a fluorochrome-labelled antibody for the transcription factor of interest is added to the oligonucleotide-bead / lysate mix, and the Median Fluorescence Intensity (MFI) of the beads is assessed by flow cytometry as a quantitative measure of binding efficiency (Fig 3A).
To validate this new technique, a positive control oligonucleotide (Motif+) was designed to incorporate the previously identified consensus sequence [54] surrounded by non-specific sequence (S3 Table). A negative control oligo (Motif-) incorporated mutations of two key residues within the motif. OligoFlow was conducted with lysate from either the YT human cell line, which constitutively expresses T-bet [59], or lysate from primary human CD4+ cells polarised under Th1 conditions in culture. The positive and negative control oligonucleotides showed a clear difference in MFI (Fig 3A) and thus OligoFlow can successfully discriminate positive and negative transcription factor binding events.
We then proceeded to test a subset of our T-bet hit-SNPs that were also associated with H3K27ac or near genes of immunological interest. SNPs that showed differential binding were tested at least five times. Within each experiment, the MFI of each allele was normalised to the MFI of the negative control and significantly altered binding between alleles across all experiments was assessed using a paired t-test. Three T-bet hit-SNPs exhibited significantly different binding to the two alleles; rs1465321, located within the second intron of IL18R1, rs1006353, 22.5 kb upstream of MTIF3, and rs11135484, within an intron of ERAP2 (Fig 3B). Differential T-bet binding to the two alleles of rs1465321 were confirmed by traditional oligonucleotide pull-down (S2 Fig). All 3 SNPs are [A/G] with A as the minor allele. In each case, allele A is also in LD with alleles associated with for the trait under investigation. rs1465321 is in high LD with multiple SNPs associated with celiac disease, including rs13015714 and rs917997, identified as the strongest risk alleles for celiac disease in 2q12.1 [18,60], with the disease-associated alleles linked to reduced IL18RAP expression [60]. rs1465321 and rs11135484 have also been associated with Crohn’s disease [18,22,60,61], but not in a more recent study [21]. For rs1465321 and rs1006353, the minor disease-associated A allele binds T-bet less strongly than the G allele (Fig 3C–3F). In contrast, for rs11135484, the A allele binds T-bet more strongly than the G allele (Fig 3G and 3H). We conclude that disease-associated genetic variants can alter T-bet binding to DNA in vitro.
Motif analysis has often been used to predict transcription factor binding sites affected by genetic variants. We previously derived a consensus T-bet motif from T-bet binding sites in human Th1 cells [54] and repeated this analysis with duplicate T-bet ChIP-seq data (Fig 4). The three T-bet hit-SNPs that showed altered binding in OligoFlow were then examined for whether they disrupted such a T-bet binding motif. In the case of rs1006353, the G allele formed part of a T-bet binding motif, whereas the A allele abolished this binding site (Fig 4B). However, neither of the other two SNPs, rs1465321 and rs11135484, overlapped a predicted T-bet binding motif (Fig 4B). Thus, over-reliance on motif analysis can result in SNPs with the potential to alter transcription factor binding sites being missed and highlights the importance of using experimental validation to confirm binding of the relevant transcription factor.
We next sought to confirm that T-bet exhibited differential binding to disease-associated SNPs in vivo. We focused on rs1465321, because it lies within the IL18R1/IL18RAP gene locus that we have previously identified as a T-bet target [54] and because disease-associated alleles in high LD are associated with reduced IL18RAP expression and disease [60]. Primary naive CD4+ T cells were purified from the peripheral blood of two individuals heterozygous for this SNP and were polarised into the Th1 lineage. We then performed ChIP-seq for T-bet in these cells, as previously described [54]. We aligned the reads for the T-bet ChIP-enriched DNA and input controls to the reference human genome and then counted the number of reads matching the major or minor alleles in the inputs and ChIP samples. In the input DNA samples, there were approximately equal numbers of reads for the two alleles in both individuals. In comparison, the T-bet ChIP reads showed significantly lower enrichment for the minor A allele in both donors (Fig 5A and 5B). There was also a significant allelic imbalance for T-bet binding at the neighbouring SNP rs2058622, which is in high LD (r2 = 1.0) with rs1465321 (Fig 5A and 5B). To determine whether T-bet exhibited allelic imbalanced binding at any other loci, we identified all SNPs that exhibited heterozygosity in both individuals. Of the heterozygous SNPs that overlapped a T-bet binding site, 19 exhibited significant allelic imbalanced binding in both donors after adjustment for multiple hypothesis testing (Fig 5C, S4 Table). These included the IBD-associated SNPs rs1551398 and rs1551399 [21], situated 86bp apart and downstream of TRIB1 (Fig 5C, S1 and S3 Figs). We conclude that the two alleles of rs1465321 exhibit different levels of T-bet binding in vivo, with the disease associated A allele bound significantly less, and that the credible IBD variants rs1551398 and rs1551399 also influence T-bet binding.
Having identified rs1465321, rs2058622, rs1551398 and rs1551399 as disease associated SNPs that modulate T-bet binding in vivo, we next determined whether there was a functional relationship between T-bet binding and the genes associated with these SNPs. rs1465321 and rs2058622 are in high LD with SNPs associated with low expression of IL18RAP in celiac disease [60]. The IBD-associated SNPs rs15513998 and rs1551399 are associated with TRIB1 [21]. To determine whether there was a functional relationship between T-bet binding and IL18RAP and TRIB1 expression, we compared gene expression profiles of wild type and T-bet-/- naïve CD4+ T cells polarised in Th1 conditions. As expected, genes known to be positively regulated by T-bet were significantly downregulated in T-bet-/- cells, including Interferon-γ (Ifng) and Tim-3 (Havcr2), while the housekeeping genes Gapdh, Actb and Hprt remained unchanged (Fig 6A). Il18rap was also significantly downregulated in the absence of T-bet, implying a positive regulatory role for T-bet in modulating its expression (Fig 6A). In contrast, Trib1 was significantly upregulated in T-bet-/- cells, implying that T-bet functions to repress this gene. Consistent with a direct role for T-bet in regulating Il18rap and Trib1 expression, multiple T-bet binding sites were located within intronic regions of murine Il18rap and downstream of Trib1 (S4 Fig). Thus, these data support a direct role for T-bet binding in the regulation of IL18RAP and TRIB1 expression.
We next explored whether the genotype of rs1465321 could control the expression of nearby genes and how this potential eQTL related to celiac disease susceptibility (Fig 6B). Celiac disease association was based on a case control association study of 12,041 celiac disease cases and 12,228 controls [23]. Using a gene expression dataset of 1,214 samples [62] we found a strong correlation between rs1465321 genotypes and IL18RAP expression level (p<10−100, Fig 6C). No other gene showed a significant association with rs1465321. However, this SNP did not display the greatest eQTL association compared with other variants in the region, which could suggest a lack of a causal role. Moreover, using a previously developed methodology [63], we established that the eQTL and disease association signals in the IL18RAP regions were unlikely to be driven by the same genetic variant (posterior probability supporting a shared variant < 1%, Fig 6C). However, a stepwise regression analysis of the eQTL data shows that after accounting for the primary eQTL signal (conditional on rs1985329), a second eQTL association was clearly detectable (p<10−30). This suggested that at least two independent variants, with distinct biological mechanisms, are controlling IL18RAP mRNA expression. Interestingly, this secondary eQTL signal co-localized with the celiac disease risk signal (Fig 6D, posterior probability supporting a shared variant > 99%). Moreover, rs1465321 is one of the most strongly associated genetic variants for this secondary eQTL signal, with the disease-associated A allele, which exhibited reduced T-bet binding, associated with reduced IL18RAP expression. Therefore, our combined fine-mapping disease eQTL data are consistent with rs1465321 affecting IL18RAP expression through altered binding of T-bet.
We have found that IBD and celiac disease-associated SNPs are significantly enriched at T-bet binding sites. Surprisingly, this association is not observed for RA or psoriasis, suggesting it may be specific for mucosal inflammatory disease. Furthermore, we have identified genetic variants that alter T-bet binding to DNA, both in vitro and in vivo, including rs1465321, which we also identify as an eQTL for IL18RAP and celiac disease. Thus, these data provide a mechanistic explanation for why a single base change at this locus is associated with changes in gene expression and disease risk.
Although some studies have identified sequence variants that modulate transcription factor binding, alterations in the binding of Th lineage-specifying factors at disease-associated variants has not previously been identified. Our discovery that SNPs associated with IBD and celiac disease alter T-bet occupancy reveals that genetic variants can have a significant impact on the function of key master regulator transcription factors that govern cell fate. The strong association of T-bet binding sites with mucosal autoimmune/inflammatory diseases suggests that other disease-associated variants also act to alter the binding of this critical immune regulator, with important consequences for T cell polarisation and lineage-specific gene expression.
That T-bet binding sites are associated with mucosal autoimmune disease, but not with RA or psoriasis is somewhat surprising, because all of these diseases have been linked to aberrant Th1 responses [2]. However, mucosal disease is more strongly associated with aberrant Th17 responses, which are repressed by T-bet [3,4,64,65], providing a mechanistic rationale for our findings. We and others have recently shown that T-bet plays a critical and non-redundant role in the function of ILCs [2,7,11–15]. It is therefore feasible that the association of mucosal autoimmune disease-associated SNPs with T-bet binding sites reflects alterations to T-bet binding in ILCs, which have a key regulatory role at mucosal surfaces. Expanding our fGWAS analysis to other autoimmune conditions will be necessary to fully establish the specificity of the association of T-bet with SNPs associated with mucosal disease.
Significantly, we have demonstrated that T-bet binding is enriched at disease-associated SNPs that have high posterior probabilities [21]. This suggests that more T-bet bound variants will be discovered when other IBD loci are subjected to fine-mapping analysis. We further found that the disease-associated alleles of rs1551398 and rs1551399 both reduce T-bet binding in vivo. These SNPs are located upstream of TRIB1, a gene that is upregulated in the mucosa of both UC and CD patients [66]. Consistent with this, we find that T-bet functions to repress Trib1 expression, suggesting that the disease-associated alleles may increase disease risk by abrogating T-bet-mediated repression of this gene. T-bet also binds at 2 other sites near TRIB1 (rs28510097 and rs1551400) and, together, these 4 SNPs account for 55% posterior probability of association for this locus [21].
We also identified rs1465321, located within an intron of IL18R1, to exhibit allele-imbalanced T-bet binding. This SNP is an eQTL for IL18RAP and celiac disease risk, with the minor disease-associated allele linked with reduced T-bet binding and IL18RAP gene expression. IL18RAP and IL18R1 together form the IL-18 receptor. Signaling through this receptor, IL-18 synergizes with IL-12 to induce IFNγ. rs1465321 is in high LD with the lead SNP in this locus for celiac disease [60]. Although our data are consistent with rs1465321 altering IL18RAP expression through altered binding of T-bet, we cannot rule out that variants in strong LD with rs1465321 could also be causal, such as rs2058622 that also exhibits allele-imbalanced T-bet binding. Given that T-bet acts through multiple sites to regulate its target genes [52, 54, 67, 68], it is likely to be the combined effect of the haplotype that is relevant. ChIP-seq for T-bet in individuals heterozygous for other disease-associated SNPs will likely reveal further examples of genetic variants that modulate T-bet binding.
Our finding that there are two independent eQTLs for IL18RAP, and that only one of these is associated with celiac disease (Fig 6), suggests that the level of IL18RAP expression may not be functionally relevant for disease susceptibility. Alternatively, it is possible that the two independent eQTLs for IL18RAP represent different enhancers that mediate transcriptional activation in different cells or in response to different stimuli, and that IL18RAP expression level is only relevant for celiac disease in one cell type or in response to a particular signal.
Attempts to determine the likely effect of non-coding sequence variants have mostly focused on identifying overlapping transcription factor binding motifs or overlapping sites of transcription factor binding, DNase I hypersensitivity or DNA and histone modification. Our analysis of allele-specific T-bet ChIP-seq data shows that genetic variants within transcription factor binding sites do not necessarily alter transcription factor binding. Similarly, genetic variants that do impact transcription factor binding do not necessarily lie within the predicted motif. Thus, confirmation of allele-specific binding events is necessary to confirm that a SNP does indeed impact transcription factor function and provides a mechanistic link between genetic variation and disease risk.
We have established the feasibility of using flow cytometry to assay allelic effects on transcription factor binding, and validated this technique through both the traditional pull-down assay and allele-specific ChIP-seq. As flow cytometric methods can be easily automated, this method provides a more rapid means to assay large numbers of allelic variants compared to traditional pull-down methods. Using this OligoFlow method, we identified alterations in T-bet binding at rs11135484, in high LD with a SNP associated with Crohn’s disease and with rs1006353, the closest neighbor of which is MTIF3, associated with body mass index [69]. Interestingly, T-bet has been linked with regulation of insulin sensitivity and visceral adiposity [70].
In summary, we have identified a specific association between T-bet binding sites and mucosal autoimmune disease variants and determined that such genetic variants modulate T-bet binding in cells. This suggests that altered binding of T cell master regulators can predispose individuals to specific autoimmune and inflammatory conditions. This study establishes a scalable method that can be used to explore the impact of genetic variation on the function of other lineage-specifying transcriptional factors. These insights will identify molecular mechanisms that underlie the genetic basis of autoimmune diseases and suggest new therapies for their treatment.
ChIP-seq for T-bet in human Th1 cells was performed previously [53–54] (GEO accessions: GSE31320 and GSE62486) and binding sites were identified from the merged dataset with MACS 1.4 (p<10−7) [71]. The positions of T-bet peaks were identified relative to gene transcription start sites annotated in RefSeq. The GWAS catalogue was downloaded from the NHGRI [55] on December 12th 2014. SNPs were checked against dbSNP and 4 SNPs called ‘suspect’ removed. SNPs that had been merged with other IDs were checked against HapMap3 and the ID given in HapMap3 used in downstream analysis. SNPs not in HapMap3 were removed, giving 13,936 autosomal SNPs in the final analysis. Data were analysed using the bioconductor snpMatrix programme (recently updated to snpStats) [72, 73]. SNPs in high LD (r2> 0.8 with a SNP from the GWAS catalogue) were obtained from HapMap3 [74], giving a total of 127,594 SNPs. These were then overlapped with the T-bet binding sites. To identify the number of independent LD blocks were represented by the 926 T-bet bound SNPs, we used the SNPclip module of LDlink to reduce any SNPs in high LD to a single tag SNP, using a R2 threshold of 0.8 and a MAF (Minimum Allele Frequency) threshold of 0.01.
ChIP-seq data for IgG, H3K27ac and total H3 were taken from GSE62486. Sequence reads were trimmed to remove low quality bases and to remove adapters and aligned using Bowtie (default settings) to hg19. Peaks of H3K27ac were identified with MACS (p<10−7) [71]. DHS data were obtained from ENCODE (GEO accession GSM736592) [75,76]. Average binding profiles were calculated across 4 kb windows centred on hit-SNPs using ngsplot [77]. Data were visualized on the UCSC genome browser by calculating tag density in 10bp windows, normalizing to reads per million total reads and subtracting background (input for T-bet and H3 for H3K27ac), as described [54].
Individuals heterozygous for rs1465321 were identified from the Twins UK cohort at the Guy’s and St Thomas’ NHS Foundation Trust (GSTT) Bioresource, where HumanHap610Q Illumina array data is available for all registered participants. The Illlumina calling algorithm [78] was used to assign genotypes from array data. Before imputation, quality controls were applied, with exclusion of all samples with: (1) call rate <98%, (2) heterozygosity across all SNPs ≥2 standard deviations from the sample mean; (3) evidence of non-European ancestry as assessed by PCA comparison with HapMap3 populations; (4) observed pairwise IBD probabilities suggestive of sample identity errors. We also corrected zygosity based on IBD probabilities. Quality controls were also applied to each individual SNP using the following exclusion criteria: (1) Hardy-Weinberg p-value <10−6 (assessed in a set of unrelated samples); (2) MAF <1% (assessed in a set of unrelated samples); (3) SNP call rate <97% (SNPs with MAF ≥5%) or < 99% (for 1% ≤ MAF < 5%). Finally all the alleles were aligned to the forward strand of HapMap2. After completion of both sample and SNP quality controls checks, imputation was performed using the IMPUTE software package (v2) [79] using HapMap2 as a reference panel (HapMap2, rel. 22, combined CEU+YRI+ASN panels). Heterozygous SNPs were selected using PLINK (version 1.0.7) [80] “—recode-rlist” option on the imputed dataset. A final QC stage was applied on all the heterozygous SNPs, excluding all those polymorphisms with an imputation quality score ≤ 0.8.
In accordance with the Department of Health’s Research Governance Framework for Health and Social Care, ethical approval for this study was gained from the South London Research Ethics Committee (Ref:15/LO/0151), and from the Department of Research and Development at GSTT NHS Trust (Ref:RJ115/N122). Approval was also gained from the GSTT National Institute of Health Research (NIHR) Bioresource for recruitment of individuals registered on the Bioresource and heterozygous for rs1465321. All of the subjects in this study gave written consent. Blood was taken from two individuals heterozygous for the desired SNP. CD4+ T cells were purified from whole blood leukocytes using CD4 microbeads (Miltenyi Biotec) and naïve CD4+ T-cells sorted by FACS selection for CD4+ CD45RA+ CD4RO- CD25- CCR7+ cells. Sorted naïve T-cells were activated with anti-CD3/CD28 and polarized under Th1 conditions (IL2, IL12 and anti-IL4) for 7 days [54]. Cells were then crosslinked and ChIP-seq for T-bet performed with a custom-made polyclonal antibody [54]. Libraries were quantified using the KAPA library quantification kit and sequenced (150 bp single-end) with an Illumina NextSeq. Sequence reads were trimmed and aligned to hg19 as before. Peak regions for both donor 1 and 2 were identified separately using MACS 1.4. Broad shallow peaks were filtered, intersecting peaks identified with Bedtools (n = 8185), and then narrowed to the central 400 bp. Potential SNP variants within these intersecting peak regions were extracted from dbSNP version 138 (assembly hg19, n = 490,310). SNP sites for further analysis were determined from the Bowtie aligned bam files as containing >1 reads with both Ref and Alt bases in both ChIP and Input samples from both donors (n = 9058). This list was then compared to the set of heterozygous SNPs identified by the SNP array analysis (n = 2621 high confidence heterozygous SNPs). Reads surrounding these sites were extracted into R using the Bioconductor Rsamtools and GenomicRanges packages. The reads were split by Ref and Alt alignment for visualization using the GenomicAlignments package.
To test whether T-bet exhibited allelic imbalanced binding at rs1465321 and at SNPs in high LD, we used a binomial test. Donor 1 and 2 p-values were combined using the Fisher method. rs1465321 and rs2058622 showed significant allelic imbalance (p<0.01) in the T-bet ChIP samples and allelic balance (p>0.1) in the Input samples. To identify other heterozygous SNPs that exhibited allelic-imbalanced T-bet binding, we used a binomial test to identify heterozygous SNPs at which significantly more reads were reported for one allele compared to the other in both T-bet ChIP samples (Benjamini-Hochberg adjusted p<0.05) but not imbalanced in the Input samples from either donor (unadjusted p>0.4). This produced a list of 19 additional SNPs (S4 Table).
T-bet ChIP-seq data from donors 1 and 2 heterozygous for rs1465321 are available at GEO under accession GSE81881.
Data-sets for wild-type and T-bet deficient CD4+ T cells polarised in Th1 and Th2 conditions were obtained from GEO (GSE38808). Raw reads were aligned to the mm10 build of the murine genome using Subread [81], and subsequently mapped to RefSeq genes using featureCounts [82]. DESeq2 was used to normalise read counts by size factors, and call differentially regulated genes using an empirical Bayes model and the Wald test followed by Benjamini-Hochberg correction for multiple testing [83].
The presence of T-bet motifs was assessed using FIMO [84] using previously compiled matrices for T-bet binding obtained by ChIP-seq [54]. Sequences for T-bet binding sites were obtained from the hg19 reference genome and SNPs were manually altered to the alternative allele.
fGWAS analysis was performed as described in [56] using fGWAS version 0.3.3 with case control setting. Data were prepared for fGWAS using R and the GenomicRanges package to compute overlap between binding sites and SNPs. Publicly available GWAS data were downloaded from the websites of the relevant consortiums for UC, Crohn’s disease (http://www.ibdgenetics.org/downloads.html), coronary artery disease (http://www.cardiogramplusc4d.org/downloads/), and rheumatoid arthritis (http://plaza.umin.ac.jp/~yokada/datasource/software.htm). Psoriasis data are from [85]. T-bet binding sites were identified as described above. ENCODE ChIP-seq, FAIRE-seq and DNaseI hypersensitivity datasets were obtained from the ENCODE website in bed format (http://ftp.ebi.ac.uk/pub/databases/ensembl/encode/integration_data_jan2011). The complete ENCODE datasets combines DNaseI (125 annotations), FAIRE-seq (24 annotations), histone marks (117 annotations) and transcription factor binding site datasets (S2 Table). In addition, we included GATA3 binding sites in Th1 and Th2 cells (GSE31320) [54], FOXP3 binding sites in Tregs [86], NF-κB binding sites in lymphoblastoid cells [47] (GSE19486), and H3K27ac [58] and DHS [57] in Th1 and Th2 cells.
Celiac disease association was based on a case control association study of 12,041 celiac disease cases and 12,228 controls [23]. Gene expression data were taken from [62]. eQTL analysis was performed as described [63]. eQTL p-values were obtained by fitting a linear trend test regression between the expression of each gene and all variants 200 kb upstream and downstream from each probe. Posterior computation was performed as described [63]. The regional association plots for the eQTL and Biomarker datasets were created using LocusZoom [87] (http://csg.sph.umich.edu/locuszoom/).
Colocalisation analysis was performed using the R package COLOC [63] based on single variant summary statistics (log odds ratio, standard error for the log odds ratio for case control and effect size and standard error for effect size for eQTL study, in addition to MAF and physical position for each variant) and with the default settings provided with the R package.
Human CD4+ cells were isolated from buffy coats (UK National Blood Service, used under REC reference number 10/H0804/65 from SE London Research Ethics Committee 2) using RosetteSep human CD4+ T cell enrichment cocktail (STEMCELL Technologies) according to manufacturer’s instructions and polarised towards a Th1 phenotype in supplemented RPMI as described in above. Cells were harvested after a total of seven days of culture. YT cells were cultured in RPMI medium (PAA) supplemented with 50 units/ml penicillin, 50 μg/ml streptomycin (Gibco), 10 mM HEPES buffer solution (Fisher Scientific), 1 mM sodium pyruvate (Gibco), 1 × minimum essential medium-non essential amino acids (Gibco), 2 mM L-glutamine (Gibco) and 10% foetal bovine serum (PAA). All cells were maintained at 37°C in 5% CO2.
Forward and reverse single-stranded oligos (Integrated DNA Technologies, S3 Table) for each allele of each SNP were annealed by incubating at 94°C for 5 mins, 65°C for 10 mins, 25°C for 10 mins and 4°C thereafter in annealing buffer (50 mM Tris pH 8, 7 mM MgCl2 and 1 mM DTT). For pull-down and western blot, 20μl of streptavidin agarose beads (Sigma) were used per sample. For OligoFlow, 50 μl of Sphero streptavidin polystyrene particles (Spherotech #SVP-100-4) were used per sample. Beads were washed twice in PBS and then once in annealing buffer. Beads were then incubated with double-stranded oligonucleotides for 1 hr at 4°C, washed twice in oligo buffer (10 mM Tris pH 8, 100 mM NaCl, 0.1 mM EDTA, 1 mM DTT, 5% glycerol, 1 mg/ml BSA Fraction V, 20 μg/ml dI/dC (Sigma, P4929) and Complete protease inhibitor (Roche) and finally resuspended in 450 μl oligo buffer. Cells (30 million per sample) were washed twice in PBS and lysed in 1 ml hypotonic buffer (20 mM HEPES pH 8, 10 mM KCl, 1 mM MgCl2, 0.1% Triton X-100, 5% glycerol, 1 mM DTT and Complete protease inhibitor) on ice for 5 mins. Lysed cells were pelleted and resuspended in 150 μl hypertonic buffer (20 mM HEPES pH 8, 400 mM NaCl, 1 mM EDTA, 0.1% Triton X-100, 5% glycerol, 1 mM DTT and Complete protease inhibitor). Debris was pelleted, 180 μl of supernatant containing nuclear extract added to the beads and incubated on a rotor for 1 hour at 4°C. For western blotting, samples were then washed three times in oligo buffer and resuspended in SDS loading buffer. For OligoFlow, 0.25 μg of anti-T-bet Alexa647 antibody (clone 4B10, BioLegend) was added and samples incubated for a further 1 hr at 4°C. Data (at least 30,000 events) were acquired on a FACSCanto flow cytometer (BD Biosciences).
Oligonucleotide pull-down samples were heated in SDS loading buffer before transfer to nitrocellulose membrane. Samples were blocked in 5% milk in TBS-T (1 hr, RT) and incubated with 1:1000 anti-T-bet (clone eBio4B10 (eBioscience); 4°C overnight). Blots were washed before addition of anti-mouse-HRP (GE Healthcare) and visualised with Enhanced Chemiluminescent Substrate (PerkinElmer) and exposed to film.
|
10.1371/journal.pcbi.1003721 | Correlated Inter-Domain Motions in Adenylate Kinase | Correlated inter-domain motions in proteins can mediate fundamental biochemical processes such as signal transduction and allostery. Here we characterize at structural level the inter-domain coupling in a multidomain enzyme, Adenylate Kinase (AK), using computational methods that exploit the shape information encoded in residual dipolar couplings (RDCs) measured under steric alignment by nuclear magnetic resonance (NMR). We find experimental evidence for a multi-state equilibrium distribution along the opening/closing pathway of Adenylate Kinase, previously proposed from computational work, in which inter-domain interactions disfavour states where only the AMP binding domain is closed. In summary, we provide a robust experimental technique for study of allosteric regulation in AK and other enzymes.
| Most proteins contain several domains, and inter-domain motions play important roles in their biological functions. Describing the various inter-domain orientations that multi-domain proteins adopt at equilibrium is challenging, but key for understanding the relationship between protein structure and function. When more than two domains are present in a protein, correlated domain motions can be of fundamental importance for biological function. This type of behaviour is typical of molecular machines but is extremely challenging to characterize both from experimental and theoretical viewpoints. In this paper, we present a hybrid experimental/computational approach to address this problem by exploiting the information on molecular shape contained in nuclear magnetic resonance experiments to determine accurate conformation ensembles for the multi-domain enzyme adenylate kinase with help of advanced simulation methods.
| Conformational heterogeneity as a consequence of dynamics is an intrinsic feature of proteins linked to biological function.[1] An important aspect for our understanding of protein dynamics is a molecular characterization of the structural states that are populated and of how correlated conformational changes can mediate biological functions such as allostery and signal transduction.[2], [3] The characterization of correlated conformational changes requires a description of how local structural heterogeneity translates into global conformational changes through collective motions. Recent developments in the analysis of various NMR parameters at atomic resolution, such as spin relaxation rates[4] and residual dipolar couplings (RDCs),[5]–[7] have enabled the description of local structural heterogeneity. However, inferences of collective conformational changes from local correlated motions have relied on force fields or motional models.[7]–[9]
RDCs are NMR parameters that report on both the local and global structural properties of weakly aligned macromolecules and, under the assumption that alignment does not alter the properties of the protein, can be used to study the amplitude of dynamics, especially when combined with simulations.[5], [10] Alignment can be induced by steric and/or electrostatic interactions of the macromolecule with external media. Specifically, for a given conformation, RDCs depend on the geometrical properties of the environment of the nuclei in the molecular frame and on the direction and degree of alignment[11], [12] (Eq. 1, Fig. 1). (1)
In Eq. 1, for two nuclei P and Q, φiPQ is the angle between the inter-nuclear vector and axis i of the molecular frame, Sij is an element of the alignment tensor (S), rPQ is the distance between P and Q, γX is the gyromagnetic ratio of nucleus X, µ0 is the magnetic susceptibility of vacuum and h is Planck's constant. The alignment tensor S is given by the alignment mechanism and, under steric alignment conditions, can be computed accurately and is closely related to the shape of the aligned macromolecule.[12]–[14] RDCs measured in steric alignment combine the local information contained in NMR parameters with the shape information contained in relaxation rates[15]–[17] and small angle X-ray scattering (SAXS).[18] Since inter-domain motions alter the shape of proteins RDCs measured in steric alignment have, as we will show, the potential to act as reporters of inter-domain structural heterogeneity.
RDCs can be used to study the structural heterogeneity of globular and disordered proteins by analysing the effect of (sub-ms) motional averaging on this NMR parameter. The effect of fast (sub-ns) local motions that do not directly alter the magnitude and main directions of alignment can be analysed in the molecular frame defined by the alignment tensor of the average structure.[19], [20] The analysis of motions that change the shape of the protein, that are typically slower than the timescale of alignment (0.5 to 5 ns) [21] needs on the contrary to explicitly consider that the various conformations in fast exchange that contribute to the measured RDC can have different alignment tensors.[20], [22] Here we show that for a molecule undergoing conformational changes in two distal sites the RDCs depend on the degree of correlation of such changes. We exploit this to characterize the degree of correlation of the inter-domain conformational changes occurring in the substrate-free state of E. coli Adenylate Kinase (AKe), an enzyme that undergoes conformational changes involving shape changes.
Our approach is based on the determination of ensembles that collectively agree with RDCs. The generation of the ensemble is divided in two steps (see Methods and Supporting Information). First, an enumeration (ca. 105) of inter-domain orientations is performed using unrestrained simulations. Secondly, the ensemble of minimal size that best fits the RDCs is identified by a genetic algorithm (see Figs. S1 and S2).[23] To maximize the coverage of conformational space in the first step we used PELE, an all atom Monte Carlo simulation algorithm with a move set designed to explore normal modes (see Methods and Supporting Information).[24] The RDCs of each conformer are calculated via Eq. 1 using two independent methods to compute the S from knowledge of the structure.[12], [13] In this case we used only the steric tensor because it is related to shape but it is principle possible to use also the electrostatic tensor, particularly when conformational changes modulate the charge distribution.
AKe is an essential enzyme that catalyses the reversible conversion of ATP and AMP into two ADP molecules. The role of AKe is to maintain the energy balance in the cell, which is essential. AKe is a modular enzyme composed of three sub-domains: a CORE domain responsible for thermal stability[25], [26] and two flexible substrate-binding domains referred to as LID and AMPbd (Fig 2a). Crystallographic structures of AKe have been solved for the open (inactive) and closed (active) states.[27] Substrate-free AKe has been shown to sample a closed-like state using single-molecule Föster resonance energy transfer (smFRET) experiments.[28], [29] We previously reported an analysis of AKe using RDCs in the substrate-free (open) and substrate-bound (closed) states where we fit the RDCs to the corresponding crystallographic structures.[30] We found that the fit of the closed state was better than that of the free state, which we interpreted as evidence for inter-domain dynamics in the substrate-free enzyme.
To identify major conformational states sampled by substrate-free AKe we used backbone NH RDCs measured in steric alignment to select ensembles of conformations from a pool derived from simulations started from the open and closed X-ray structures[27],[31] (see Methods). The ensembles collectively agree with the RDCs (Q≈0.26) even though the structures they contain do not (Q>0.6) when considered individually. An ensemble size of 4 to 64 accounted for the data and equivalent distributions were observed regardless of ensemble size and method used to calculate the RDCs (Figs S3 to S5). It is worth noting that the LID domain of AKe is in equilibrium with a locally unfolded state with a population of ca. 5% at 37°C;[32] under the conditions used in this study (25°C) this state has a low population (<1%). Similarly, cracking of the AMPbd along the closing mechanism may play a role.[26], [33] However, cracking occurs at the transition state, and therefore has a very low population. The states obtained for AKe are shown in Fig. 2b. An analysis of the results indicated that substrate free AKe samples three major states corresponding to i) a closed-like state (θLID ∼105°, θAMPbd ∼60°, population ∼0.5±0.1), where the LID is closed and the AMPbd slightly opened ii) the open state (θLID ∼146°, θAMPbd ∼74°, population ∼0.25±0.1), where both nucleotide binding domains are open, and iii) conformations in which the degree of domain opening is intermediate between that of the fully closed and fully open. The estimate of the fraction closed AKe from our analysis is 0.5, which is in agreement with numbers from single molecule FRET studies where fractions of 0.6[29] and 0.3[28] has been enumerated.
We performed additional control simulations to assess the robustness of the distributions obtained. Scrambling the RDCs lead to a list of restraints that could not be fit to any distribution (Fig. S6), showing that the distribution obtained is encoded in the data and not biased by the population of the conformers in the pool. The distribution shown in Fig. 2a was found to be robust to various sources of error, including errors in the experimental RDCs (Fig. S7) and in the prediction of the alignment tensor (Fig. S8). We also performed computational experiments to ensure that a single set of RDCs suffices to distinguish between ensembles with a distinct number of states and identify correlated structural changes (see Figs. S9 to S13).
An analysis of the conformational ensemble allowed us to assess the presence and degree of inter-domain correlation, an important property of AKe. As shown in Fig. 2b, the ATP and AMP domain movements are correlated, disfavouring conformations where the LID is open and the AMPbd is closed. Our findings agree with free energy calculations[34], molecular dynamics simulations[35] as well as with normal mode analysis.[33], [36] Although it is impossible to derive the closure pathway from equilibrium data, the conformational states observed favour a step-wise mechanism in which LID closure takes place before AMPbd closure. Such a mechanism would be beneficial because the initial closure of the LID, together with the substantially higher affinity of this domain for its substrate (ATP) as compared to AMP (50 vs 1700 µM),[30] would reduce the probability of non-productive binding of AMP to the LID.[37]
Single-molecule FRET[28], [29] and NMR[28] have shown that the open and closed states of AKe are in equilibrium and that nucleotide binding shifts it towards the closed state.[30] Of particular interest are smFRET experiments, as these can resolve conformational states and provide a qualitative validation of the ensemble. Two different reaction coordinates, corresponding to distances between residues in the LID and the AMPbd (Aquifex AK)[28] and between residues in the LID and the CORE domain (AKe),[29] have been studied. In the former the authors resolved two states in which open conformations were populated to ca. 70%, whereas in the latter the authors found that the open conformation was disfavoured, with a population of ca. 40%. These results can be rationalized based on the ensemble. Following the approach of Beckstein et al.,[38] where distances corresponding to the open and closed states are mapped in domain angle space, we estimated open populations along the LID-AMPbd and LID-CORE smFRET reaction coordinates >50% and ∼30%, respectively (for LID-AMPbd large uncertainties are found due to difficulties in assigning open and closed states as quantified by smFRET). Overall, the ensemble is able to qualitatively reconcile two apparently contradictory smFRET studies, which also suggests that steric alignment does not significantly alter the structural heterogeneity of AKe.
Here we have used the dependence of RDCs on global molecular shape via S to determine ensembles reporting on the amplitude and degree of correlation of inter-domain motions. Although RDCs are uniquely suited for this, this is a challenging endeavour due to their intrinsic degeneracies.[39] It is therefore relevant to discuss the factors that allowed the approach used in this work to alleviate them. The first factor is that our approach accounts for how shape changes alter the value of the RDC by computing S for each inter-domain orientation i.e. two inter-domain orientations that would have degenerate RDCs when the tensor is assumed to be constant can be differentiated if they have sufficiently large differences in shape (Fig. S14).[39] A second factor is the presence of structural constraints due to the covalent linkage, where steric clashes between domains restrict the available conformational space;[40], [41] AKe is a well-structured proteins where this effect is particularly important but for multi-domain proteins with flexible linker sequences RDCs measured in multiple alignment media will be required. Alternatively, RDCs could be complemented with structural restraints derived from smFRET or SAXS data.
Correlated motions in proteins are thought to mediate biochemical functionality.[42] Recently, we have identified weak long-range correlated motions in a surface patch of ubiquitin involved in molecular recognition.[7] Here, we move a step forward and find correlated domain movements in a representative multi-domain enzyme. The observed inter-domain correlation suggests functional roles in allowing ligand access, in adopting the inter-domain orientation necessary for catalysis as well as in binding allostery. Our results, therefore, reinforce that correlated inter-domain motions in proteins can mediate important biochemical processes.
NH RDCs used in this work for free AKe were measured in stretched polyacrylamide gels.[30] Two independent methods, PALES [12] and ALMOND[13], [13]were used to calculate the alignment tensor using Eq. 1. They gave equivalent distributions as shown in Figs. S3–S5. In Figs. 2 we provide the distributions that lead the best agreement between calculated and experimental RDCs [43] (Tab. S1). For AKe it was obtained using ALMOND. The agreement between calculated and experimental (or reference) RDCs was assessed by the quality factor Q [43] defined in Eq. 2. (2)
The calculated RDCs were scaled, after averaging across the ensemble, to minimize Q to account for the difficulty of predicting the absolute concentration of alignment medium. We used RDCs corresponding to structured regions in the protein (see Fig. S15). For AKe, we found that RDCs in loop regions were more difficult to fit and therefore we decided to exclude them. However, we note that the ensemble derived including RDCs in the loop regions remains unaffected. We also evaluated the impact of the alignment media by monitoring changes in chemical shifts induced by the presence of the stretched polyacrylamide gel. As shown in the Supporting Figure S16, no significant chemical shift perturbations were observed when apo AKe was immersed into the anisotropic phase. As a reference the chemical shift perturbations induced by binding of the inhibitor Ap5A to AKe are also displayed. A thorough analysis of the robustness of the procedure to various source of error is provided as Supporting Information. To select the optimum ensemble size we monitored the agreement with RDCs used to guide the genetic algorithm and to RDCs left out of this process or free RDCs. Ten sets of randomly removed RDCs (20%) or free RDCs sets were used and 20 independent calculations were run for each set of randomly removed RDCs. The ensemble was qualitatively validated against two independent smFRET experiments (see results and discussion).
An extensive set of inter-domain conformations (ca. 105) was generated for AKe. The X-ray structures 1ake [31] and 4ake [27] were used as seeds. Trajectories were generated where either the AMP or ATP binding domain open (or closed). From each state along the trajectory a second trajectory was generated where the other domain was forced to open (or to close). These simulations were performed with the molecular simulation package CHARMM c35.[44] A time step of 1.0 fs was used. Simulations were performed at 300K. A shape-term (biasing) potential on the backbone atoms was used. The CHARMM commands "CONStraint HARmonic BEStfit COMParison" and the IC table tool (CONStraint IC BOND ANGL IMPR DIHE) were used to sample inter-domain orientations between open and closed conformations. The initial force was set to 0.1 and exponentially increased by a factor of 1.05 during 200 cycles of 100 steps each. The degree of opening ranged between ∼30–90° and ∼90–160° for the AMP- and ATP-binding domains (see angle definitions for AKe). The angles sampled covered the range observed by known X-ray structures of AKe and related proteins. [38] An advanced sampling strategy based on the PELE [24] method was used (see Protein Energy Landscape Exploration, see Supporting Information). This allowed identifying high-energy conformations which are unlikely to be of sufficiently low energy to be present in the crystalline state or sampled by conventional MD.
An in-house genetic algorithm GA was developed to efficiently search within the pool of structures used to determine the experimental inter-domain orientation distributions of T4L and AKe. Initially 1000 ensembles were generated with structures randomly selected from a pool of ca. 105 conformations (see reference pool of conformations above). Ensembles of size N = 1, 2, 4, 8, 16, 32 and 64 were used. The 103 ensembles of size N were submitted to evolution through 1000 steps. At each step two new sets of 103 ensembles were generated by mutation and crossing operations, leaving 3×103 ensembles. At each step the best 103 ensembles, based on the value of Q (see below), were retained. After 1000 iterations the best ensemble was saved and the complete procedure was repeated 200 times. Because too high mutation and recombination rates may lead to loss of good solutions and to premature convergence of the GA, these parameters evolved during the calculations. Initially, the mutation and crossing rates were set to 100% and 2%, respectively. At each step new mutation and crossing rates were determined by dividing/multiplying them by a factor of 1.001, respectively. In Figs. S1 and S2 a scheme illustrating the GA and the convergence of the method are shown. The error in the population of the states observed were determined from the influence of RDC experimental uncertainty by performing 200 calculations with random Gaussian error added to each experimental RDC. We used a standard deviation of 1.0 Hz for AKe, three times the estimated experimental error.
The definition used by Beckstein et al. [38] was adopted in this work. Briefly, the angle formed between the AMP-binding domain (AMPbd) and the core domain (CORE) was determined as the angle formed by two vectors: Vector 1 connects the centers of mass (Cα atoms) of CORE residues 90–100 with residues 115–125 (“hinge region” between the CORE and LID domains). Vector 2 connects the centers of mass of CORE residues 90–100 and AMPbd residues 35–55. The angle formed between the ATP-binding domain (LID) and the CORE was defined equivalently using the angle formed between two vectors: Vector 1 connects the centers of mass of residues 115–125 (“hinge region” CORE–LID domains) with CORE residues 179–185. Vector 2 connects the centers of mass for residues 115–125 (“hinge region” CORE–LID) with LID residues 125–153.
|
10.1371/journal.ppat.1006796 | Influenza A virus hemagglutinin glycosylation compensates for antibody escape fitness costs | Rapid antigenic evolution enables the persistence of seasonal influenza A and B viruses in human populations despite widespread herd immunity. Understanding viral mechanisms that enable antigenic evolution is critical for designing durable vaccines and therapeutics. Here, we utilize the primerID method of error-correcting viral population sequencing to reveal an unexpected role for hemagglutinin (HA) glycosylation in compensating for fitness defects resulting from escape from anti-HA neutralizing antibodies. Antibody-free propagation following antigenic escape rapidly selected viruses with mutations that modulated receptor binding avidity through the addition of N-linked glycans to the HA globular domain. These findings expand our understanding of the viral mechanisms that maintain fitness during antigenic evolution to include glycan addition, and highlight the immense power of high-definition virus population sequencing to reveal novel viral adaptive mechanisms.
| Seasonal influenza A viruses (IAV) cause tens of thousands of deaths and tens of billions of dollars in economic costs every year in the U.S. alone, despite widespread pre-exposure and vaccination. IAV persists within the human population by evading herd immunity through the continual accumulation of immune escape substitutions in a process known as antigenic drift. Unfortunately, the specific mechanisms that facilitate IAV antigenic drift remain unclear, making it difficult to design more efficacious, escape-proof vaccines. Here, we used a high-definition population sequencing approach to understand how IAV tolerates the accumulation of immune escape substitutions during antigenic drift without losing fitness. We found that the virus rapidly acquired N-linked glycan structures on the hemagglutinin (HA) protein following escape from a panel of neutralizing antibodies. Glycosylation of HA alleviated the deleterious effects of immune escape substitutions on interactions with host cell receptors. These results expand our understanding of the mechanisms that facilitate IAV immune evasion, and highlight the immense power of high-definition virus population sequencing to reveal novel viral adaptive mechanisms.
| Influenza A virus (IAV) persists in human populations by continuously evolving to escape herd immunity. Protective immunity is predominantly mediated by neutralizing antibodies (Abs) specific for the viral surface glycoprotein hemagglutinin (HA). HA mediates both target cell attachment, by binding terminal sialic acids (SA) on cellular membrane components, and fusion between viral and cellular membranes following virion internalization. Neutralizing Abs mainly target the five highly variable immunodominant antigenic sites on the globular head domain of HA, blocking either HA-mediated attachment or fusion [1–4]. Viruses can escape neutralization by amino substitutions in HA that reduce antibody affinity and/or function. The fitness costs imposed by Ab escape, and the mechanisms by which the virus compensates remain poorly understood, yet play a central role in governing HA antigenic evolution [5,6].
Influenza virus escapes from neutralizing antibody via a number of defined mechanisms involving HA mutations. The simplest is an amino acid substitution in a cognate epitope that diminishes antibody affinity [2]. Less commonly, distant substitutions can affect antibody binding via allosteric effects on antibody access to its epitope [7–9]. Amino acid substitutions in the sialic acid receptor site, or other regions of the globular domain that increase affinity for cellular SA receptors, enable Ab-escape by shifting the binding equilibrium towards virus binding to cells versus antibody [10–13]. The most drastic alterations in overall antigenicity typically result from amino acid substitutions that create a N-linked glycosylation site in the globular domain, as the attached glycan can sterically block the binding of Abs to multiple sites [5,14–19].
The ability of viruses to exploit these escape routes is limited by their costs to viral fitness. The sialic acid receptor binding site is nestled among the immunodominant antigenic sites, and escape substitutions often alter viral fitness by changing HA receptor binding avidity. Amino acid substitutions that change the net charge of the globular domain typically alter cell binding avidity [13,20]. We previously demonstrated that selection of substitutions that add N-linked glycosylation sites within the globular domain to escape antibody neutralization often reduces receptor avidity [5]. Receptor avidity and HA antigenicity are thus intimately linked, and must be continuously balanced to maintain fitness during antigenic evolution [13,21].
Here, we use the primerID method for error-correcting virus population sequencing to reveal a surprising new role for N-linked glycosylation in facilitating IAV immune escape. Beyond its canonical role in blocking Ab binding, we show that glycosylation can also compensate for fitness costs imposed by escape substitutions elsewhere on HA, thus increasing the viability and subsequent emergence potential of Ab-escape variants.
We previously modeled IAV antigenic evolution by sequentially selecting for virus escape from different mouse anti-HA monoclonal Abs (mAbs) under over-neutralizing conditions using the allantois-on-shell (AOS) system [21]. Twelve rounds of mAb escape by PR8 yielded a virus (SV12) carrying twelve amino acid substitutions in HA that collectively mediated near total escape from polyclonal serum raised against PR8 in mice, guinea pigs or chickens. When propagated in embryonated chicken eggs, PR8 and SV12 reached similar titers in the absence of Ab pressure, suggesting that the selected constellation of escape mutations imposed minimal fitness costs on the virus.
An alternative possibility that we did not examine in detail is that the accumulation of antigenic escape substitutions imposed significant costs, but strong selection for compensatory mutations or reversions in SV12 rapidly restored fitness during the initial rounds of SV12 expansion in eggs; the original SV12 isolate had been expanded 2–3 times in the absence of antibody in embryonated chicken eggs after the 12th sequential selection in the AOS system [21]. To test this possibility, we generated a recombinant PR8 clone carrying the consensus SV12 HA gene segment and compared the ability of WT PR8 and the consensus SV12 clone to replicate in eggs (Fig 1). This revealed that antigenic escape substitutions present within SV12 HA imposed significant fitness costs (greater than 10-fold decreased titer) when the virus population had limited time to recover fitness via mutational compensation.
Compensatory mutations that increased the fitness of SV12 may not have been identified with conventional Sanger sequencing if multiple variants were selected that collectively were a significant proportion of the population, but that individually remained below the wild-type SV12 consensus frequency at each codon. To identify compensatory mutations or reversions in HA that potentially emerged following sequential antigenic escape, we used the primerID method for error-correcting virus population sequencing to identify minor sequence variants that emerged during the limited expansion passaging of SV12 [22,23].
PrimerID sequencing is based on tagging individual viral cDNAs with random unique sequence barcodes during reverse transcription [22,23]. These barcodes are maintained through downstream PCR and sequencing steps and are used to assemble all sequencing reads derived from a single cDNA sequence. This permits consensus-based reconstruction of the original cDNA sequence, enumeration of actual viral genomes sampled, and analysis of linkage between specific mutations within reads. Importantly, primerID lowers the background sequencing error rate of the Illumina platform 10-100-fold (to 10−3–10−4 non-consensus events/nt), by correcting the high-frequency PCR and base-calling errors that occur during standard next-generation sequencing. PrimerID can also correct for variant frequency skewing due to PCR resampling imbalance [22,23].
To precisely quantify the advantages conferred by primerID sequencing over shotgun sequencing approaches in the context of IAV population sequencing, we directly compared results obtained with the primerID method to those obtained by shotgun sequencing, as analyzed through the ViVan pipeline [24]. Variant calls with a moderately high frequency (>10−3) show consistent results whether an amplicon/primerID or a shotgun sequencing approach were used (S1 Fig); however, evaluating variant frequencies in the range of 10−4 to 10−3 shows that primerID gives a higher signal to background noise ratio compared to shotgun sequencing (S2 Fig). The noise reduction of primerID consensus reads compared to unmerged amplicon reads ranges from 9.2-fold to 18.9-fold (S1 Table). Similarly, the background noise level, estimated by the median combined minor allele frequency (MAF) in primerID is about 3-10x10-5, which is 4.5 to 12-fold lower than background in shotgun sequencing (S2 Table). Altogether, primerID allows for more conservative variant calling, and likely more accurate estimations of variant frequencies, than algorithm-based error correction approaches [24,25].
To identify emergent HA variants within an approximately 143 a.a. region (115 to 258) that encompassed the receptor binding pocket and most of the SV12 substitutions in the original expanded SV12 virus stock, we performed paired-end sequencing of primerID libraries on the Illumina MiSeq platform and analyzed data using a custom software pipeline. Sequencing revealed amino acid substitutions at 7 sites that emerged at frequencies of >0.1% during 2–3 passages of SV12. These substitutions were (including their frequencies) K123N (1.7%), K123E (1.1%), N133T (25%), N133S (8%), K144N (9%), K144E (0.5%), N145D (0.3%), K174E (0.3%), K222T (0.17%), and G225D (3.7%) (all H3 HA numbering) (Fig 2A). Linkage analysis revealed that substitutions at positions K123, N133, K144, K145, and G225 represented distinct variants (S2 Table), and thus collectively made up approximately 50% of the population of the SV12 stock.
Each substitution was observed at sites with potential functional roles in SV12 (Fig 2B). Substitutions at two SV12-defining sites were observed: G225D was a reversion to the PR8 wild-type amino acid identity, while the most abundant substitution at N145 was D (0.3%), rather than the PR8 wild-type S (0.01%). The other substitutions (K123N/E, N133T/S, K144N/E, K174E, and K222T) were at new sites, but were adjacent to SV12-defining sites or rimmed the receptor binding site (Fig 2B). Thus, a combination of forward mutations and reversions rapidly emerged on the SV12 background following antigenic escape.
Three of the amino acid substitutions that emerged in the original expanded SV12 virus stock (K123N, N133T, and K144N) created new potential N-linked glycan sites (PNGS) within the HA sequence (at positions 123, 131 and 144, respectively). To determine whether these substitutions indeed affected the glycosylation status of virion-associated HA[26], we introduced each substitution individually into the recombinant consensus SV12 HA background (rSV12-HA) and rescued viruses carrying the 7 other gene segments from WT PR8 by reverse genetics. We generated purified, detergent-disrupted virus preparations of rSV12-HA, rSV12(K123N)-HA, rSV12(N133T)-HA, and rSV12(K144N)-HA and digested with either Endo H (cleaves mannose-rich N-linked glycans, but not complex N-linked glycans) or PNGase F (cleaves all N-linked glycans). We compared the migration of digested and undigested HA on a denaturing gel by immunoblotting (Fig 3A). All three compensatory PNGS clearly slowed the migration of undigested, but not PNGase F-digested, HA on the gel, demonstrating that each mutation results in the addition of a glycan chain to virion-associated HA.
We next tested whether compensatory glycosylation site additions increased the replicative capacity of rSV12-HA in multi-step growth assays in eggs (Fig 3B). The impact of K144N on titer was marginal, but K123N and N133T both resulted in 10–100 fold higher viral titers compared with parental rSV12-HA virus by 36 hpi, demonstrating a clear role for each of these glycosylation site additions in restoring fitness in eggs following SV12 antigenic escape.
A clue to how these glycan additions restore SV12 fitness came from their close proximity to the HA receptor binding pocket (Fig 2B). We hypothesized that the glycans modulate HA receptor binding avidity, correcting for detrimental effects of escape substitutions. To test this, we compared the association (Kon) and dissociation (Koff) rates of purified WT PR8, rSV12-HA, rSV12(K123N)-HA, rSV12(N133T)-HA, and rSV12(K144N)-HA virion preps binding to the model sialic acid receptor 3-SLN using bio-layer interferometry (BLI) (Fig 4, Table 1). rSV12-HA exhibited a 3-fold increase in binding avidity, compared with WT HA. The addition of the K123N, and N133T substitutions to SV12 HA altered binding constants toward WT levels, consistent with our hypothesis. The K144N substitution reduced binding avidity to a much greater degree than K123N or N133T, resulting a 2-fold decrease in avidity relative to WT HA. This effect may have crossed a line beyond the beneficial effects of K123N and N133T, resulting in a detrimental loss of binding avidity (compared with WT) that may explain the comparatively slower growth kinetics of the K144N variant (Fig 3B).
To determine the reproducibility (and potentially the predictability) of the compensatory mechanisms that we observed, we repeated the passage experiment using recombinant, reverse genetics-derived versions of either WT PR8 or SV12-HA (rSV12-HA) virus as the parental virus. We generated each virus in 293T cells from transfected plasmids and then passaged each virus in triplicate in chicken eggs, using a dose of 104 EID50 (based on allantois-on-shell infectivity) to initiate each passage. We harvested each passage at 16 hours post-infection (h.p.i.) to minimize defective interfering particle formation. After the third passage, we sequenced the entire HA coding region of all six passaged populations, along with the parental seed stocks, using an optimized primerID protocol, (see Methods). To overcome the sequence read length limitations of the Illumina Miseq platform (<550nt read length limit), we used four amplicons to cover the entire HA coding region.
Both SV12 and PR8 (WT) parental populations harbored numerous missense mutations at frequencies above background in HA (Fig 5A). Five such mutations resulted in amino acid substitutions that were present in the WT PR8 parental population at frequencies over 0.1%, with a K123T substitution dominant at 2.5% (Fig 5B). Four of the WT variants likely emerged due to stochastic genetic drift resulting from bottlenecking during the rescue process, as indicated by minimal change in frequency by passage 3 [27]. One variant enriched in the WT parental population, K174E, increased slightly in frequency in all three passage lines, and may represent a tissue culture adaptive variant as it was also present in the parental rSV12-HA population. No other variants exhibited a consistent pattern of emergence during passaging of the WT populations.
By contrast, the parental rSV12-HA population carried nearly three times as many missense mutations (14 vs. 5) at frequencies >0.1% (Fig 5C), primarily located in the globular domain of HA. Only K174E was also present in the WT parental population. Five of the 14 variants increased or were maintained at high frequency during passaging in all three SV12 populations: N133T, T136S, K144E, N145D, and G225D (Fig 5D). Of these, N133T exhibited the most dramatic increase in frequency and approached fixation in all three passage lines. Strikingly, four of the five variants carried substitutions at residues that had elevated variant frequencies in our first experiment: N133, K144, N145, and G225, indicating that compensation for antigenic escape can follow predictable genetic pathways under highly controlled conditions.
Emergence of an amino acid variant within a population may result from positive selection driven by specific beneficial effects of the variant on fitness. Alternatively, the variant may be neutral or deleterious but hitchhike with another variant under positive selection to attain higher frequency. The potential for intragenic hitchhiking is high with IAV due to extreme rarity of intragenic recombination during infection [28]. PrimerID sequencing allows the direct quantification of linkage between amino acid positions present in individual viral cDNAs, thereby defining the relative contributions of direct selection vs. hitchhiking in amino acid variant emergence.
We assessed the extent of genetic linkage between the five high frequency variants that emerged in all three rSV12-HA passage populations (N133T, T136S, K144E, N145D, and G225D) (Fig 5D). Examination of linkage at passage 3 revealed that 4 of the variants, N133T, K144E, N145D, and G225D, were mutually exclusive within single HA variants (with just a handful of exceptions) (Fig 6A, S3 and S4 Tables). In contrast, T136S was only observed as paired with N133T. These results were consistent with the nearly mutually exclusive relationships determined for the substitutions observed in the first sequencing experiment carried out on the original SV12 stock (Fig 2, S2 Table).
These findings suggest that each emergent, mutually-exclusive substitution was selected individually for its ability to compensate for the fitness defects of SV12, and that their combination in the same variant may substantially reduce fitness. The data, however, are also consistent with the predicted infrequent recombination within IAV genome segments and a strong role for clonal interference in governing the emergence of beneficial alleles during adaptation [29,30].
To test whether combining compensatory HA glycan additions decreases viral fitness, we generated a N133T, K144N double mutant (rSV12(N133T, K144N)) using reverse genetics. After confirming the double mutant did incorporate an additional glycan compared to the single mutants by SDS-PAGE (S3 Fig), we examined its growth in eggs. The double mutant achieved significantly higher titers than rSV12-HA, but lower than rSV12(N133T) (Fig 6B). This result can be explained by the binding activity of the double mutant, which is comparable to rSV12(N133T) based on BLI (Table 1). Together, these data suggest that clonal interference rather than reduced fitness may be the primary explanation for the mutual exclusivity of compensatory substitutions that we observe.
IAV has persisted in humans despite widespread vaccination and infection, while vaccination has driven other RNA viruses with similarly high mutation rates, such as measles virus and poliovirus, to near-extinction. Understanding the specific features of IAV that enable effective immune escape is critical for designing more effective vaccines and therapeutics.
Our findings reveal a novel role for glycosylation in viral immune evasion, independent of its well-known steric effects on epitope shielding [14–19]. Previous studies clearly showed that glycan addition generally impairs HA function while dramatically increasing fitness in the presence of neutralizing antibodies [5]. Our results demonstrate that glycan addition can also play the opposite role during antigenic drift: increasing the fitness of antigenic escape variants by restoring optimal HA receptor interactions, even in the absence of antibody selective pressure. This raises the intriguing possibility that both mechanisms contribute to the steady evolutionary accretion of N-linked glycans within the globular domain observed during human circulation of H1 and H3 viruses [17].
Importantly, the fact that we selected for glycan addition in the absence of Ab selection does not diminish the profound effects that these glycans have on the antigenicity of HA. In particular, the predominant variant selected, N133S/T, also emerged during the circulation of seasonal H1N1 in humans, and has been demonstrated to have enormous effects on antigenicity and sensitivity to neutralization [17]. These results highlight the self-perpetuating nature of HA antigenic evolution. Hensley et al. found that infection of naïve hosts with antigenic escape variants selected for compensatory mutations that were themselves antigenically significant [13]. Similarly, our results demonstrate that antigenic escape can lead to the emergence of compensatory mutations that push HA even further in antigenic space. These factors make it difficult to infer the nature of selective pressures resulting in given HA sequence changes in human IAV isolates.
Defining the specific properties of the HA molecule that facilitate mutational tolerance is an area of intense study. Studies in vitro found that the HA globular head was highly tolerant of both random substitutions and 5-amino acid block insertions [3,27,31,32]. These results suggested that the HA globular head may exhibit a high degree of mutational robustness, such that the median fitness effects of substitutions are relatively small [32–34]. Alternatively, mutations may generally be costly, but are compensated by epistatic mutations that restore protein function and fitness[6,35]. Our data, along with other studies that demonstrate fitness costs associated with antigenic variation and/or rapid emergence of compensatory mutations following HA antigenic escape, suggest that compensatory epistatic interactions play a critical role during antigenic drift [6,13,36,37].
If rapid compensation of antigenic escape mutations is important for maintaining fitness during antigenic drift, then defining compensatory mechanisms is essential. Many antigenicity-determining residues scattered in the HA globular domain also strongly influence avidity and specificity for sialic acid receptors, and thus Ab escape substitutions at these residues can negatively affect receptor interactions[13]. As a result, many of the compensatory mutations that have been described in association with antigenic escape affect receptor interactions, and likely serve to restore optimal receptor binding properties. The need to accumulate compensatory mutations in order to combine antigenic novelty with fitness and transmissibility has been proposed as an explanation for the discordance between the high in vitro mutation rates and rapid Ab-based selection of influenza viruses in vivo and the relatively slow rates of population-level antigenic evolution seen with influenza viruses[13,38]. Our results indicate that glycosylation can play an important role in this process.
The availability of multiple mutational pathways to restore fitness increases the odds that a fit variant will emerge under selection [39]. Our discovery that glycan addition beneficially tunes receptor interactions adds to previously described mechanisms, including charge transitions in the globular HA domain and alterations in NA activity [13,21,36,40]. While difficult to test experimentally, it is intriguing to consider that IAV may have evolved to maximize the number of compensatory pathways available during antigenic drift.
Our unexpected discovery of a new role for glycosylation in compensating for IAV immune evasion highlights the power of virus population sequencing for uncovering novel mechanisms of viral adaptation through high-definition forward genetics screening. The ability to accurately measure changes in the viral population through methods such as primerID can provide a comprehensive, unbiased profile of the genetic pathways to increased fitness under defined selection conditions.
The primerID approach offers several advantages over conventional deep sequencing protocols. First, the ability to generate a multi-read consensus of the original cDNA sequence reduces the background nucleotide error rate from ~1% to ~0.01%, greatly increasing the sensitivity of minor variant detection. In theory, this background represents the bona fide mutation frequency within the population, combined with the mutation rate of the reverse transcriptase used to generate the cDNA pool (~0.01%)[23]. Second, assembly of amplicon reads into ancestral cDNA sequences eliminates the distortion of measured variant frequencies by the stochastic PCR re-sampling that occurs during conventional amplicon sequencing. Third, the number of unique primerID barcodes used to generate consensus sequences represents the actual depth of the viral population sequenced. Fourth, the ability to phase substitutions into authentic haplotypes enables quantitation of linkage between substitutions within a given genome segment. Finally, primerID requires far less virus input than other methods of highly accurate population sequencing such as CirSeq, facilitating a wider range of experimental applications, including the analysis of samples collected from infected animals or human subjects [41,42].
Altogether, we reveal a surprising new role for glycosylation in facilitating IAV immune evasion by compensating for the fitness costs of antigenic escape mutations. These results broaden our understanding of the critical function of glycosylation during antigenic drift and shed light on the unique mutational plasticity of influenza HA. Further, these studies highlight the remarkable potential of highly accurate virus population sequencing and high-definition forward genetics for exploring viral evolution.
We obtained chicken eggs from a commercial vendor.
The original SV12 stock was generated as previously described [21], and passaged 2–3 times in eggs in the absence of antibodies prior to RNA extraction and sequencing. We generated the A/Puerto Rico/8/1934 (PR8) strain using the eight-plasmid rescue system (plasmids generously provided by Dr. Adolfo Garcia-Sastre; Icahn School of Medicine at Mt. Sinai, New York). We generated seed virus by cotransfecting 293T cells (obtained from ATCC) with the eight gene segment-encoding plasmids. Seed virus stocks were expanded once in 10-day-old embryonated, specific pathogen-free (SPF) eggs. Allantoic fluid was collected 48 hours post infection, and clarified by centrifugation. A reverse genetic construct expressing the consensus HA gene segment sequence from SV12 was generated by RT-PCR amplification of the HA-encoding segment from SV12 virions, and cloning it into the pDZ vector. Virus mutants were generated by site-directed PCR mutagenesis of the relevant reverse genetics constructs. All infectious virus titers were determined by end-point dilution on MDCK cells (obtained from ATCC) in Gibco minimal essential medium (MEM) supplemented with 1 μg/mL trypsin treated with l-(tosylamido-2-phenyl) ethyl chloromethyl ketone (TPCK-treated trypsin), 1mM HEPES buffer (Corning), and gentamycin. TCID50 titers were determined using the Reed–Muench method.
To compare peak titers of molecular clone-derived mutants, 500 TCID50 of each virus were inoculated into 10 day old embryonated eggs in triplicate. Allantoic fluid was collected 48 hours post infection, and titers determined based on end-point dilution using MDCK cells as described above.
MCDK cells were infected with rSV12-HA and each molecular clone-derived mutant at an MOI of 5 TCID50/cell. After 1 hour, virus supernatant was removed and replaced with Gibco MEM supplemented with 7.5% fetal bovine serum, after washing cells with PBS containing CaCl2 and MgCl2. Six hours post infection, cells were washed twice with PBS, and then exposed to lysis buffer containing 1% sodium dodecyl sulfate, 50mM Tris-HCl pH 7.5, 10mM dithiothreitol (DTT), 15U/mL DNase1, and mini-complete protease inhibitor. Lysate was then digested either with EndoH or PNGase F or left untreated. Control or digested lysates were then immunoblotted using the mouse HA2 chain specific mAb RA5-22[43].
Recombinant PR8 and SEQ12HA viruses were serially passaged three times in 10-day old embryonated chicken eggs (Charles River) at a titer of 104 AOS ID/egg at 35°C. At 16 hours post infection, infected allantoic fluid was harvested, clarified and titrated using the AOS method for the subsequent passage [44].
In experiments using recombinant viruses, hemagglutinin was sequenced using four overlapping amplicons spanning the entire coding sequence. Viral RNA was extracted from allantoic fluid using the QiaAmp Viral RNA Mini Kit. Viral RNA was reverse transcribed with Accuscript Hi-Fi Reverse Transcriptase (Agilent) with the following primers bearing random 12-mer barcodes (All Ns hand-mixed to achieve ~25:25:25:25 ratio): PR8_SV12 Amp 1F 5’-ACACTCTTTCCCTACACGACGCTCTTCCGATCTNNNNNNNNNNNNCAGGAAAATAAAAACAACCAAA ATG-3’, PR8_SV12 Amp 2F 5’-ACACTCTTTCCCTACACGACGCTCTTCCGATCTNNNNNNNNNNNNCACCAAAGAAAGCTCATGGCCC-3’, PR8_SV12 Amp 3F 5’-ACACTCTTTCCCTACACGACGCTCTTCCGATCTNNNNNNNNNNNNCAACACGAAGTGTCAAACACCC -3’, PR8_SV12 Amp 4F 5’-ACACTCTTTCCCTACACGACGCTCTTCCGATCTNNNNNNNNNNNNCAGGATTTCTGGACATTTGGAC -3’. The reverse transcription reactions were purified using the PureLink PCR Purification kit (Invitrogen) using the high-molecular-weight cutoff buffer. First-round PCR was performed using Platinum Taq DNA Polymerase High Fidelity (Invitrogen) with an estimated 105 copies of cDNA templates, a universal forward primer (5’-AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACAC-3’) and amplicon-specific reverse primers: PR8_SV12 Amp 1R 5’-GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTNNNNATGGGAGCATGCTGCCGTTA-3’, PR8_SV12 Amp 2R 5’-GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTNNNNGGGAGACTGCTGTTTATAGC-3’, PR8_SV12 Amp 3R 5’-GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTNNNNCAGAGTCCTTTCATTTTCCA-3’, PR8_SV12 Amp 4R 5’-GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTNNNNATATCTCTGAAATTCTAATC-3’.
Sample indices were finally added using KAPA HiFi HotStart PCR Kit, 2.5μL of the first round PCR products and indexed primers (5’- CAAGCAGAAGACGGCATACGAGAT(8-mer sample ID)GTGACTGGAGTTCAGACGTGTG-3’). Amplicons were gel purified using the Qiagen Gel Extraction Kit and quantified on a Nanodrop 1000. Amplicons were pooled into a single library, quantified with the KAPA Library Quantification Kit, and sequenced using the MiSeq Reagent Kit V3 with 325x325 paired-end cycles. The raw sequencing data is available at the NCBI Short Read Archive under accession numbers SRR5428773-SRR5428801.
The primerID library preparation methods used for the original SV12 stock (Fig 2) differed from the optimized pipeline described above in several substantive ways. A single amplicon spanning HA amino acids 115 to 258 was generated using an RT primer with a random 10-mer barcode and the Superscript III RT enzyme (Thermo Fisher). Paired-end reads of the amplicon did not overlap, leaving a gap involving codons 184–186. RT primer (HAF381): ACACTCTTTCCCTACACGACGCTCTTCCGA TCTNNNNNNNNNNCAGCTAAGAGAGCAATTGAGCTCAGTGTCATC. Reverse PCR primer HAr813: GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTNNNN GATGCCGGACCCAAAGCCTCTACTCAGTGC. Subsequent library preparation was as outlined above.
For the sequence analysis of passaging experiments using the recombinant PR8 and SV12 viruses, an optimized analysis pipeline was employed. In this pipeline, the first 4 bases of R2 reads were trimmed using fastx_trimmer from the FASTX Toolkit (v. 0.0.13; http://hannonlab.cshl.edu/fastx_toolkit/). Overlapping paired-end reads were merged into a single read using PANDAseq (v. 2.9)[45]. Merged reads were processed using filter_fastq_by_primerid_length.pl to identify and extract the primerID from each read, with options—post CA—removepost, and place the primerID sequence in the sequence id line for each read. Each FASTQ file was split into separate amplicon libraries using Btrim64[46] with options -u 2 -v 2 -S -B -e 300; a list of primers for each amplicon was also supplied. Off-target sequences were removed with get_majority_block_bam.pl, a wrapper script that maps the reads using BWA MEM (v. 0.7.12-r1039[47] with options -B 1 -M, converts the output to BAM using Picard SortSam (v. 1.130; http://broadinstitute.github.io/picard/), and retains only the reads that have the most common mapped start and stop coordinates using BEDTools (v. 2.19.1)[48]. Reads were collapsed into consensus sequences using merge_primerid_read_groups.pl with options—ambig 600—min_freq 0.75 to require that the consensus base called makes up at least 75% of the bases for that position within the read group, otherwise an ambiguous base is called. The minimum group size (-m) to use in the merging process was computed for each amplicon within each sample (i.e., each library), using a formula derived for 12 bp primerID length, based on the simulation model proposed by Zhou et al.[23] in order to reduce the effect of offspring primerID reads from large primerID groups on smaller primerID groups. (To derive the formula for 12 bp primerIDs, the script consensus_cutoff.rb was run with length_of_primer_id set to 12.) If the computed minimum group size was less than 5, then -m was set to 5. Consensus reads were converted to frequency tables for nucleotides, codons, and amino acids at each position within the amplicon using convert_reads_to_amino_acid.pl.
Sequence analysis of the original SV12 stock library (Fig 2), generated using a slightly different barcode primer design, substantively differed from the optimized analysis approach described above as follows: PrimerID groups required a minimum of 2 reads, and the intra-sample consensus was used as a tiebreaker.
To determine linkage between high frequency variants (MAF > 0.5%), we performed a Fisher’s exact test for every pair of variants within each replicate and we derived a combined p-value from these individual replicate p-values using Fisher’s method.
All software used for the optimized primerID analysis is available on Github: http://github.com/niaid/primer-id-progs.
|
10.1371/journal.pntd.0003468 | Differences in Type I Interferon Signaling Antagonism by Dengue Viruses in Human and Non-Human Primate Cell Lines | In vitro studies have shown that dengue virus (DENV) can thwart the actions of interferon (IFN)-α/β and prevent the development of an antiviral state in infected cells. Clinical studies looking at gene expression in patients with severe dengue show a reduced expression of interferon stimulated genes compared to patients with dengue fever. Interestingly, there are conflicting reports as to the ability of DENV or other flaviviruses to inhibit IFN-α/β signaling.
In order to determine the relative inhibition of IFN-α/β signaling by DENVs, a method combining flow cytometry and a four-parameter logistic regression model was established. A representative isolate from DENV-1, -3 and -4 and seventeen representative isolates encompassing all DENV-2 genotypes were evaluated. All of the DENVs evaluated in this study were capable of inhibiting IFN-α/β signaling. Most of the strains were able to inhibit IFN-α/β to a degree similar to DENV strain 16681; however, DENV-2 sylvatic strains demonstrated an increased inhibition of phosphorylated signal transducer and activator of transcription (pSTAT1). Surprisingly, we were unable to observe inhibition of pSTAT1 by DENV-2 sylvatic strains or the Asian strain 16681 in non-human primate (NHP) cell lines. Analysis in primary Rhesus macaque dendritic cells suggests that DENVs are capable of inhibiting IFN signaling in these cells. However, contrary to human dendritic cells, production of IFN-α was detected in the supernatant of DENV-infected Rhesus macaque dendritic cells.
The ability of DENVs to inhibit IFN-α/β signaling is conserved. Although some variation in the inhibition was observed, the moderate differences may be difficult to correlate with clinical outcomes. DENVs were unable to inhibit pSTAT1 in NHP cell lines, but their ability to inhibit pSTAT1 in primary Rhesus macaque dendritic cells suggests that this may be a cell specific phenomena or due to the transformed nature of the cell lines.
| Dengue is a viral illness acquired through the bite of an infected mosquito. This flu-like illness, which in rare instances can be fatal, threatens more than half of the world’s population. Both in vitro and clinical studies looking at how the virus operates have consistently found that the interferon response is modulated by the virus during infection. We looked at the ability of dengue virus (DENV) strains to inhibit phosphorylated signal transducer and activator of transcription (pSTAT1) after IFN-β stimulation and observed that contrary to earlier published reports; all DENVs are capable of inhibiting IFN-α/β signaling. Strains from the DENV-2 sylvatic genotype, which mainly infect non-human primates (NHP), displayed an increased ability to inhibit pSTAT1 compared to the Asian strain 16681. To our surprise, DENVs were only capable of inhibiting pSTAT1 in human cell lines, but not in NHP cell lines. Inhibition of pSTAT1 is observed in both human and NHP primary dendritic cells. These results have important implications in the use of NHP cell lines for studies of IFN-α/β inhibition by DENV in vitro and may be a relevant consideration when using NHPs for DENV pre-clinical studies.
| More than half of the world’s population is at risk of acquiring an acute mosquito-borne illness known as dengue [1]. Infected individuals can be asymptomatic or display a range of clinical features. Many symptomatic dengue patients experience a mild fever, however, some develop severe dengue complications resulting in plasma leakage, hemorrhage, and organ impairment [2].
Dengue virus (DENV) contains a ∼10.7 kb positive strand RNA genome that encodes 3 virus structural proteins (C, prM, and E) and seven nonstructural (NS) proteins (NS1, 2A, 2B, 3, 4A, 4B and 5) [3]. There are four serotypes of DENV (DENV-1, -2, -3, & -4) and each is further sub-classified into genotypes. Some studies have observed differences in virological characteristics and clinical outcomes that associate with certain genotypes [4–7]. So far, these correlates of disease severity have been most extensively studied in the DENV-2 genotypes. The key elements hypothesized to contribute to disease outcome come from both virus molecular determinants and host factors [5,8–10].
The acute nature of DENV infections suggests that the innate immune system plays a vital role in its elimination. Type I interferon (IFN-α/β) is produced in response to the detection of DENV RNA by various pathogen-recognition receptors [11,12]. The IFN-α/β produced can bind cell surface receptors and cause dimerization of the IFN-α/β receptor subunits [13]. As a result, the JAK/STAT pathway is activated. The phosphorylation of STAT1 creates binding sites that allow homodimerization of STAT1 and heterodimerization of STAT1–2 [14,15]. STAT1 or STAT1–2 dimers are joined with IRF9/p48 to form a trimeric complex named ISGF3 [16,17]. The mature ISGF3 complex functions as a transcription factor that enters the nucleus and binds to promoter sequences in DNA containing interferon stimulated response elements. Over three hundred interferon stimulated genes are induced by IFN-α/β signaling [18]. These genes encode products that help uninfected cells to establish an antiviral state [19–22].
DENV has evolved to thwart the IFN-α/β antiviral response. Initial studies demonstrated that established DENV infections in vitro were refractory to the inhibitory effects of IFN-α/β on replication [23]. Gene array studies of clinical samples from DENV infected individuals have highlighted the relevance of the interferon system with regards to pathogenic outcomes. These studies have shown that the interferon system contains some of the most highly regulated genes during infection. Patients with severe dengue, such as dengue hemorrhagic fever or dengue shock syndrome, show suppression of IFN-α/β-stimulated genes compared to those with dengue fever [24,25].
Pathologically relevant flaviviruses such as JEV, WN, KUN, and TBE also inhibit the signal transduction cascade exerted by IFN-α/β [26–29]. Some studies have suggested that not all DENV or flaviviruses are capable of blocking IFN-α/β signaling [30]. Furthermore, studies of JEV and WNV strains have suggested a correlation between disease severity and the ability to inhibit IFN-α/β signaling [31–33]. Our aim in this study was to determine if these observations could be extended to DENV. However, the variable plaque morphology and growth characteristics make it difficult to compare the differential effect flaviviruses may exert on host signaling pathways. In this study we developed a quantitative method that allowed us to determine the relative IFN-α/β blocking ability among DENV strains. Comparisons were made with low-passage clinical isolates from all DENV serotypes and with representative DENV-2 strains of the American, Asian, American/Asian, cosmopolitan, and sylvatic genotypes. In contrast to previous observations with DENV and other flaviviruses, we show that all DENVs are capable of blocking IFN-α/β signaling. Most strains suppressed pSTAT1 to levels similar to those observed with DENV strain 16681, except for the DENV2 sylvatic genotype. The degree to which DENVs block IFN-α/β signaling did not correlate with their replication. The correlation of the strength of IFN-α/β signaling inhibition and pathogenic potential will likely be difficult to demonstrate due to the modest variations observed between DENV strains.
In this study we show that there are differences in inhibition of pSTAT1 by DENVs in human and non-human primate (NHP) cell lines. However, DENVs are capable of inhibiting pSTAT1 in both human and Rhesus macaque primary dendritic cells. Previous studies with human dendritic cells showed that DENVs inhibit IFN-α/β production. Our results show that Rhesus macaque primary dendritic cells readily produce IFN-α when infected. This is the first study to suggest possible differences in the innate immune response to DENVs in humans and NHPs.
Aedes albopictus C6/36 cells (ATCC #CRL-1660, Manassas, VA) and A549 human lung epithelial cells (ATCC # CCL-185, Manassas, VA) were cultivated in DMEM (Invitrogen, Grand Island, NY) supplemented with 10% heat-inactivated fetal bovine serum (FBS), L-glutamine, and nonessential amino acids. C6/36 cells were maintained at 33°C and A549 cells at 37°C with 5% CO2. Vero African green monkey cells (ATCC # CCL-81, Manassas, VA) were cultivated in M199 medium (Invitrogen, Grand Island, NY) supplemented with 10% heat-inactivated FBS L-glutamine and nonessential amino acids and maintained at 37°C.
Twenty DENV strains were used in this study. Four low-passage strains representing each of the DENV serotypes (DENV1: 101–001/PR1998, DENV2: BID-V681, DENV3: BID-V1610, and DENV4: BID-V2442) were obtained from the Centers for Disease Control Dengue Branch Passive Dengue Surveillance System (San Juan, Puerto Rico) [34]. For studies using DENV-2 genotypes, representative strains TH/DB052/2003, 203–001/TW1987, 201–001/PR2006, and BID-V585 were obtained from the Centers for Disease Control Dengue Branch Passive Dengue Surveillance System (San Juan, Puerto Rico); DENV-2 strains 16681, and PR-159 were obtained from the Centers for Disease Control dengue reference strain collection; DENV-2 strains 131, IQT2133, Ven2, K0049, Mara3, 1349, and ArA6894 were kindly provided by Dr. R. Rico-Hesse (Baylor College of Medicine, Houston, TX); DENV-2 strains DakAr510, DakAr75505, and DkD811 were kindly provided by Dr. Nikos Vasilakis (University of Texas Medical Branch, Galveston, TX).
Plaque assays were performed as described previously [35]. Images of stained plaques were taken in a Bio-Rad Molecular Imager Gel Doc XR+ System (Bio-Rad, Hercules, CA). The area (mm ± standard deviation) of each plaque was determined by means of an image analyzing program (Digimizer® version 4.2.1, MedCalc Software, Mariakerke, Belgium).
DENV stocks were titered by flow cytometry using a previously described method that yields results similar to the standard plaque assay [36]. Infected cells were harvested, fixed with BD Cytofix, permeabilized with BD Cytoperm (BD Biosciences, San Jose, CA) and stained with the Alexa 647 conjugated monoclonal antibody (MAb) 2H2 (prM-specific, dengue-complex cross-reactive MAb). Unconjugated 2H2 was kindly provided by Dr. Robert Putnak, Walter Reed Army Institute of Research. Cell infectivity was quantified with a BD FACS Calibur and using BD Cell Quest software. The titer of the virus was determined using the following formula: fluorescence-activated cell sorting (FACS) infectious units/ml = [(% of infected cells) × (total number of cells per well) × (dilution factor)]/ (volume of inoculum added to cells).
DENV strains were used to infect A549 cells at an MOI of 2 for 24 hours. Cells were then stimulated for 30 minutes with 500 U/ml IFN-β (PBL interferon source, Piscataway, NJ) and fixed with 2% paraformaldehyde. After methanol permeabilization, cells were co-stained with anti-pSTAT1 Alexa 488- and anti-DENV prM Alexa 647-conjugated antibodies. Cell fluorescence was measured on a BD FACS Calibur and data analysis was conducted using BD Cell Quest software. DENV infected (prM+) cells were gated and analyzed to determine the percentage of pSTAT1(+) cells.
The pSTAT1 inhibition assay is performed akin to an ELISA assay with a standard curve. To construct the graph in Figs. 1 and 2, the percent of cells that stained DENV positive was calculated and a gate was placed on the DENV positive population at each of the virus dilutions to analyze the percent of cells that also stained positive for pSTAT1. The data points for graphs 1 and 2 display as curved shape, therefore, instead of using the formula y = Mx + b to calculate the expected percent of inhibition from DENV strains by linear regression analysis a 4-parameter logistic curve fit was performed to obtain the expected percent inhibition. The expected DENV inhibition of pSTAT1 was calculated using a 4-parameter logistic (4PL) model developed from a standard curve of serially diluted DENV strain 16681 infected cells that were stimulated with IFN-β. Serial 1:3 dilutions of 16681 were performed starting with an MOI = 6. The expected inhibition of pSTAT1 by DENV strains at the obtained infectivity was calculated using the four-parameter logistic regression (4PL) model fit using the following equation:
f(x)=c+d−c1+exp(b(log(x)−log(e)))
x = expected % inhibition of pSTAT1
b = slope factor
c = the response at zero infection
d = the response at infinite % infection
e = mid-range inhibition of pSTAT1 by DENV 16681
Using Fig. 1 as a crude example, we extrapolate the data to have a rough estimate of how much inhibition of pSTAT1 would be expected if a strain blocked IFN to the same degree as 16681. For example, if strain A had the same ability to inhibit pSTAT1 as 16681 and approximately 45% of the cells were infected, we would expect to see only 25% of the DENV positive cells to also be pSTAT1 positive (see Fig. 1). In this assay, any strain can be used as the reference strain.
The relative inhibition of pSTAT by the DENV strain was calculated by: (Expected pSTAT1 inhibition by 16681 at the obtained percent of infection; as determined above)–(observed pSTAT1 inhibition by DENV strain).
A549 cells were infected with representative DENV strains at an MOI = 0.1 for 1 hour at 37°C. After removal of the virus inoculum, cells were washed four times with PBS and grown in DMEM containing 2% FBS. Supernatants from infected cells were collected at 12, 24, and 48 hours post-infection. Samples (280 μl) were processed for RNA extraction on an automated M48 BioRobot using the MagAttract Viral RNA Kit (Qiagen, Valencia, CA). The changes in DENV amplification over time were assessed using a standard curve of in-vitro transcribed RNA generated from plasmids containing amplicons for DENV1 (NS5 gene), DENV2 (E gene), DENV3 (M gene), and DENV4 (M gene) using AmpliScribe T7-Flash Transcription Kit (Epicentre Biotechnologies, Madison, WI) per manufacturer's protocol. Contaminating plasmid DNA was removed with the Ambion TURBO DNA-free Kit (Life Technologies, Grand Island, NY) and RNA was quantified using a NanoDrop ND-1000 Spectrophotometer. Starting with 100 μg of in-vitro transcribed RNA, eight serial 1:10 dilutions were performed. We performed real time RT-PCR to quantify DENV genome copy equivalents/ml of the sample using a previously described method[37].
Human (A549, Huh7) and monkey (LLCMK2, Vero) cell lines were infected with DENV-2 16681 at an MOI of 5 for 24 hrs. After a 30 minute stimulation with IFN-β, cell lysates were prepared with RIPA buffer containing a protease inhibitor and phosphatase inhibitor cocktail (Roche Diagnostics GmbH, Mannheim, Germany). Samples were clarified by centrifugation at 12,000 x g for 10 min. at 4⁰C and protein concentrations were quantified using the bicinchoninic acid reagent (Thermo Scientific Pierce, Rockford, IL). The volume required for 10 μg of protein was loaded and separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis and transferred to nitrocellulose. The primary antibodies in this study were used at the manufacturer’s recommended dilutions: anti-pSTAT1 and anti-GAPDH (Cell Signaling Technologies, Beverly, MA), anti-STAT1 (Sigma-Aldrich, St. Louis, MO), anti-pSTAT2 and anti-STAT2 (R&D Systems, Minneapolis, MN), anti-DENV NS4B (Abcam, Cambridge, MA). Horseradish peroxidase-conjugated secondary antibodies (1:5,000 dilution; Jackson ImmunoResearch, West Grove, PA) were used to visualize bound primary antibodies with the Supersignal chemiluminescence substrate (Thermo Scientific Pierce, Rockford, IL).
DENV strains were compared to 16681 as a control by one-way analysis of variance (ANOVA) followed by Dunnett's Multiple Comparisons Test in GraphPad 5 (GraphPad InStat, GraphPad Software, San Diego, CA). Statistical differences in IFN-α/β inhibition between individual viruses were assessed by unpaired two-tailed T-test analysis. Differences were considered significant when p < 0.05.
The phylogenetic trees of DENV strains utilized in this study were created with Mega 5, using maximum likelihood analysis based on nucleotide sequences of the complete envelope gene.
One of our objectives was to compare the ability of DENV strains to inhibit IFN-α/β signaling. To help account for differences in strain infectivity when using the same MOI, we developed a method for quantifying inhibition of IFN-α/β signaling by DENV through inhibition of STAT-1 phosphorylation (pSTAT1). We selected pSTAT1 as a marker of IFN antagonism based on its relatively high expression compared to other molecules in the JAK/STAT pathway and because detection can be readily achieved by flow cytometry with commercially available MAbs. To establish our assay we used the prototypical DENV strain 16681, previously shown to inhibit IFN-α/β stimulation, as our reference virus [38]. The inhibition of IFN signaling was assessed by flow cytometry of A549 cells that were infected with serially diluted 16681 and treated with IFN twenty-four hours post-infection. Only cells that stained positive for DENV M protein were gated and analyzed for STAT1 phosphorylation at each dilution. A decrease in pSTAT1 expression indicated an inhibition of the IFN signaling cascade. Our results show that inhibition of pSTAT1 was observed in all of the virus dilutions. The DENV blockade was still observed at the lowest MOI (0.025) where approximately 10% of the cells were infected. However, a comparison between the different MOIs demonstrates that the magnitude of the inhibition, or percentage of pSTAT1(+) cells, depends on the percentage of infected cells. The percentage of pSTAT1(+) cells was inversely correlated with the percentage of DENV infected cells (Fig. 2 and Table 1). These results suggest that only analyzing DENV(+) cells is insufficient for making comparisons between strains and that the level of infection must be accounted for when determining the relative IFN antagonism potential. In order to normalize for differences in infectivity we created a dose response curve from A549 cells infected with serial dilutions of 16681 and stimulated with IFN-β. The dose-response curve was fit using a 4PL logistic regression and an equation was created from the data sets to obtain the expected percent of pSTAT1(+) cells at any given amount of infection (see Materials and Methods). This allows us to use the 16681 IFN dose-response curve as a reference to which we can compare the relative IFN antagonism of DENV strains.
Using our previously described method we sought to determine the ability of isolates from the four DENV serotypes to prevent STAT1 phosphorylation after stimulation with IFN-β. To determine the relative percent inhibition of pSTAT by DENV strains compared to 16681 we calculated the difference between the expected pSTAT1 inhibition at the obtained percent of infection and the observed pSTAT1 inhibition. A phylogenetic analysis of the sequenced DENV strains from our Passive Dengue Surveillance System (PDSS) was performed and we selected those that represented ancestors of recently circulating viruses (S1A-S1D Fig. and S1 Table). A description of these viruses can be seen in Table 2 (see S2 Table for plaque sizes). Our results show that all of the DENV serotypes evaluated had the capacity to block the IFN-α/β response (Fig. 3A). However, there was some variability in the levels of inhibition of pSTAT1 by these DENV serotypes compared to 16681. DENV-1 & 2 had slightly lower levels of inhibition at-16% and-10% respectively. The lowest levels of inhibition were observed with DENV-3 at-34% and the highest levels were observed with DENV-4 at 10% (Fig. 3A). The largest (44%) difference in inhibition was observed between DENV-3 and-4.
To explore the possibility of differences in IFN-α/β antagonism by DENV strains associated with different pathogenicities, we studied several strains from the clinically well-characterized genotypes of DENV serotype 2 (DENV-2). A549 cells were infected with strains from the American, Asian, American/Asian, Cosmopolitan, and sylvatic genotypes. As observed with the four DENV serotype strains examined, the levels of inhibition of pSTAT1 by most of the DENV-2 strains was similar to that observed with strain 16681. Most of the strains did not display a consistent relative increase or decrease in inhibition of pSTAT1 by genotype (Fig. 3B). Interestingly, the DENV-2 sylvatic genotype consistently displayed higher levels of pSTAT1 inhibition (33–61%) compared to 16681 and all other strains (p < 0.01). The examination of all of the DENV strains in this study suggests that all DENVs are capable of inhibiting pSTAT1 and that this function, albeit with small variations, is highly conserved.
The replication rate of all the DENVs used in this study was measured to determine if it could play a role in the observed IFN-α/β antagonism of these viruses. Only DENV-1 displayed lower growth than the other three DENV serotypes (Fig. 4A). However, the level of pSTAT1 inhibition of DENV-1 was not as low as DENV-3, which displayed a similar replication rate as DENV-4, the best pSTAT1 inhibitor. Measurement of the replication rate of DENV-2 strains showed that the Asian strains replicated to higher levels than American/Asian and American strains (Figs. 4B, 3C, 3D & 3E). The replications rates of strains in the Cosmopolitan and sylvatic genotypes were variable and inconsistent within the genotype (Fig. 4E & 4F). DENV DkD811 displayed one of the lowest replication rates. Together, the data demonstrates that the replication rate of DENVs does not cause interference in our model assessing the IFN inhibitory ability of these viruses.
It is hypothesized that sylvatic DENVs, mainly detected in West Africa and Malaysia, are transmitted in an enzootic cycle most likely between NHPs and arboreal Aedes spp. mosquitoes. Therefore, we wanted to determine whether the increased inhibition of STAT1 phosphorylation by sylvatic DENVs was also observed in NHP cells. Surprisingly, none of the sylvatic DENV strains used in this study were able to prevent STAT1 phosphorylation. Results were comparable in LLCMK2 (Rhesus macaque) and Vero (Cercopithecus aethiops) cell lines (Fig. 5).
The unexpected finding that sylvatic DENVs were unable to inhibit STAT1 phosphorylation in NHP cell lines led us to question whether this observation was unique to these viruses. For this reason, we compared inhibition of IFN signaling by DENV-2 strain 16681 in both human (A549, Huh7) and NHP (LLCMK2, Vero) cell lines. Flow cytometry analysis of STAT1 phosphorylation in the DENV+ population revealed that there is a slight reduction in NHP cells compared to uninfected cells stimulated with IFN-β (Fig. 6A). As expected, DENV 16681 inhibited phosphorylation of STAT1 and caused STAT2 degradation in human cell lines. However, in NHP cell lines, STAT1 was phosphorylated in infected cells after IFN stimulation, but degradation of STAT2 was observed (Fig. 6A & B). This demonstrates a significantly reduced ability of DENVs to inhibit STAT1 phosphorylation in NHP cell lines.
In order to determine if the observed reduction in pSTAT1 is also observed in primary cells we differentiated CD14+ monocytes into myeloid dendritic cells. Analysis of human dendritic cells infected with or without dengue and stimulated with IFN revealed results comparable to those obtained with human cell lines (A549, Huh7). Phosphorylation of STAT1 was observed in uninfected dendritic cells after stimulation with IFN. These levels of pSTAT1 did not increase in DENV infected human and Rhesus macaque dendritic cells after stimulation with IFN (Fig. 7A). An increase in pSTAT1 was observed in human and Rhesus macaque dendritic cells that were infected with DENV, but not stimulated with IFN, suggesting that IFN was being produced. Testing of Rhesus macaque dendritic cell supernatants twenty-four hours post-infection revealed the presence of IFN-α. No IFN-α was detected in human dendritic cell supernatants (Fig. 7B).
There is mounting evidence for conserved mechanisms among flaviviruses to counteract the IFN-α/β response [26–29,39]. However, some studies suggest that not all flaviviruses are capable of inhibiting STAT1 phosphorylation and in some instances it resulted in reduced pathogenicity [30–33]. Most of these comparative studies are limited in the number of virus strains analyzed and the clinical data available. Therefore we aimed to do a more exhaustive study evaluating the capacity of DENVs to interfere with IFN-α/β signaling and if any observed difference could correlate with disease outcome. A previous study on DENV-induced STAT2 degradation and our unpublished observations between DENV-2 strains shows that this inhibitory mechanism is conserved among all DENVs [40]. Therefore, we focused on phosphorylation of STAT1 as a marker for differences in IFN-α/β signaling antagonism by DENVs.
Signaling though the Jak/Stat pathway can occur through STAT1 and STAT2 homodimers or heterodimers. The presence of both molecules is not required for activation of the Jak/Stat pathway to occur. However, the ISG response is more potent when both molecules are present. This general observation has been studied in the context of dengue virus as well. Studies by Shresta et al. demonstrated that a STAT1-independent IFN response confers protection from DENV infection in mice [41]. This suggested that signaling through STAT2 homodimers could mount an antiviral response that was sufficient to combat DENV infection. Furthermore, studies by Perry et. al. showed that both STAT-1 and STAT-2 single-deficient mice can survive a DENV challenge whereas STAT1/2 double deficient mice succumbed to early death [42]. This showed that STAT1 homodimers could also mount an antiviral response that was sufficient to combat DENV infection. These results suggest that there is a compensatory mechanism that protects against DENV in the absence of either one of the STAT proteins. Additional experiments done by Perry et. al suggest that STAT1 plays a larger role in the anti-DENV response than STAT2. Part of this evidence is supported by the fact that the viral load in STAT2 deficient mice was lower than their STAT1 counterparts at multiple time points after infection [42]. Together these data suggest that inactivation of either STAT1 or STAT2 alone is not sufficient to eliminate the antiviral response.
Studies evaluating differences in the cellular response to flaviviruses frequently encounter the difficulty of performing an accurate comparative analysis between strains. DENV isolates from humans or mosquitoes can differ in their morphology, plaque size and replication rate [43–45]. This makes it difficult to normalize for infectivity when comparisons are made between virus strains. For some DENVs, viral quantification by plaque assay can be difficult. Viral plaques can sometimes be very small and/or difficult to discern. For this reason, we developed a method akin to an ELISA assay for quantifying antagonism of IFN-α/β signaling based on the measurement of pSTAT1 inhibition in DENV(+) cells using flow cytometry. Comparing inhibition of IFN-α/β signaling molecules between DENV strains by RT-PCR, luciferase assays and Western blot analysis in infected cells can lead to misleading results since uninfected cells will be a confounding factor in the results. The use of Western blot analysis can also be problematic when MAbs to variable regions or polyclonal antibodies are used as a control for virus proteins. In these cases, even when plaque size and infectivity are same, the virus band signal intensity is greater in viruses that are more similar to the virus used for antibody development. Our data shows that when the percent of infected cells is not taken into account it can lead to an over- or underestimation of the virus’s antagonistic activity. The measurement of pSTAT1 in the DENV(+) population without standardization was inadequate for a comparative analysis between strains because the level of pSTAT1 inhibition increases with the percentage of infected cells. The virus replication rate did not have a confounding effect in our analysis as it did not correlate with the ability of DENVs to block phosphorylation of STAT1.
To our knowledge, we present the first concurrent comparison of all four DENV serotypes for their ability to block IFN-α/β signaling. All of the representative clinical isolates representing the four DENV serotypes were capable of inhibiting STAT1 phosphorylation to a significant degree, with low variability. Prospective clinical studies will be needed to determine whether this moderate reduction in inhibition of IFN-α/β signaling by DENV-3 results in pathogenic differences between DENV serotypes or strains. DENV-2 viruses, which have been better described clinically as to their pathogenic potential than other DENV serotypes, did not vary greatly in their ability to inhibit pSTAT1 [4,7]. Although DENV-2 Asian strains have been shown to cause more severe disease compared to American strains, their ability to inhibit pSTAT1 did not differ significantly from the other genotypes.
Viruses from the DENV-2 sylvatic genotype distinguished themselves by consistently displaying an increased capacity to inhibit STAT1 phosphorylation. The replication rate did not have a confounding effect in our analysis as they did not correlate with the ability of DENVs to block phosphorylation of STAT1. Our findings are consistent with previous observations showing that the DENV-2 Asian genotype has a similar level of replication, but attains higher virus titers than all other genotypes. Therefore, the increased peak viral load observed in strains of the DENV-2 Asian genotype are likely capable of causing an overall increased capacity to inhibit IFN-α/β signaling and induce more severe illness than other genotypes. This is simply due to the fact of increased presence of the virus and not due to a viral genetic determinant conferring increased virulence because of an improved capacity to antagonize IFN-α/β.
Documentation of human cases with viruses of the sylvatic genotype is not common, but spillover events have been observed in West Africa and Malaysia [46–49]. The cases described in West Africa are mostly of mild disease. Only two cases of dengue hemorrhagic fever caused by sylvatic dengue serotype 2 viruses have been documented. One occurred in West Africa and the other in Malaysia [47,50]. Therefore, it is difficult for us to address the possible implications this increased inhibition could have on clinical outcomes of human infections. Sylvatic DENV strains are considered ancestors of the current circulating genotypes that infect humans and are maintained through a mosquito-NHP cycle. Although we are not certain whether the lack of inhibition of pSTAT1 by DENVs occurs in non-hematopoietic primary NHP cells as observed in NHP cell lines, we did observe an increased ability of sylvatic strains to inhibit pSTAT in human cell lines. Inhibition of pSTAT by DENV remained intact in primary Rhesus macaque dendritic cells. One limitation of our study is that we were unable to assess expression of ISGs due to lack of expression of antiviral proteins in the NHP cell lines. The lone study to date that utilized a sylvatic DENV strain to infect NHPs showed that the viremia was similar to that of human-endemic DENV-2 but lasted for a shorter period of time and no differences in pathogenicity were observed [51]. Although the sample size was small in this study, it suggests that the increased ability of sylvatic strains to inhibit pSTAT1 does not result in any obvious pathogenic differences in NHPs when compared to infections by human-endemic DENVs.
We observed production of IFN-α in DENV-infected primary myeloid Rhesus macaque dendritic cells that had not been stimulated. In contrast, no production of IFN-α was detected from DENV-infected primary myeloid human dendritic cells. Our observations in human dendritic cells are in concordance with those obtained by Rodriguez-Madoz et. al. under similar experimental conditions in which they also used DENV-2 16681 in their study and did not observe production of IFN-α/β at 24 hours post-infection with different MOIs ranging from 0.2–25. These observation together with the high levels of IFN-α detected in NHP dendritic cell supernatants suggests that inhibition of IFN-α production by DENV in NHPs does not occur. These results highlight a key difference in the IFN response between humans and NHPs in primary cells.
These limited, but interesting observed differences between human and NHPs cells deserve further study as they may add to the growing body of evidence to the distinct immune responses in primate hosts. It is known that humans are more severely affected than NHPs to a large number of diseases. Examples include HIV infections progressing to AIDS, Plasmodium falciparum infection progressing to malaria, and adverse complications following infections with hepatitis B, C and DENV [52,53]. The differences in susceptibility are thought to arise from inter-species differences in the immune response to infections. For example, in sooty mangabey monkeys which are unaffected by SIV or Yellow Fever, the innate and adaptive T cell proliferative responses are limited compared to Rhesus macaques and humans [54,55]. In a comparison of genome-wide gene expression levels between humans, chimpanzees, and rhesus macaques, it was observed that the innate response associated with viral infections is often lineage-specific [56]. Moreover, even though expression of PAMPs is similar in human and Rhesus macaque myeloid dendritic cells, the cytokine response differs [57]. This suggests that although humans and NHPs have similar innate immune signaling pathways, there are differences that will result in different outcomes in terms of expression and pathogenic outcome.
Studies by Umareddy et. al [30] suggested that some DENV strains are unable to suppress pSTAT1. Among these was TSV01, a cosmopolitan strain with 98.9% amino acid similarity to 1349 that was used in our experiments. Our results showed that 1349 was able to inhibit pSTAT1 at a level relatively higher than strain 16681. Our comparison of the NS4B amino acid sequence between TSV01 and 1349 shows that they are identical. These results suggest that either NS4B from TSV01 should be capable of inhibiting pSTAT1 or that another dengue protein aside from NS4B plays a more prominent role in preventing STAT1 phosphorylation in these strains.
The use of targeted mutagenesis of IFN-α/β-antagonizing flavivirus proteins to make IFN-α/β-sensitive viruses has been proposed as a method of vaccine attenuation [32]. Our results suggest that contrary to what has been published for other DENVs and flaviviruses, most, if not all DENVs have the capacity to inhibit STAT1 phosphorylation. This suggests that the selection of naturally occurring IFN-α/β-sensitive DENVs is not likely and attenuation must be achieved through genetic means. Targeted sites should be chosen carefully when designing DENV vaccine candidates and sensitivity to IFN-α/β evaluated through a rigorous method such as the one we have presented here.
|
10.1371/journal.pcbi.1005666 | Spread of hospital-acquired infections: A comparison of healthcare networks | Hospital-acquired infections (HAIs), including emerging multi-drug resistant organisms, threaten healthcare systems worldwide. Efficient containment measures of HAIs must mobilize the entire healthcare network. Thus, to best understand how to reduce the potential scale of HAI epidemic spread, we explore patient transfer patterns in the French healthcare system. Using an exhaustive database of all hospital discharge summaries in France in 2014, we construct and analyze three patient networks based on the following: transfers of patients with HAI (HAI-specific network); patients with suspected HAI (suspected-HAI network); and all patients (general network). All three networks have heterogeneous patient flow and demonstrate small-world and scale-free characteristics. Patient populations that comprise these networks are also heterogeneous in their movement patterns. Ranking of hospitals by centrality measures and comparing community clustering using community detection algorithms shows that despite the differences in patient population, the HAI-specific and suspected-HAI networks rely on the same underlying structure as that of the general network. As a result, the general network may be more reliable in studying potential spread of HAIs. Finally, we identify transfer patterns at both the French regional and departmental (county) levels that are important in the identification of key hospital centers, patient flow trajectories, and regional clusters that may serve as a basis for novel wide-scale infection control strategies.
| Hospital-acquired infections (HAIs), including emerging multi-drug resistant organisms, threaten healthcare systems worldwide. Efficient containment measures of HAIs must mobilize the entire healthcare network. Thus, to best understand how to reduce the scale of potential HAI epidemic spread, we explore patient transfer patterns in the French healthcare system. We construct and compare the characteristics of three different patient transfer networks based on data on transfers of patients with diagnosed HAIs, suspected HAIs, or of all patients. Our analyses show that these healthcare networks, the patient populations that comprise them and the patient movement patterns are heterogeneous and centralized. Despite the differences in patient populations, the HAI-specific and suspected-HAI healthcare networks have the same underlying structure as that of the general healthcare network. We identify key hospital centers, patient flow trajectories, at both the regional and department (county) level that may serve as a basis for proposing novel wide-scale infection control strategies.
| The emergence and spread of multi-drug resistant organisms threatens healthcare systems worldwide.[1] This is particularly true concerning methicillin-resistant Staphylococcus aureus, vancomycin-resistant enterococci, and multi-resistant gram-negative bacteria such as carbapenemase-producing Enterobacteriaceae (CPE). Spread of CPE is now a global public health problem associated with patient transfers between healthcare facilities within the same country as well as across national borders, as shown in many recent studies.[2–7]
In recent years, patient transfer or referral data has been used to construct “healthcare networks” to propose innovative approaches for hospital infection prevention and control. Healthcare networks are cooperative healthcare systems where hospitals and other healthcare centers are linked by shared patients through secondary transfers or referral.[8, 9] Rather than being exclusive to one sole hospital, as Ciccolini et al. argue, the extent of hospital-acquired infection (HAI) spread is dependent on the healthcare network connected by inter-institutional patient transfers.[8] Heterogeneous hospital patient populations and the interactions that occur between them and with the community are important in the understanding of the spatial spread of HAI between hospitals across geographic regions.[9]
As early as 2007, studies applied more complex social network analysis approaches to reconstructed healthcare networks in order to demonstrate that infection control measures that take into account network properties can decrease the risk for outbreaks.[8, 10] Lee et al. consider network properties to assess the individual influence of different hospitals and the impact of hospital proximities on HAI spread on a regional scale.[11] Many studies show that healthcare networks display a community structure.[8, 12–14] Network analysis is especially effective in the identification of sensor hospitals for surveillance of HAIs.[15, 16] In addition, mathematical models of healthcare networks may serve to inform decision-makers on enhanced coordinated regional and national approaches to infection control strategies, in a context where increasingly centralized healthcare systems favor the spread of HAIs.[8, 15, 17]
Although national healthcare networks are informative regarding novel HAI control strategies, the impact of reconstructing these networks based on a general patient population rather than a HAI-diagnosed patient population has rarely been addressed. In this study, we assess and compare French healthcare networks based on either patients diagnosed with HAIs or the general patient population, in order to better understand the potential implications in terms of HAI spread predictions. To that aim, we perform social network analyses to describe the different patient flow patterns, network topology characteristics, and community clustering structure.
We analyzed and compared three different networks built using transfer data from an exhaustive database of all hospital discharge summaries in France in 2014: (1) a network based on transfers of patients with diagnosed HAI (HAI-specific network); (2) a network based on transfers of patients with suspected HAI (suspected-HAI network); and (3) the network of all patient transfers (general network).
More than 10 million hospital transfers were recorded in France in 2014, for a total of 2.3 million transferred patients, creating a hospital network of 2063 hospitals (nodes) and 50026 patient trajectories (edges) linking them (Table 1). Patients with a HAI-specific diagnosis created a healthcare network of 1266 hospitals and 3722 connections for 13627 patient transfers. A larger population of patients suspected to have an HAI infection formed a healthcare network of 1975 hospitals and 18812 connections for a total of 128681 patient transfers. With the increasing number of patient transfers, the networks increased from an average 5.88, 19.05, and 48.05 average connections per hospital (average degree k¯) and an average 2.31, 4.92 to 14.02 transfers per connection (average strength s¯) for the HAI-specific, suspected-HAI, and general healthcare networks respectively (Table 1).
Overall, the three networks displayed “scale-free” and “small-world” characteristics that indicated the presence of a small number of very highly connected hospitals with high degrees, referred to as “hubs.” Analyses of the degree, strength, and shortest path length distributions in addition to the small-world characteristics of the healthcare networks are discussed in S1–S3 Texts and S1–S7 Figs. Compared to random networks, we also showed the general network was more clustered and efficient in transferring patients (S4 Text, S1 Table). We identified several high degree hospitals in all three networks with a consistent outlier–the Assistance Publique—Hôpitaux de Paris (AP-HP)–a conglomerate of 39 hospitals predominately in Paris and the Ile-de-France region represented as one hospital code in our database.[18] AP-HP also acted as the most important intermediary hospital system in the networks due to having the highest betweenness centrality measure.
The hospitals involved in the patient transfers recorded in the three networks were of various types, including private rehabilitation and postoperative care facilities, acute-care hospitals or clinics, and hospital centers (Table 2). However, the majority of hubs, defined as the top 5% of hospitals by their degree, were large hospitals providing both acute and postoperative or rehabilitation care (67%, 65%, and 88% in the general, suspected-HAI, and HAI-specific networks respectively). In addition, in the general and suspected-HAI networks, hubs were mostly acute-care hospitals or clinics, hospital centers, or university hospitals centers, with many concentrated in the Ile-de-France, Marseille, and Lyon metropoles (31%, 33%, 28%, and 32%, 28%, 30% respectively). In contrast, university hospital centers rather than acute-care facilities dominated the hub hospitals of the HAI-specific network, representing 48% of hubs (Table 2). The hub university healthcare centers, which provided highly specialized services, included the AP-HP, Hospices Civils de Lyon, and the Assistance Publique—Hôpitaux de Marseille (AP-HM); among them there were also university hospitals of other major cities in France.
To better understand the role of hub hospitals across the networks, the shared hospitals between the networks were ranked based on their degree, closeness, and betweenness (Fig 1). Overall, when comparing the degree, betweenness, and closeness, the hospital rankings did not differ between the complete set of 1266 HAI-specific network hospitals and these same hospitals in the general network (p = 0.81, p = 1, p = 0.99 respectively, Wilcoxon rank sum test), or between the 1975 suspected-HAI network hospitals and the same hospitals in the general network (p = 0.99, p = 1, p = 0.99, Wilcoxon rank sum test). For comparison and illustration purposes, we showed that random rankings for degree, betweenness, and closeness of all hospitals differed significantly between patient specific networks and the general network (p < 0.05 respectively, Wilcoxon rank sum test) (Fig 1).
Suspecting that the differences between rankings might exist between subsets of hospitals, we tested the differences between rankings on an increasing subset of shared hospitals, starting with the highest rank, adding the next ranked hospital, and testing for significant differences. As a result, we determined the range of hospital rankings across the networks where the rankings significantly differed. We defined significant differences as Wilcoxon rank sum test p-values under the 5% alpha risk which we represent as a grey area in Fig 1. Distributions of these p-values are provided in S8 and S9 Figs. For the HAI-specific network, the range of statistically significant degree ranking differences were observed between the 24th ranked hospital to the 1159th ranking hospital. For the suspected-HAI network, statistically significant degree ranking differences were observed between the 405th ranked hospital to the 1078th ranked hospital.
For hospital rankings based on betweenness and closeness centrality measures, the hospitals ranked with highest and lowest centralities in the general network were also the hospitals ranked with highest and lowest centralities ranking in the HAI-specific and suspected-HAI networks. Even though hospital rankings of all hospitals did not differ, the majority did differ for betweenness ranks between the 33rd highest ranking to the 1183rd ranking in the HAI-specific network and the 71st highest ranking to the 1757th ranking in the suspected-HAI network (p < 0.05, Wilcoxon rank sum test). Closeness rankings differences were observed for almost all rankings after the first 3 rankings in the HAI-specific network and after the first 6 in the suspected-HAI network. The lack of statistically significant differences for the highest rankings may have been only due to insufficient power and for lowest hospital rankings due to a series of repeating small closeness values. With this method, we highlight that differences do exist for subsets of hospitals, but we also observe that the most highly connected hub hospitals were consistently highly connected across the networks, irrespective of the different patient population that connected them.
To further assess patient movement patterns in the networks, we investigated how our healthcare networks displayed “community” or hospital clustering structure. We compared hospital communities detected with two different community clustering algorithms: 1) the Greedy algorithm [19] that selected members of the communities to maximize the density of links between vertices as it reconstructed the network one vertex at a time and 2) the Map Equation algorithm [20], based on network structure-induced movement using a flow-based and information-theoretic method, detecting communities by measuring probability flows by taking into consideration the directionality and weight of the edges. In general, we detected fewer communities with the Greedy algorithm given that it seeks to maximize modularity–a value that measures the density of links inside communities by comparing the fraction of edges within the communities to the fraction in a random network; a maximum value of 1 corresponds to a network structure with the highest strength possible–as a result, the algorithm searched for the repartitions that maximized the density of the edges.[21–23] The Greedy algorithm considered pairwise interactions and the formation of the network whereas the Map Equation considered the interdependence of links and the dynamics of an already formed network.
For each network, we calculated the modularity, the number of communities, community size, and average community clustering distance using the Greedy and Map Equation community detection algorithms (Table 3). For each community, the pairwise clustering distance was calculated as the average geographic distance between all pairs of hospitals of the same community in kilometers. Compared to the general healthcare network, the patient-specific networks had more communities. In the HAI-specific network, there were on average 35.17 hospitals per community (SD = 44.31) and 31.40 kilometers between pairs of hospitals in the same community (SD = 25.60 km). In the larger networks, the larger community sizes resulted in a higher average distance between community hospitals (41.60 km (SD = 34.71) and 39.01 km (SD = 45.63) for the suspected-HAI and general healthcare network respectively). For the Map Equation-based communities, as the number of communities decreased from the HAI-specific to suspected-HAI to the general healthcare network, the average community size and average community distance between hospitals of the same community increased (Table 3). Overall, the suspected-HAI network was more similar to the general network than the HAI-specific network in terms of community structure (S5 Text).
The regional community clustering using the Greedy algorithm in the three networks are represented in Fig 2. The hospitals in communities were geo-localized, color-coded, and labelled across the networks according to the administrative region(s) they encompassed. We observed that the Greedy-based communities accurately reflected the French regional administrative structure (Fig 2). The identified community clusters formed hospitals communities in which most of the patients were shared between hospitals of the same region frequently centralized towards the hub acute-care centers, university hospital centers, and general hospital centers. On the other hand, the Map Equation-based communities displayed geographic community clustering at the French “departmental” or county level–the administrative division between the administrative region and the municipalities, similar to “counties” or “districts”; of which 96 departmental divisions are present in continental France. The vast majority of these departmental-level community clusters were acute-care centers followed by university hospitals centers and long-term care facilities.
To further understand the community structure, we constructed intercommunity networks by combining patient flows between hospitals of the same community and across communities. The Greedy-based intercommunity network was composed of 18 nodes representing the sum of all patient transfers that occurred between hospitals of each community with 306 regional transfer trajectories (Fig 3A). Out of the 22 French metropolitan regions in 2014, 4 pairs of 8 metropolitan regions were combined in this intercommunity network (Picardie and Champagne-Ardenne, Auvergne and Limousin, Aquitaine and Poitou-Charentes, and Bourgogne and Franche-Comté). The network was completely connected. All regional communities were connected to one another with an average of 4590 patients moving within these intercommunity trajectories over the year. Certain trajectories played a larger role in patient movement, notably Ile-de-France which admitted the largest number of patients from neighboring regions Picardie and Champagne-Ardenne (4772 transfers) and Centre (3205 transfers) where healthcare hubs were most concentrated. The subsequent largest traffic came from the Rhone-Alpes, the second largest regional center around the city of Lyon, which discharged patients to its neighboring regions (1482 transfers to neighboring Bourgogne and Franche-Comté and 1342 transfers to neighboring Provence-Alps respectively). Nonetheless, the greatest amount of transfers (93%) occurred within the communities themselves on average with up to 98% of transfers occurring within Ile-De France for instance. Although most of these transfers occurred within the communities, the regions remained highly interconnected and certain trajectories played an important role in the interregional and nation-wide movement of patients in France.
Building the intercommunity network where community affiliation was determined by the Map Equation algorithm allowed us to consider communities based on the directionality of patient flow, which was overlooked by the Greedy algorithm. The intercommunity network was composed of 113 community nodes with 3215 trajectories with an average degree of 57 and an average of 2597 patients moving between these connections (Fig 3B). Map Equation-based intercommunity communities demonstrated more comprehensive department-level patient flow. Communities were composed of hospitals from many different departments within and across regions; however, the majority of communities were of hospitals within the same department where most of the patient exchange occurred. Concerning the most important transfer routes with the highest traffic, discharged patients coming from many neighboring departments were preferentially going to hospitals in one or a few number of departments, indicating that there was interdepartmental centralization of patient movement. For example, a community composed of 200 hospitals from 9 Ile-de-France departments sent the largest number of transfers (3137 patients) to 28 hospitals of which 24 were from one department (Val-d’Oise). In exchange, this 28-hospital community sent back 2772 patients to the larger community. Overall, patient transfers in Map-Equation communities displayed departmental clustering, but also demonstrated asymmetric movements of patients, concentrating towards small communities of hospitals usually in one department, illustrating the different nature of the communities.
Patient sharing patterns and community clustering in the networks were also analyzed based on patient age groups in which new communities were identified (S6 Text, S10–S12 Figs). Moreover, analysis of monthly temporal dynamics of the networks showed that monthly communities may be less clustered and patients may not visit all of the hospitals each month but they still retained the same regional patient sharing patterns seen in the annual network (S7 Text).
Having assessed the role of hospitals, hospital communities, and patient trajectories in each network, we considered if the differences in the patient-specific networks and the general networks are due to the number of patient transfers that could lead to structural differences between the specific patient population flows. We first compared the general patient network to two sets of 1000 networks built from a subset of randomly chosen patients: in the first set we selected the same number of patients as the HAI-specific network (21276 patients) at random and in the second set the same number as in the suspected-HAI network (394859 patients) at random 1000 times and reconstructed each network. Overall, both sets of random patient networks (RP) were smaller in size compared the general network in terms of the number of nodes, edges, edge weight, and as a result average degree (Table 4). In addition, most of the diameters and all average path lengths were larger in the RP networks. The diameters and path lengths of the RP networks are skewed and not normally distributed (p< 0.001, Shapiro-Wilk normality test). As a result, the number of patients used to reconstruct the networks did have an impact of network characteristics.
We then compared the characteristics of the HAI-specific and suspected-HAI networks to the RP networks with the same number of patients to assess if HAI patients modified network structure differently than other patients. Overall, the RP networks were larger than their HAI-specific and suspected-HAI healthcare networks analogues meaning that HAI patients were transferred to fewer hospitals than other patients (Table 4). However despite these differences, for some networks measures such as diameter, average path length, and global clustering coefficient, there was less of a difference between the RP networks and the HAI networks than the RP networks and the general network. For example, 63% of suspected-HAI-like RP networks had a diameter equal to or less than that of the suspected-HAI network (64) while 1.9% of these networks had a diameter equal to or less than that of the general network (30). The average diameter (61.59) and the average path lengths (3.78) of these RP networks approached more of that of the suspected-HAI network than the general network. Thus having controlled for the number of patients and thus the size of the network, the differences observed between the suspected-HAI and the general network diameter and average path length may have been due to the suspected-HAI network being a subset of the healthcare network rather than due to differences between HAI patient transfer patterns and non-HAI patient transfers.
In this study we show that the French healthcare networks have heterogeneous patient flows, demonstrate characteristics of small-world and scale-free networks, and are characterized with highly centralized movement of patients towards hub hospital centers. Hub hospitals are characterized as university hospitals and private hospitals in the large metropoles that dominate patient flow. The healthcare networks displayed two-level community clustering: regional community clustering reflecting the French administrative structure, and department or county-level clustering. Certain patient transfer trajectories play a more important role in transferring patients between the regional and departmental communities. Despite differences in the patient population and size, both the HAI-specific and suspected-HAI specific healthcare networks seem to rely on the same underlying structure as that of the general healthcare network.
Due to weak sensitivity and specificity of the PMSI database to detect nosocomial infections with the sole ICD-10 Y95 diagnostic, the HAI-specific network is not reliable in demonstrating the real patient movement patterns for those infected with an HAI.[24–28] There was no confirmation if an infection was absent during admission and if an infection appeared during the first 48 hours of their stay. We suspect that the degree ranking differences and the low percent of acute-care facilities, notably private hospitals, in the HAI-specific network may be due to differences in coding practicing among hospitals rather than the epidemiology of HAIs. The suspected-HAI network reflects a standardized list of diagnoses used by the French HAI surveillance network which has been shown to be more specific and sensitive at detecting patients with HAIs.[28, 29] Having considered the network size differences in the patient-specific networks and the general network, we show that despite the differences in size of the patient population, both the HAI-specific and suspected-HAI specific healthcare networks seem to rely on the same underlying structure as that of the general healthcare network. Indeed, patient-specific networks are a subset of the general patient network and are subject to the same network dynamics.
Public university hospital centers and private hospitals in the main metropoles of France dominate patient flow. A study conducted in the Bourgogne region of France has shown that patient flow was centered towards the university hospital that admitted patients from the entire region and based on the regional proximity of the patients’ residence and patients also sought care in two of the closest main healthcare hubs for specialized care (Paris or Lyon).[30] Highly connected hospitals may harbor more MRSA and MRSA bacteremia cases and may have the most potential to transmit HAIs in the entire network.[12, 13, 31, 32] HAIs may spread at a higher rate than expected at random due to the centralization of patient movement and due to the small average number of transfers required for patients to move throughout the network. A 2012 point prevalence study has shown that HAIs are most prevalent in cancers centers, university hospitals, and armed forces.[33] HAI prevalence was high in the Ile-de-France region which has many hubs, and the north-eastern regions which were not reflected by a higher number of transfers in the patient specific.[33] Albeit some difference in prevalence and patient transfer patterns, hubs should be proposed as targets for sentinel surveillance in addition to priority targets of HAI control strategies where HAI is most prevalent to achieve the most effective reduction in transmission across the country.[15]
Regional community clustering patterns as a form of network connectedness are also important in the development of strategies for coordinated HAI control.[8, 13] Our regional community clustering findings are consistent with that of the healthcare network of England in which communities tend to share more patients among clusters of hospitals in addition to patient flows centered towards a university hospital within the community.[13] Important intermediary trajectories may play a key role in the spread of HAI between hub hospitals and between communities. A study has shown that modifying the number of patients moving between communities may reduce the spread of MRSA.[34] Furthermore, we demonstrated that a two-tier hospital community exists. Depending on the clustering algorithm used, we identified clustering of healthcare communities at the regional level, consistent with the French administrative regions, and department-level communities and inter-departmental hospital clusters that took into account the directionality of patient flow. Coordinated department-level control such as screening of patients based on the identification of key department-level cluster admissions may be the first line of defense against HAI spread within the regions before spread reaches the hub university hospitals. We identified differences between department-level communities of the suspected-HAI and the general network that were overlooked at the regional community level. This may be important in distinguishing hospitals with higher potential to harbor HAI patients, with possible consequences in terms of HAI spread prediction.
Studies have proposed reducing hospital connectedness in order to reduce the risk of epidemic spread of HAI in networks.[13, 35] Decentralization of the healthcare system and more specifically human resource and specialized health services towards the regional and department levels may help reduce the high connectedness of hubs in the metropole centers and redirect patient transfers. France has moved towards regionalization strategies with the creation of regional hospital agencies, albeit not very effective.[36, 37] In addition, the number of university hospitals may be insufficient, below that of the UK, a country with a similar population size. We recommend increasing the number facilities providing specialized services and distributing them at the local level to help redirect patient flow and potentially avoid large-scale HAI dispersal.
We considered temporal dynamics, masked in a network constructed with data for the entire year of 2014, in which observed that monthly healthcare networks were smaller and displayed less centralized patient flow; hence, infection control strategies–for short-term control–should rely more on the local department-level dynamics to minimize hospital-level outbreaks and transmission to neighboring hospitals. In the long term, regional community dynamics may give us clues regarding the gradual propagation of specific HAI pathogens over time assuming HAI carriage patterns follow that of patient flow patterns in the healthcare networks. Further studies are required to assess the temporal dynamics of HAI spread in networks to identify any potential seasonality patterns of flow and how to prevent emerging multi-drug resistant bacteria from becoming endemic.
Our study was subject to certain limitations which should be considered. Many of the university hospitals represent more than one public hospital or healthcare facility due to sharing the same identification number. For example, the largest outlier hub in Paris (AP-HP) represented 39 hospitals, 12 hospitals and 2 specialized health facilities constituted Hospices Civils de Lyon, 9 hospitals make up the university hospital of Toulouse, and 4 hospitals make up the APHM of Marseille. Consequently, university hospital centers accommodated a larger patient population than hospital centers or local hospitals, influencing the network characteristics, which may have led us to overestimate the specific patient movement patterns to and from these centers. However, the high concentration of other hospitals especially hub private hospital centers, armed forces hospitals, cancer centers, psychiatric hospitals, and private post-operative and rehabilitation centers within proximity of these public hospital hubs demonstrates that despite this, major cities such as Paris play the most important role in connecting patients in the national network and that the French healthcare network is a highly centralized system.
The healthcare networks did not include patient flow from private nursing homes that have been shown to play an important role in HAI spread.[38–42] Without private nursing homes included in our study, our results only describe the network topology of hospital patient populations which may be both younger, have shorter duration stay, and may spread HAI differently than the complete nursing home population. As a result, transmission dynamics in our networks may over or underestimate average hospital centrality measures, the volume of patient movements, and the speed at which HAI can spread.
By considering all HAIs as a whole, our networks and recommendations reflect action for a broad spectrum of HAIs; however, one should consider that specific HAIs can vary in terms of carriage and transmission patterns. In addition, recommendations based on our networks would overlook the potential exposure to community-acquired infections, although these may later spread in hospital settings, leading to healthcare-associated outbreaks. Future studies should consider all potential components of patient exposure to both community-associated and healthcare-associated infections and account for individual exposure histories to these infections.
Despite these limitations, our study provides a first description and analysis of the healthcare networks in France. The identified characteristics and community structures may greatly improve future inter-hospital HAI control strategies. The general patient network responds best to informing regional control strategies targeting key patient trajectories and hub hospital centers. We show that the scale-free structure, the number of communities, and their distribution over the country remain qualitatively similar across all networks and that patient-specific networks rely on the underlying structure of the general patient network. Future studies should take into consideration network topology in the prediction of HAI spread and should consider the potential impact of different community definitions for multi-level infection control strategies.
The Programme de Médicalisation des Systèmes d’information (PMSI) database, a comprehensive French medico-administrative database of hospital activity and patient discharge information, is used to construct the hospital networks.[24, 25] The PMSI database has been used for epidemiological and medical research regarding HAIs.[24–28, 43] A lack of sufficient specificity and sensitivity of the PMSI to detect HAIs is highlighted in these studies. Comparison between laboratory data and hospital data shows that the PMSI has limited coverage of detecting nosocomial conditions.[25–28]
Hence, Gerbier et al. [28] use patient discharge summaries from the PMSI to detect nosocomial infections in the University Hospital of Lyon in 2006 and 2007 for the identification of HAIs in surgery, intensive care and obstetric units. They compare the PMSI data to a gold standard by systematic review of patient files for those classified under surgery, the Centre de Coordination de la Lutte contre les Infections Nosocomiales (CClin) Southwest surveillance network for ICU patients, and a combination of surveillance data from CClin and patient information data for obstetrics. The list of ICD-10 codes related to nosocomial conditions, which we entitle “suspected-HAIs,” can be found in S1 Annex. Gerbier et al. find a sensitivity and specificity for case identification of nosocomial infections to be 26.3% (95% CI 13.2–42.1) and 99.5% (95% 98.8–100.0) for the identification of surgical site infections (78.9% and 65.7% by expanding the number of diagnostic codes) respectively; 48.8% (95% CI 42.6–55.0) and 78.4% (95% CI 76.1–80.1) in intensive care respectively, and 42.9% (95% CI 25.0–60.7) and 87.3% (95% CI 85.2–89.3) for identification of postpartum infections respectively.[28]
Using patient transfer data from 2014, three healthcare networks are reconstructed based on the following criteria:
Only direct transfers of patients who are discharged from a hospital and sent to another in another jurisdiction (“transfer”) or those who are discharged from one medical unit and move to another in the same hospital jurisdiction (“mutation”) are included. The hospital discharge summaries reflected the overall hospital stay of patients and a single diagnosis made them eligible without specification if it occurred during admission or at discharge. Patients who are discharged to their residence or deceased in a hospital are excluded. Patients hospitalized in non-continental European departments are also excluded.
First, the networks of hospitals and healthcare centers are re-built in silico using patient transfer data to model the potential movement of patients with HAIs from one hospital to another. In the PMSI database, each patient discharge summary contains information on the hospital facility of stay. Each hospital facility is identified by its unique FINESS number (Fichier National des Etablissements Sanitaires et Sociaux).[44] For this study, the FINESS and stay number of each patient discharge summary are used to merge two PMSI databases: one for acute-care and one for long-term care hospitals. Each patient stay is also numbered by order of stay across different hospitals. To create the logical sequence of patient movement, we sort each discharge summary: by patient ID and patient stay number for all observations.
The adjacency matrix [21], a graph of N nodes and E edges can be described by its’ N × N adjacency matrix A defined as:
Aij={=1ifiandjareconnected=0otherwise
In our patient transfer network, nodes (N) are defined as hospitals and edges (E) as the patient trajectories that connect hospitals. We computed origin i and target j hospitals for each patient stay by assessing if for each discharge the patient entered the hospital i as a transfer or mutation and left hospital i as a transfer or mutation. The same is computed for each j hospital. Using the iGraph package for R statistical software, we create the adjacency matrix of all i and j hospitals, including i and j if i did not transfer out any patients but received them and vice versa for j.[45]
We also compute the number of patients moving between hospitals i and j, as wij. The sum of the edge weights of the adjacent edges, the weight strength, is given by:
siw=∑j∈Γ(i)wij
in which Γ(i) is the set of neighbor hospitals of i.[21] Edge weights represent the number of patients within the trajectories between two healthcare facilities.
To identify the most important hospitals of a network, a series of centrality measures are calculated. The degree of a hospital, k, is the number of hospitals one hospital is connected to through its patient trajectories [21] defined as:
ki=∑jAij
The average degree of a network[21] is given by:
〈k〉=1N∑iki=2EN
In addition, Aij is a directed graph in which the directionality of patient transfers from one hospital to another is taken into account. Consequently, we can calculate the indegree (deg-) and outdegree (deg+) of any given node in which the degree sum formula is given by:
∑n∈Ndeg+(n)=∑n∈Ndeg−(n)=|E|
Betweenness centrality measures the importance of hospital acting as an intermediary between other hospitals defined as:
g(i)=∑s≠tσst(i)σst
Where betweenness centrality g(i) is equal to the sum of the σst the number of shortest paths going from s to t through hospital i measuring the importance of hospital i to the organization of flow in the network.[21] The same measure is calculated for patient trajectories defined as:
g(e)=∑e∈Eσst(e)σst
where edge betweenness centrality g(e) is equal to the sum of the σst the number of shortest paths going from s to t through edge e measuring the importance of edge e to the organization of flow in the network.[21]
Two community detection algorithms are used to assess community clustering for each network, which both take into account weighted graphs.[45] A common measure of the quality of partitions of a network into communities of densely connected nodes is modularity. Modularity is a scalar value between the vales of -1 and 1 that measures the density of links inside communities compared to links between them.[21, 22] The modularity and different communities of our network are defined using a community detection algorithm. The Greedy algorithm developed by Clauset et al.[19] optimizes modularity as the algorithm relies on network formation and as a result, computes a smaller range of communities as modularity approaches 1; however, the Greedy algorithm does not take into account edge directionality and we detect communities for undirected graphs of the healthcare networks. On the other hand, the Map equation algorithm developed by Rosvall et al. detects communities based on patterns of flow and takes into account edge directionality and the directed graphs are assessed.[20] This algorithm detects communities based on network structure and how it influences the system’s behavior.
Based on the community partitioning for each network, the mean geographic distance between hospitals of the same community is measured. To geo-localize hospitals, we used public government data on French hospital facilities and postal code addresses (https://www.data.gouv.fr/). Using an online batch geocoding server (http://www.findlatitudeandlongitude.com/), the hospitals’ addresses were converted to latitude and longitude coordinates. A distance matrix was calculated using the haversine formula to measure great-circle distances between all hospitals.[46]
Two intercommunity matrices were developed to assess patient sharing between different communities 1) Greedy algorithm-based communities 2) Map Equation-based communities. Based on the algorithm, each hospital node is assigned a community number. A matrix summing the individual hospitals transfers for hospitals that share the same community is constructed and converted into a directed graph. In addition, the mean latitude and longitude are calculated for each community from individual geocodes of the member hospitals. For the Map Equation intercommunity network, the Greedy algorithm is applied to identify the number of communities present when modularity is maximized.
Hospitals were ranked by their degree, betweenness, and closeness centrality measures for each network. When the centrality measures were equal, we replaced the rankings by the mean rankings. We tested the differences between rankings on an increasing subset of shared hospitals with the Wilcoxon rank sum test. The test was conducted as follows: starting with the highest ranked hospital in the general network, adding the next ranked general network hospital, and testing for significant differences between the general network rank and either the HAI-specific or suspected-HAI network rank of the same hospital until we compared all shared hospitals. As a result, we determined the thresholds where hospital rankings across the networks start to significantly differ which was defined as Wilcoxon rank sum test p-values under the 5% alpha risk.
To compare the networks between each other, we built 1000 random patients networks from the general network. We selected the same number of patients as either the HAI (21276 patients) or suspected HAI networks (394859 patients) from the general patient network at random and reconstructed these networks using their hospital discharge summaries. We calculated various network measures and the proportion of random patient networks that had values greater than, equal to, or less than the general patient network and the respective patient-specific networks.
|
10.1371/journal.pbio.1001935 | The Genomic Landscape of Compensatory Evolution | Adaptive evolution is generally assumed to progress through the accumulation of beneficial mutations. However, as deleterious mutations are common in natural populations, they generate a strong selection pressure to mitigate their detrimental effects through compensatory genetic changes. This process can potentially influence directions of adaptive evolution by enabling evolutionary routes that are otherwise inaccessible. Therefore, the extent to which compensatory mutations shape genomic evolution is of central importance. Here, we studied the capacity of the baker's yeast genome to compensate the complete loss of genes during evolution, and explored the long-term consequences of this process. We initiated laboratory evolutionary experiments with over 180 haploid baker's yeast genotypes, all of which initially displayed slow growth owing to the deletion of a single gene. Compensatory evolution following gene loss was rapid and pervasive: 68% of the genotypes reached near wild-type fitness through accumulation of adaptive mutations elsewhere in the genome. As compensatory mutations have associated fitness costs, genotypes with especially low fitnesses were more likely to be subjects of compensatory evolution. Genomic analysis revealed that as compensatory mutations were generally specific to the functional defect incurred, convergent evolution at the molecular level was extremely rare. Moreover, the majority of the gene expression changes due to gene deletion remained unrestored. Accordingly, compensatory evolution promoted genomic divergence of parallel evolving populations. However, these different evolutionary outcomes are not phenotypically equivalent, as they generated diverse growth phenotypes across environments. Taken together, these results indicate that gene loss initiates adaptive genomic changes that rapidly restores fitness, but this process has substantial pleiotropic effects on cellular physiology and evolvability upon environmental change. Our work also implies that gene content variation across species could be partly due to the action of compensatory evolution rather than the passive loss of genes.
| While core cellular processes are generally conserved during evolution, the constituent genes differ somewhat between related species with similar lifestyles. Why should this be so? In this work, we propose that gene loss may initially be deleterious, but organisms can recover fitness by the accumulation of compensatory mutations elsewhere in the genome. To investigate this process in the laboratory, we investigated 180 haploid yeast strains, each of which initially displayed slow growth owing to the deletion of a single gene. Laboratory evolutionary experiments revealed that defects in a broad range of molecular processes can readily be compensated during evolution. Genomic analyses and functional assays demonstrated that compensatory evolution generates hidden genetic and physiological variation across parallel evolving lines, which can be revealed when the environment changes. Strikingly, despite nearly full recovery of fitness, the wild-type genomic expression pattern is generally not restored. Based on these results, we argue that genomes undergo major changes not simply to adapt to external conditions but also to compensate for previously accumulated deleterious mutations.
| Deleterious, but non-lethal mutations are constantly generated and can hitchhike with adaptive mutations [1]. Consequently, such deleterious alleles are widespread in eukaryotic populations [2],[3]. For example, as high as 12% of the coding SNPs in yeast populations are deleterious [2]. Many of the observed functional variation in this species yield proteins with compromised or no activities [2], or lead to complete loss of genes with significant contribution to fitness (Text S1). Deleterious loss-of-function variants may occasionally revert to wild type, eventually perish from the population, or become compensated by mutations elsewhere in the genome. The third possibility, termed compensatory evolution, is the focus of our study. Theoretical works suggest that mutant subpopulations can cross fitness valleys by the simultaneous fixation of a compensatory mutation in the population [4],[5]. This process can also work in large populations and is facilitated by linkage of the two alleles [5].
Compensatory evolution appears to be common at many levels of molecular interactions. It is involved in the maintenance of RNA and protein secondary structures, it mitigates the costs of antibiotic resistance [6],[7], and allows rapid fitness recovery in populations with accumulated deleterious mutation loads [7]–[9]. Compensatory regulatory mutations also act to stabilize gene expression levels across species [10],[11], and conserve DNA-encoded nucleosome organization [12]. The most detailed experimental analyses on compensatory mutations for fixed deleterious mutations were performed in DNA bacteriophages [8],[13]–[15], bacteria [16],[17], and Caenorhabditis elegans [7],[9]. Three major patterns emerged from these studies. As the target size for compensatory mutations is typically much larger than that for reversion, compensation is more likely than reversion of deleterious mutations [13]. The rate of compensatory evolution increased with the severity of the deleterious fitness effects, and was not limited to functionally interacting partners of the mutated gene [15].
As regards the potential pleiotropic effects of compensatory evolution, our knowledge is rather limited, not least because it demands detailed exploration of the underlying molecular mechanisms of compensation. Compensatory mutations may enhance fitness either by reducing the need for the gene with the compromised function, or by restoring the efficiency of the affected molecular function [18]. For compensation of fitness costs of antibiotic resistance conferring mutations, restoration of function was the most common mechanism [18], but in other systems the relative importance of functional substitution and restoration is unknown. In the case of functional restoration (e.g., by enhanced dosage of a redundant duplicate of the disrupted gene), one might expect limited pleiotropic fitness effects of compensatory mutations across environmental conditions.
Compensatory evolution following gene loss is of special interest [17]. Gene loss may be initiated by genetic drift and/or selection through antagonistic pleiotropy [17],[19]. As reversion to the wild-type state is less likely, gene loss may promote genetic changes that drive the populations to new adaptive peaks (Figure 1). It's reasonable to assume that compensatory mutations are generally specific to the gene defect, and multiple molecular mechanisms can restore fitness. Therefore, independently evolving populations carrying an inactivated gene are expected to diverge from each other. Moreover, if compensation mainly proceeds by reducing the need for the disrupted molecular function then compensatory evolution could have a large impact on cellular physiology and survival upon environmental change. Accordingly, the beneficial effects of compensatory mutations may frequently be conditional, and subsequent changes to the environment can reveal the hidden genetic variation across populations (Figure 1). The goal of the current study was to test this hypothesis by an integrated systems biology approach. Specifically, we aimed to determine the potential of the Saccharomyces cerevisiae genome to compensate for gene loss through compensatory evolution and to explore the long-term consequences of this process.
We initiated laboratory evolutionary experiments with 187 haploid single gene knock-out mutant strains, all of which initially showed slow (but non-zero) growth compared to the wild-type control in a standard laboratory medium (Figure 2A, for selection criteria, see Materials and Methods). These genes cover a wide range of molecular processes and functions (Table S1). Populations were cultivated in parallel (four replicate populations for each null mutation), resulting in 748 independently evolving lines. 0.5% of each culture was diluted into fresh medium every 48 hours, and populations were propagated for approximately 400 generations. To control for potential adaptation unrelated to compensatory evolution, we also established 22 populations starting from the isogenic wild-type genotype, referred to as evolving wild types. Next, all starting and evolved populations were subjected to high-throughput fitness measurements by monitoring growth rates in liquid cultures.
Fitness may increase during the course of laboratory evolution as a result of general adaptation to the environment and/or accumulation of compensatory mutations that suppress the deleterious effects of gene inactivation. Under the assumption that compensatory evolution was the dominant force in our experiments, fitness should not increase by the same extent in all lineages: genotypes that carry deleterious null mutations are further away from the optimal state and are hence expected to show large fitness gains (Figure 2A); this was indeed so. On average, the evolving wild-type control populations showed a small, but significant 5% fitness improvement. By contrast, the fitness of populations carrying a deleterious null mutation improved by 23% on average (Figure 2B), and many of them approximated wild-type fitness (Figure 2C; Table S1). On the basis of fitness measurements at multiple time points during laboratory evolution (see Methods), we also report that individual fitness trajectories often showed a saturating trend during the course of laboratory evolution (Figure S1).
The difference in fitness improvement is not due to the elevated mutation rate of mutant genotypes for two reasons. First, a previous study conducted a genome-wide screen with the aim to identify genes in S. cerevisiae that influence the rate of mutations [20]. While a large number of such genes have been found, only four of them were present in our gene set (Δrad54, Δrad52, Δmre11, and Δrad50). Second, fitness improvements of the corresponding single gene knock-out strains did not differ from the rest of the dataset (one-tailed Wilcoxon rank sum test, p = 0.89).
As previously [16], we defined compensatory evolution as a fitness increase that is disproportionally large relative to that in the evolving wild-type lines. Using this definition, 68% of the genotypes showed evidence of compensatory evolution (i.e., at least one of the four independently evolving populations fulfilled the above criteria). The corresponding genes cover a wide range of molecular and cellular processes (Table S1).
Next, we compared the fitness improvements between evolved lines founded from the same gene deletion genotype versus those founded from different genotypes. This analysis revealed that not all genes were equally likely to be compensated as fitness gain differed significantly across genotypes (ANOVA, F(186) = 3.9, p<10−14) (see also Figure S2). It has been previously suggested that as mutations with especially large fitness effects tend to disrupt a broader range of molecular processes [21], such mutations may influence the number of mutational targets where compensatory evolution can occur [13]. We compiled three datasets that estimate different aspects of gene pleiotropy [22], including fitness under diverse environmental conditions (environmental pleiotropy), the number of protein-protein interactions (network pleiotropy), and the number of biological processes associated with a gene (multifunctionality). The extent of evolutionary compensation did not depend on any of the above mentioned features (Figure 2D). However, consistent with results of prior small-scale bacterial and viral evolutionary studies [13],[16], null mutations with more severe defects were more likely to be compensated (Figure 2E). This pattern probably reflects that the availability of compensatory mutations across the genome strongly depends on the fitness effect of the deleted gene. We provide a simple explanation of this phenomenon in the Discussion.
To investigate the genomic changes underlying compensatory evolution, we re-sequenced the complete genomes of 41 independently evolved lines and the 14 corresponding ancestors, all of which showed large fitness improvements (Table S1). We focused on de novo mutations that accumulated during the course of laboratory evolution. Large-scale duplications (including segmental or whole chromosome duplication) were observed in 22% of the laboratory evolved lines. On average, six point mutations and 0.5 small insertions or deletions per clone were detected (Figure 3A; Table S2). The ratio of non-synonymous to synonymous mutations was significantly higher than expected by chance (p = 0.003, see Materials and Methods), indicating that the accumulation of these mutations was driven by adaptive evolution. On average, pairs of evolutionary lines founded from the same genotype shared 5.3% of their mutated genes, while the same figure was 0.1% for lines founded from different genotypes (Table S2). This result is in contrast to results of a prior bacterial study [23], where a strong signature of parallel evolution emerged at the gene level across parallel evolving laboratory populations. Despite the rarity of parallel evolution at the molecular level, a major unifying trend emerged: evolution preferentially affected genes that are functionally related to that of the disrupted gene (Figure 3B). Moreover, when the null mutation affected a protein complex subunit, another subunit of the same complex was mutated 10 times more often than expected by chance (Figure 3B). Taken together, these results indicate that deletion of any single gene drives adaptive genetic changes specific to the functional defect incurred.
Although duplicated genes with partially overlapping function are frequent in the yeast genome, we found no evidence that genetic changes affecting a duplicate of the disrupted gene provide a general mechanism of compensation in our evolved lines. First, our dataset contains 128 genes showing evidence for compensation, and only 25% of these genes have a duplicate in the yeast genome (i.e., at least 30% amino acid similarity between the two copies). This figure is a gross overestimate, as it includes very distant duplicates that most likely diverged functionally (Materials and Methods). Second, the subset of genes with a gene duplicate were not more likely to be compensated during laboratory evolution than the rest of the dataset (Chi-squared test, p = 0.54). Third, genome sequence analysis of the evolved lines revealed only one clear example where evolution proceeded through increasing the dosage of a gene duplicate with redundant function of the deleted gene (Figure 3C). All three studied evolved lines of Δrpl6b showed an increased copy number of the left arm of Chromosome XIII (Figure 3C). RPL6B is a non-essential gene and encodes a ribosomal 60S subunit protein L6B. The duplicated genomic regions of Δrpl6b evolved lines carry RPL6A, a duplicate copy of RPL6B. The two genes share 94% amino acid identity, have highly overlapping functions, and deletion of both genes confer a synthetic lethal phenotype [24]. On the basis of these observations, we propose that doubling the copy number of RPL6A through segmental duplication could be partly responsible for the improved fitness in the evolved lines carrying the RPL6B deletion. The hypothesis was tested by increasing the copy number of RPL6A in wild-type and Δrpl6b genetic backgrounds, respectively. As expected, an enhanced copy number of RPL6A substantially improved the fitness of Δrpl6b, but not that of the wild type (Figure 3D).
Compensatory evolution may restore wild-type physiology or generate novel alterations with respect to prior physiological states [25]. To investigate the relative contribution of these processes, eight genotypes carrying a deleterious gene deletion and one corresponding evolved line were selected for transcriptome analysis (see Materials and Methods for selection criteria). Using DNA microarrays, the global gene expression states were compared between the wild-type, the ancestral line, and the evolved lines carrying the same gene deletion (Figure 4A and 4B). As expected from prior studies [26], inactivation of genes with high fitness contribution altered the expression of a large number of genes across the genome (ranging between 81 to 588) (see Table S3). Next, the transcriptomic profiles were compared by calculating all pairwise combinations of Euclidean distances. The wild-type, the ancestral line, and the corresponding evolved lines generally showed substantial differences in their transcriptome profiles (Figure 4B), indicating that compensatory evolution drives the cell towards novel genomic expression states. Importantly, transcriptome profile distances between different genotypes was always higher than distances between replicate measurements of the same genotype (Figure 4B), implying that the substantial differences observed between evolved lines and wild-type cannot be attributed to measurement noise. As a further support, typically only 10%–30% of the genes with altered expression in the ancestral lines showed significant shift towards the wild-type expression level in the corresponding evolved lines (Figure 4C). Hence, despite substantial fitness improvements (>75% for all cases investigated), the majority of the gene expression changes due to gene deletion remained unrestored during evolution. These patterns were not attributable to growth rate regulated gene expression or copy number variation in the evolved lines (Figure S3).
Taken together, compensatory evolution following gene loss did not restore wild-type genomic expression and promoted genomic divergence across populations. Are these evolutionary outcomes phenotypically completely equivalent? This problem was first addressed by monitoring the fitness of 237 evolved populations in 14 environmental settings, including previously tested nutrients and stress factors [27]. Prior to evolution, genotypes carrying a gene deletion generally displayed slow growth in most environments (Table S1). The situation was far more complex following laboratory evolution. Considering all possible pairs of population-environment combinations, fitness improved in 52%, and declined in 8% of the cases (Figure 5A). Moreover, independently evolved populations carrying the same disrupted gene showed more fitness variation across the 14 tested conditions than in the environment they had been exposed to during laboratory evolution (Figure 5B, p<10−7), while evolved wild-type populations did not show such a difference (p = 0.93, coefficient of variations compared by Z-test). Furthermore, the degree of fitness variation across conditions was especially high for gene deletions that showed large fitness gains during compensatory evolution (Spearman rho = 0.36, p = 10−4) (Figure 5C). These results indicate that the level of discernible heterogeneity in fitness was relatively low in the evolved populations founded from the same genotype, but the variation can be uncovered upon environmental change.
Finally, our analysis revealed a few instances where the laboratory evolved lines displayed significantly higher than wild-type fitness in specific environments (Table S1). Most notably, the evolved Δrpl6b and Δatp11 lines displayed 24%–26% fitness increase compared to that of the wild type in a medium containing sodium chloride (Table S1), a result that was confirmed by additional independent colony size assays with high replicate number (n = 20, Wilcoxon rank-sum test p<10−4 in all cases). Moreover, the fitnesses of these lines in this medium surpassed all that of the 22 evolved wild-type controls. These results are all the more remarkable, as the corresponding ancestral Δrpl6b and Δatp11 strains showed fitness values significantly lower than wild type under all environmental conditions considered. These preliminary results indicate that gene loss can promote adaptive evolution towards novel environments, a possibility that will be explored further in a future work.
Next, we conducted an in-depth genetic analysis with the MDM34 deletion with the aim of deciphering the molecular mechanisms and/or potential fitness costs of compensatory mutations (Text S1). This gene codes for a component of the ERMES protein complex, and is involved in the exchange of phospholipids between mitochondria and the endoplasmatic reticulum (Figure 6A). Disruption of this gene yields impaired cardiolipin synthesis [28], as an insufficient amount of unsaturated fatty acids reaches the mitochondria (Figure 6A). Laboratory-evolved lines carrying deletion in this gene substantially improved fitness in the medium of selection (Table S1), but the putative cellular mechanisms of compensation were remarkably different across populations (Figures 6A and S4). The native copy of MDM34 was reinserted into the ancestral line and four evolved lines carrying the same deletion (Δmdm34). The analysis revealed that the net effect of mutations in three evolved lines were deleterious in the presence of MDM34 (Figure 6B). Next, we concentrated on a specific mutation observed in MGA2, a gene involved in the regulation of unsaturated fatty acid biosynthesis (Figure 6A; Text S1). Inserting the observed mutations (mga2-1) into wild type and Δmdm34 resulted in very similar conclusions. mga2-1 and Δmdm34 showed strong sign-epistasis [29]: they were independently deleterious but significantly less so when they occurred together (Figure 6C). Moreover, the capacity of mga2-1 to compensate the loss of MDM34 was restricted to non-acidic conditions (Figure 6C), probably because of the misregulation of the corresponding stress-induced pathway under low pH (Text S1).
Our dataset contains 21 independent point mutations that occurred during laboratory evolution and generated in-frame stop codons. Most notably, a mutation in WHI2 emerged in an evolving Δrpb9 line, which shortened the coding region from 480 to 133 codons, and hence most likely resulted in a non-functional protein.
To test the impact of loss of WHI2 function on fitness and compensation, Δwhi2 was introduced into Δrpb9 cells using synthetic genetic array methodology (Figure 7A and 7B) [30]. In agreement with expectation, deletion of WHI2 partly suppressed the harmful effect of the RPB9 deletion (Figure 7B). RPB9 is an RNA polymerase II subunit, and its deletion leads to elevated transcriptional error rate [31] and in turn, to proteotoxic stress [32], which can result in cell cycle arrest [33]. WHI2 is known to be required for general stress response [34] and cell cycle arrest [35]. We speculate that less stringent cell cycle control due to WHI2 deletion is favorable in Δrpb9 (see also [36]).
Next, the fitness impact of WHI2 deletion was evaluated across 14 environments. The fitnesses of the Δrpb9 Δwhi2 strain varied strongly across conditions, and showed correlation with that of the evolved Δrpb9 line, which carried the WHI2 non-sense mutations (Spearman rho = 0.77, p<0.005) (see Figure 7C). Most notably, the compensation of Δrpb9 by Δwhi2 was completely abolished in the presence of cycloheximide (Figure 7B). We conclude that the compensatory effect of WHI2 deletion is plastic across environments.
Our work addresses one of the most long-standing debates in evolution. Since the early 1920s, Ronald Fisher pioneered the view that adaptation is by and large a hill climbing process: it proceeds through progressive accumulation of beneficial mutations [37],[38]. However, as slightly deleterious mutations are far more abundant, they have a significant contribution to genetic variation in natural populations [2]. In the long run, the wealth of such detrimental mutations is expected to promote fixation of compensatory mutations elsewhere in the genome. This work focused on a specific aspect of this problem, and asked whether deleterious gene loss events promote adaptive genetic changes and what the side consequences of such a process might be. To systematically study compensatory evolution following gene loss, we initiated laboratory evolutionary experiments with over 180 haploid yeast genotypes, all of which initially displayed slow growth owing to the deletion of a single gene, and investigated the genomic and phenotypic capacities of the evolved lines in detail. Thanks to the exceptionally large-scale analysis of our study, the following major conclusions can be drawn.
First, compensatory evolution following gene loss was pervasive: 68% of the deleterious, but non-lethal gene disruptions were compensated through the accumulation of adaptive mutations elsewhere in the genome (Figure 2B). Furthermore, in agreement with prior bacterial studies [16],[17], the process was strikingly rapid. As the set of disrupted genes are functionally very diverse (Table S1), it appears that defects in a broad range of molecular processes can readily be compensated during evolution.However, we and others [17] also found that not all genotypes are equally likely to be recovered during laboratory evolution. Therefore, future works should clarify the exact molecular, functional, and systems level gene properties that influence compensability. Second, our large-scale study indicates that the extent of fitness loss due to gene disruption is one if not the strongest predictor of compensatory evolution (Figure 2E). Although this relationship has been observed previously in small-scale studies [16], the reasons remained largely unknown. One may argue that the spread of compensatory mutations with mild beneficial effects would have taken many more than 400 generations to reach fixation [16]. Although this explanation cannot be excluded, there is another intriguing possibility [13]. Consistent with Fisher's geometric model [37],[38], fitness improvement in populations close to an optimal state can only be achieved by relatively rare mutations with small effects. However, when a population with a gene defect is further away from a fitness peak, compensatory evolution may proceed through a wider range of mutations, including ones that have deleterious side effects. Two lines of evidence are consistent with this scenario. Compensatory evolution has associated pleiotropic effects (Figures 5 and 6C). Moreover, the theory predicts that compensatory mutations should be especially frequent in the case of strongly deleterious null mutations. An analysis based on data of a prior genome-wide genetic interaction study [21] suggests that it may indeed be so (Figure 8).
Third, genomic analysis of the evolved lines revealed that deletion of any single gene drives adaptive genetic changes specific to the functional defect incurred (Figure 3B), and consequently convergent evolution at the molecular level was extremely rare. In agreement with a prior bacterial evolutionary study [17], we found that gene duplication has only a minor role during compensatory evolution following gene loss. A more general issue is the extent to which mutations that affect gene expression could alone recover fitness [17],[39]. Although genetic changes in putative promoter regions were not overrepresented in our dataset (Binomial test, p = 0.87), 21 observed point mutations generated in-frame stop codons, most likely yielding proteins with compromised or no activities (see also Figure 7). These results indicate that fitness recovery following gene loss can partly be achieved purely through inactivation of other genes.
Fourth, compensatory evolution promoted divergence of genomic diversification, and shifted the evolved population towards novel genomic expression states (Figure 4B). Despite substantial fitness improvements, the majority of the gene expression changes due to gene deletion remained unrestored during evolution. This finding is consistent with prior works arguing that no clear relationship exists between the change in mRNA expression of a gene and its requirement for growth in the same condition [40].
Fifth, independently evolved populations showed substantial fitness variation across environments that they had not been exposed to during laboratory evolution (Figure 5). These results suggest that accumulation of adaptive mutations during compensatory evolution generated substantial genetic differences between populations, and this variation can be uncovered upon environmental change.
Taken together, several lines of evidence indicate that fitness gains in the evolved lines reflect accumulation of gene specific compensatory mutations rather than a global adaptation: (i) evolving wild-type control populations showed only minor changes in fitness, (ii) the rate of adaptation was genotype specific, (ii) convergence at the molecular across genotypes was extremely rare, (iv) evolution preferentially affected genes that are functionally related to that of the disrupted gene, and (v) compensatory mutations had no beneficial impact in a wild-type genetic background.
The above results encouraged us to distinguish between two evolutionary scenarios. Organisms may attempt to restore the disrupted molecular function through mutations in genes with redundant functions (functional restoration). Alternatively, they may aim to minimize the cellular damage incurred by gene disruption (functional replacement). While the possibility of full functional restoration cannot be excluded, the rarity of compensation through mutations in gene duplicates and the plasticity of compensatory mutational effects across environments are consistent with the second scenario. Indeed, our work demonstrates that gene loss promotes genetic changes that have a large impact on evolutionary diversification, genomic expression, and viability upon environmental change. An important implication of our study is that the beneficial effects of compensatory mutations should frequently be conditional, and subsequent changes to the environment can reveal the hidden fitness effects (beneficial and detrimental alike). Lack of restoration of fitness across environments is broadly consistent with the emerging view that epistatic interactions are plastic across conditions [41],[42].
The perspective offered in this work leads to the re-formulation of several fundamental questions. First, it sheds light on an evolutionary paradox: while core cellular processes are generally conserved during evolution [43], the constituent genes are partly different across related species with similar lifestyles. We propose that gene content variation across species is partly due to the action of compensatory evolution and may not need to reflect changes in environmental conditions and the consequent passive loss of genes. Although the exact population genetic conditions facilitating this process remain to be elucidated, several observations are consistent with this view. Most notably, the phylogenetic conservation of indispensable genes depends on how easily the gene can be functionally replaced through enhanced expression of other genes [44]. Second, it has been suggested that deleterious mutations may act as stepping stones in adaptive evolution by providing access to fitness peaks that are not otherwise accessible [45],[46]. Indeed, our analysis revealed a few instances where the laboratory evolved lines displayed significantly higher than wild-type fitness in specific environments. Finally, given the prevalence of gene loss events during tumorigenesis, future work should elucidate whether similar processes drive the somatic evolution of cancer [47].
All strains used in this study were derived from the BY4741 S. cerevisiae parental strain. Non-essential single-gene deletion strains from the haploid yeast deletion collection [40] (MATa; his3Δ 1; leu2Δ 0; met15Δ 0; ura3Δ 0; xxx::KanMX4) were used to systematically identify all gene disruptions with a significant growth defect. Slow-growing mutants were identified in two steps. An earlier study identified 671 gene deletants in diploid background, which showed a significant fitness defect on both rich and synthetic media [48]. We thus measured fitness of the corresponding MATa haploid strains by recording their growth curves in liquid media. We identified 187 deletants showing at least 10% growth rate defect, which constituted the set of ancestral strains subjected to laboratory evolution (for details of growth measurements see below).
The slow-growing yeast deletants used in this study are listed in Table S1. The evolutionary experiment was conducted using rich liquid medium (YPD, 1% yeast extract, 2% peptone, 2% glucose). Solid media were prepared using 2% agar, which were found to be optimal for reproducible colony size measurement. Details on the media used in the phenotypic profiling experiment can be found in Table S4. Oleic acid and stearic acid was dissolved in DMSO as a 100 mM stock and added to the medium after autoclaving to a final concentration of 0.1 mM.
Compensatory adaptation refers to fitness gains in a gene deletion strain that are greater than fitness gains occurring in an isogenic wild-type strain. We conducted a series of laboratory evolutionary experiments using four independent populations of each of the 187 slow-growing deletants along with 22 independent lineages of an isogenic wild-type strain (referred to as evolving wild types). The YOR202W deletion strain was used as evolving wild-type control because the fitness of this strain is indistinguishable from the BY4741 parental wild-type strain [19]. Moreover, this strain carries the KanMX4 cassette in the nonfunctional his3Δ1 allele, thus it was possible to control for the reported mutation-generating effect of the KanMX4 cassette [36]. All strains were inoculated into randomly selected positions of 96-well plates. Four wells in different positions were not inoculated by cells to help plate identification and orientation. Cells were grown in standard laboratory rich media to minimize selection pressure originating from nutrient limitation. The presence of the KanXM4 cassette was not selected for during the evolutionary experiment, since G418 was omitted from the medium for two reasons. First, using G418 at 200 mg/l concentration decreases the growth rate of the unevolved wild-type control strain (unpublished data) and might lead to selection for increased resistance. Second, the usage of the drug at a growth-limiting concentration may induce mutagenesis through environmental stress response. To provide optimal growth conditions, plates were covered with sandwich cover (Enzyscreeen BV), shaken at 350 rpm, and incubated at 30°C. Using a handheld replicator, ∼105 cells (∼0.5 µl sample volume) were transferred every second day to 100 µl of fresh medium in 96-well plates resulting in ∼7.6 generations between transfers. The experiment was run for 104 days (∼400 generations total) and samples from days 0, 26, 52, 78, and 104 were frozen in 15% glycerol and kept at −80°C until fitness measurement. Cross-contamination events were regularly checked by PCR and visual inspection of empty wells (unpublished data).
We used established protocols specifically designed to measure fitness in yeast populations [49]. Growth was assayed by monitoring the optical density (OD600) of liquid cultures of each strain using 384-well microtiter plates containing YPD medium (as during the evolutionary experiments). We used relative growth rate as a proxy for relative fitness (see below). Compared to laborious competition based fitness assays, this protocol allows estimating growth rate on a relatively large scale in an environment that is nearly identical to the one used in the evolutionary experiments.
Starter cultures were inoculated from frozen samples using 96-well plates. The starter plates were grown for 48 hours under identical conditions to the evolutionary experiment. 384-well plates filled with 60 µl rich medium per well were inoculated for growth curve recording from the starter plates using pintool with 1.58 mm floating pins. The pintool was moved by a Microlab Starlet liquid handling workstation (Hamilton Bonaduz AG) to provide uniform inoculum across all samples. The median blank corrected initial OD600 of the wells was 0.027. Each 384-well plate were inoculated with four different starter plates: one plate having the unevolved wild-type control as a reference strain in all wells in order to estimate various within-plate measurement biases, and three plates containing the same set of mutants from three of the five time points of the evolutionary experiment. The 384-well plates were incubated at 30°C in an STX44 (LiCONiC AG) automated incubator with alternating shaking speed every minute between 1,000 rpm and 1,200 rpm. Plates were transferred by a Microlab Swap 420 robotic arm (Hamilton Bonaduz AG) to Powerwave XS2 plate readers (BioTek Instruments Inc) every 20 minutes and cell growth was followed by recording the optical density at 600 nm. Six technical replicate measurements were executed on all strains sampled from each time-point of the evolutionary experiment. Measurements with growth curve irregularities were automatically removed. Only those strains were further analyzed where at least four technical replicate measurements remained after this quality control step.
Growth rate was calculated from the obtained growth curves following an established procedure [49],[50]. To eliminate potential within-plate effects that might cause measurement bias, growth rates were normalized by the growth rate of neighboring reference wells that contained the wild-type controls. For each strain and each evolutionary time point, relative fitness was calculated as the median of the normalized growth rates of the technical replicates divided by the median growth rate of the wild-type controls. At day 0, the technical replicate measurements of the isogenic independently evolving lines were combined to calculate median ancestral fitness since by that time these populations had no independent evolutionary history. Stringent criteria were used to define the set of ancestor strains with substantial growth rate defect: a minimum of 10% fitness drop was required compared to the wild-type controls; significance was determined by one-tailed Wilcoxon rank sum test, p-value was corrected with a false discovery rate of 0.05.
To determine whether the fitness defect of a given knock-out strain became compensated during the evolutionary experiment two criteria must have been met: First, the growth rate improvement had to be significant (one-tailed Wilcoxon rank sum test, p-value corrected with a false discovery rate of 0.05). Second, the growth rate increment of the knock-out strain had to be disproportionally larger than that of the evolving wild-type control strains. To test whether fitness gain in a knockout is higher than those occurring in the evolving control lines, we first fitted a normal distribution to the fitness improvement values of the evolving control lines. Next, we defined a fitness improvement cutoff, so that the probability that an evolving control line would show an improvement at least that high is less than 0.05.
To evaluate the extent of evolutionary compensation, a relative compensation index was calculated according to the following formula:where WT and Δ means median normalized growth rate of the evolving wild-type control and the knock-out strain, respectively, measured before (start) and after (end) the evolutionary experiment. Thus, a relative compensation of 1 indicates that the knock-out strain reached the same fitness after evolution as the evolving wild-type control strains. See Table S1 for the whole dataset.
To study the pleiotropic effects of compensatory adaptation, we measured the fitnesses of 237 evolved lines carrying a single gene deletion, all evolved wild-type control lines along with the corresponding ancestors across various environmental conditions. As this experiment demands high-throughput analyses (over 14,000 data points), fitness was estimated by colony size on solid agar media. Moreover, it allowed direct comparison of the reliability of our measurements to results of a previous study (Figure S5).
We prepared solid agar media of 14 different compositions to expose the strains to fundamentally diverse environments and to obtain sufficient throughput. Our list of 14 growth media was primarily based on a previous study [27] and included various carbon sources and stress conditions (Table S4). A robotized replicating system was set up for colony size based fitness measurement. The system consists of a Microlab Starlet liquid handling workstation (Hamilton Bonaduz AG) equipped with a pintool with 768 pins (S&P Robotics Inc) and a custom-made pintool sterilization station. Several aspects of the replication procedure had been experimentally customized to achieve uniform, reproducible inoculation of yeast cells.
Fitness of the ancestor (day 0) and evolved strains (day 104) was approximated by measuring colony sizes of ordered arrays of strains at 768 density. First, four different 96-well plates of the evolutionary experiment were scaled up to arrays of 384 colonies: one having the unevolved wild-type control in all positions, and three different plates of the mutant set from the same time point. Then pairs of 384 arrays with corresponding strains from day 0 and 104 were combined to reach 768 density. With this set up, all evolving replicate lines derived from the same ancestral genotype from both day 0 and day 104 were grown on the same 768 plate to exclude potential plate-to-plate variations when comparing colony growth of ancestor and evolved lines. Four technical replicates of these 768 arrays were transferred into each of the 14 different media.
After acclimatization to the media at 30°C for 48 hours the plates were replicated again onto the same type of media and photographed after 48 hours of incubation at 30°C. Digital images were processed to calculate colony sizes, and potential systematic biases in colony growth were eliminated (Text S1). For each growth environment, fitness of each original knock-out genotype at day zero and each independently evolving line at day 104 was determined as the median of the size of replicate colonies. The reliability of our experimental setup and data processing was confirmed by comparing the fitness measurements of ancestral knock-out strains with the published data of Dudley and colleagues (Figure S5) [27].
To determine whether an ancestor genotype shows a significantly altered fitness compared to the wild-type control in a given environment, we used a Wilcoxon rank sum test (with p-value corrected for each condition with a false discovery rate of 0.05). The same statistical test was used to determine whether the fitness of an evolved line is different from that of its ancestor in a given environment. See result in Table S1.
To reveal the underlying molecular mechanisms of compensation, we subjected 41 strains to whole-genome re-sequencing. Our list of sequenced strains primarily included genotypes with large initial fitness defect, substantial fitness improvement and gradual fitness increase over the course of evolution. To be able to detect parallel evolution at the molecular level, we selected two to four independently evolving lines of each ancestor genotype for sequencing. Overall, 41 evolved lines from 14 deletion strains were chosen along with their corresponding ancestor strains. Candidates were re-streaked and single clones were isolated and their fitness increase was confirmed by growth curve recording.
Genomic DNA was prepared using a glass bead lysis protocol: clones were inoculated into 5 ml YPD+G418 (200 mg/l) and grown to saturation at 30°C. Cells were pelleted and resuspended in 500 µl of lyis buffer (1% SDS, 50 mM EDTA, 100 mM Tris [pH 8]). Cells were mechanically disrupted by vortexing for 3 minutes at high speed with 500 µl glass bead (500 µm, acid washed). After adding 275 µl 7 M ammonium acetate, samples were incubated at 65°C for 5 minutes, followed by a second incubation on ice for 5 minutes. The samples were extracted with chloroform∶isoamyl alcohol (24∶1) and centrifuged for 10 minutes. The aqueous layer was transferred into a new tube and precipitated with 1 ml isopropanol, pelleted and washed with 70% ethanol, and resuspended in 500 µl RNaseA solution (50 ng/ml). After 30 minutes RNaseA treatment at room temperature, samples were chloroform∶isoamyl alcohol (24∶1) extracted, precipitated with 50 µl sodium acetate (3 M [pH 5.2]) and 1,250 µl ethanol, pelleted and washed with 70% ethanol. Finally, the genomic DNA was dissolved in water. The steps of re-sequencing was done by the UD-GenoMed Medical Genomic Technologies Ltd: amplified genomic shotgun libraries were run on the Illumina HighScan SC with 1×100 bp single read module resulting in an average coverage of about 80×. Reads were aligned to the S. cerevisiae EF4 genome assembly using the BWA software package [51] having the genomic repeats masked using RepeatMasking [52]. Variant calling was performed using the GATK software package [53]. Genomic single-nucleotide polymorphisms with less than 200 phred-scaled quality score or lower than 0.3 mutant/reference ratio were ignored. Duplications of large chromosomal segments or whole chromosomes were identified as increased read coverage of certain regions. Elevated read coverage of regions with a minimum of 25 kb length were accepted as duplications if both the Control-FREEC [54] (Wilcoxon rank-sum test, p<0.01) and the CNV-seq [55] (p<0.0001) software predicted significant alteration from the read coverage of the reference genome.
Our primary aim was to analyze de novo mutational events. De novo mutations were identified as alterations from the reference genome specifically found in the evolved lines but not present in the ancestral strains. Mutations, which occurred before our evolutionary experiment but after the gene knock-out, are referred to as secondary ancestor mutations. These mutations were identified in the ancestral strains as SNPs and indels present only in the corresponding ancestor strain, not in any other ancestral strains. The rationale behind this consideration is not to classify mutations accumulated in the parental strain of the mutant library prior to the generation of the knock-out strain as a secondary ancestor mutation. The list of identified mutations can be found in Table S2.
Whole-genome re-sequencing revealed that 86% of SNPs in the coding regions were non-synonymous. To statistically test whether the ratio of non-synonymous to synonymous SNPs was higher than expected based on a neutral model of evolution, we employed the method of Barrick and colleagues [56]. Briefly, we took all different point mutations observed in protein coding regions and calculated the probability that 86% or more substitutions would result in a non-synonymous substitution if it occurred in a random coding position. The excess of non-synonymous substitution observed in the evolved genomes was significant (p = 0.003).
To test whether the extent of evolutionary compensation is influenced by the disrupted gene's pleiotropy, we used three complementary measures of gene pleiotropy. Environmental pleiotropy of a non-essential gene was defined as the number of unique conditions in which the removal of the gene resulted in a fitness defect according to Dudley and colleagues [27]. Network pleiotropy was measured as the total number of protein-protein interactions reported in the BioGRID database [57]. Finally, multifunctionality of a gene was calculated on the basis of a set of GO terms considered to be specific by yeast geneticists, as previously described [58].
To investigate whether mutations accumulated during compensatory evolution preferentially affected genes that are functionally related to the disrupted gene, we used different measures of functional relatedness: co-membership within stable protein complexes, shared functional category, genetic interaction profile similarity, co-expression, and paralogy. For protein complexes we used the manually curated dataset based on tandem affinity purification/mass spectrometry studies (YHTP2008) from the Wodak lab [59]. For functional categories, the MIPS Functional Catalogue Database was downloaded [60]. Genetic interaction profile similarities were obtained from a large-scale genetic interaction screen study [21]. The authors calculated the genetic interaction profile for a given gene deletion genotype as the list of genetic interaction scores detected across all other genes in their dataset. The genetic interaction profile similarity between two genes was defined as the Pearson correlation value of the two genetic interaction profiles [21]. For calculating co-expression data, 247 normalized microarray datasets from the M3D database [61] were used to create an expression profile for each gene. In case of multiple replicates per experiment, the average normalized values were calculated, and employed further. For each gene pair, co-expression value was calculated as the Pearson correlation coefficient between the two expression profiles.
Paralog gene pairs were identified by performing all-against-all BLASTP similarity searches of yeast open reading frames. We defined two genes as paralogs if (i) the BLAST score had an expected value E<10−8, (ii) alignment length exceeded 100 residues, (iii) sequence similarity was >30%, and (iv) they were not parts of transposons.
Eight evolved lines were selected for microarray analysis, all of them showing high fitness following evolution (at least 20% initial fitness defect compared to the wild-type control and at least 20% fitness improvement as a result of the evolutionary process). The corresponding ancestral strains and the wild-type control were also subjected to gene expression profiling. Table S3 contains the list of strains. Candidates were re-streaked and single clones were isolated and their fitness increase was confirmed by growth curve recording.
Two independent colonies of the wild-type control, evolved, and corresponding ancestor knock-out strains were inoculated into 15 ml YPD and grown overnight at 30°C. The saturated populations were diluted to an OD600 of 0.15 in 60 ml YPD and grown to early mid-log phase (OD600 0.6±0.05) in 250 ml Erlenmeyer flasks with 220 rpm shaking at 30°C. Cells were harvested by centrifugation (4,000 rpm, 3 min, 30°C) and immediately frozen in liquid nitrogen after removal of supernatant. Total RNA was prepared by hot acidic phenol extraction and cleaned up using the QIAGEN's RNAeasy kit.
All steps after RNA isolation were automated using robotic liquid handlers as described previously [62]. Dual-channel 70-mer oligonucleotide arrays were used with a common reference pool of wild-type RNA. Quality control, normalization, and dye-bias correction was performed as described earlier [62]. The reported fold change is the average of the four replicate mutant profiles versus the average of all wild-type controls. A total of 58 transcripts showed stochastic changes in wild-type profiles and were excluded from the analyses. Differentially expressed genes were defined as those showing a 1.7-fold abundance change and a p-value<0.05 when comparing two strains. The raw dataset is available online at ArrayExpress (http://www.ebi.ac.uk/arrayexpress/, accession number E-MTAB-2352).
All transcriptome comparisons of the wild-type, knockout, and evolved strains were repeated on a dataset where CNVs, genes showing expression response to aneuploidy, and growth rate related genes were excluded. CNVs were identified on the basis of the read coverage of the genome sequence data (Table S2) with the exception of one strain (Δrpl43a), which was not sequenced. In the case of Δrpl43a, whole chromosome duplication was predicted on the basis of visual inspection of expression profiles. The position of partial chromosome duplication was predicted by the Charm algorithm [63]. In evolved strains carrying aneuploid chromosomes, genes showing expression response to that particular aneuploidy were excluded from the transcriptome comparisons (data on the transcriptome effects of aneuploidy were obtained from [64]). Genes showing significant expression response to changes in growth rate were also excluded, as defined previously [65] on the basis of the growth rate measurements of Brauer and colleagues [66].
The evolved lines of Δmdm34 were chosen for in-depth genetic analysis. The fitness cost of the set of compensatory mutations accumulated in the evolved Δmdm34 lineages was measured in wild-type genetic background. To this end, the MDM34 gene was re-introduced into the ancestor and evolved Δmdm34 lineages according to the delitto perfetto method [67]. First, the KanMX4 cassette in the ancestor and evolved Δmdm34 lineages was swapped with the CORE-UH cassette, containing the KlURA3 and hyg markers. Then the MDM34 open reading frame with longer than 0.3 kb flanking regions on both sides was amplified from the unevolved wild-type control strain and transformed into the cells to replace the CORE-UH cassette. The replacement of the KlURA3 marker was counter-selected using 5-FOA containing medium. The loss of hygr was confirmed, the site and orientation of gene replacement was verified by PCR and the sequence of the MDM34 gene was determined by capillary sequencing.
In a second analysis, a point mutation identified in the MGA2 gene in one of the evolved Δmdm34 lineages was reinserted into both the wild-type and ancestor Δmdm34 background. This specific point mutation changes the 750th codon of MGA2 from GAT to TAT resulting in the incorporation of tyrosine instead of aspartic acid. We refer to the mutant allele as mga2-1. Using the delitto perfetto method [67], we introduced this point mutation into the unevolved wild-type control strain. First, the CORE-UH cassette was inserted into the genome at the desired position of the SNP. Then, two complementary oligonucleotides of 81 bp length with the sequence of the region of interest and the SNP in the 41st position were transformed. The replacement of the KlURA3 marker with the missense SNP was counter-selected using 5-FOA containing medium, loss of hygr was confirmed, and the result of the site-directed mutagenesis was verified by capillary sequencing. Attempts to introduce the mga2-1 mutation into the ancestor Δmdm34 strain in this way were not successful, presumably due to the severe slow growth of the intermediate strain that lacks both MDM34 and MGA2 gene in a functional form. To complement this, a helper plasmid with MDM34 gene (MoBY ORF Library [68]) was transformed into the cells prior to the site directed mutagenesis [69]. Because of the presence of the URA3 marker on the helper plasmid, the CORE-Hp53 cassette was used in this experiment. The steps of mutagenesis were similar as without the helper plasmid, which was removed by passaging cells through 5-FOA afterwards.
Yeast samples were grown in 20 ml YPD medium to mid-log phase (0.8 OD600 value). RNA was extracted from 107 yeast cells by acidic phenol method using TRI Reagent Protocol (Sigma-Aldrich Co). The RNA samples were concentrated by the NucleoSpin RNA Plant Kit (Macherey-Nagel), according to the manufacturer's instructions. A total of 500 ng RNA was used as a template to prepare cDNA using the Maxima First Strand cDNA Synthesis kit (Thermo Scientific). Reactions without template were set up to detect contaminations of the reagents used in the cDNA synthesis. qPCR reactions were set up in 20 µl volume, using the following templates: no template control, 10 ng non-transcribed RNA and cDNA transcribed from 10 ng RNA. The qPCR reactions were run in a Bioer LineK Gene device, using 2× Maxima SYBR Green qPCR Master Mix (Thermo Scientific). All samples had three technical replicates. Gene expression was determined in arbitrary units using a standard curve fitted on triplicates of a four-step 10-fold dilution series. OLE1 expression level was determined relative to TUB1 expression level as an internal control. All control reactions, not treated with reverse transcriptase or not having template, gave Ct values at least 10 cycles higher than the corresponding samples.
|
10.1371/journal.pgen.1003899 | A Reversible Histone H3 Acetylation Cooperates with Mismatch Repair and Replicative Polymerases in Maintaining Genome Stability | Mutations are a major driving force of evolution and genetic disease. In eukaryotes, mutations are produced in the chromatin environment, but the impact of chromatin on mutagenesis is poorly understood. Previous studies have determined that in yeast Saccharomyces cerevisiae, Rtt109-dependent acetylation of histone H3 on K56 is an abundant modification that is introduced in chromatin in S phase and removed by Hst3 and Hst4 in G2/M. We show here that the chromatin deacetylation on histone H3 K56 by Hst3 and Hst4 is required for the suppression of spontaneous gross chromosomal rearrangements, base substitutions, 1-bp insertions/deletions, and complex mutations. The rate of base substitutions in hst3Δ hst4Δ is similar to that in isogenic mismatch repair-deficient msh2Δ mutant. We also provide evidence that H3 K56 acetylation by Rtt109 is important for safeguarding DNA from small insertions/deletions and complex mutations. Furthermore, we reveal that both the deacetylation and acetylation on histone H3 K56 are involved in mutation avoidance mechanisms that cooperate with mismatch repair and the proofreading activities of replicative DNA polymerases in suppressing spontaneous mutagenesis. Our results suggest that cyclic acetylation and deacetylation of chromatin contribute to replication fidelity and play important roles in the protection of nuclear DNA from diverse spontaneous mutations.
| Mutations strongly predispose humans to cancer and many other diseases. Despite significant progress, we still do not fully understand the molecular mechanisms that protect us from mutations. Human DNA is part of a highly organized complex called chromatin. Chromatin regulates our development, metabolism, and behavior. Special enzymes modify chromatin by the addition and removal of chemical groups. Acetylation and deacetylation of chromatin have been conserved during evolution. The involvement of chromatin and its modifications in the protection of DNA from mutations is poorly understood. The yeast Saccharomyces cerevisiae is an excellent model for studying the connection between chromatin modifications and mutations. Using this model, we found that the deacetylation and acetylation of chromatin on histone H3 lysine 56 are required for preventing a wide range of spontaneous mutations. Future studies will determine whether acetylation and deacetylation of chromatin are involved in protecting DNA from mutations in human cells.
| Mutations are the prerequisites for evolution and the humoral immune response. However, mutations are often detrimental due to their ability to trigger both inherited and sporadic diseases. Base substitutions, 1-bp deletions, and 1-bp insertions are the most common mutations [1], [2]. Cells can also acquire gross chromosomal rearrangements (GCRs) [3], [4], complex mutations [5], and other genetic alterations [1], [2], [6]. Though GCRs are relatively rare mutational events, they profoundly reshape genetic information. Mutations arise as a result of replication errors, defects in DNA repair, spontaneous and induced DNA damage, and several error-prone processes including somatic hypermutagenesis, mitotic gene conversion, and break-induced replication [2], [6]–[12]. DNA replication errors produce a large fraction of spontaneous mutations [7].
The bulk of nuclear DNA is replicated by the leading-strand polymerase ε and lagging-strand polymerase δ that both possess intrinsic 3′-5′ exonucleolytic activities [13], [14]. The suppression of DNA replication errors is in part achieved by the nucleotide selectivity at the active sites of replicative polymerases that permits DNA synthesis with an error rate of 10−4–10−5 [6]. The excision of incorrectly incorporated dNMPs by the 3′-5′ exonucleolytic activity of replicative polymerases further decreases the error rate ∼100-fold. In addition, mismatch repair (MMR) promotes high-fidelity DNA replication by correcting replication errors which escaped the proofreading activities of replicative polymerases. MMR is a multifunctional process, but correction of DNA replication errors is its primary function [2], [11], [15]–[21]. Eukaryotic MMR is initiated by the binding of MutSα (MSH2-MSH6 heterodimer) or MutSβ (MSH2-MSH3 heterodimer) to a mispair. After detecting a mismatch, MutSα or MutSβ activates the endonuclease activity of MutLα (MLH1-PMS2 in humans and Mlh1-Pms1 in S. cerevisiae) in the presence of ATP, a strand break, and PCNA loaded by RFC [22]–[25]. A MutLα incision 5′ to the mismatch initiates the downstream events leading to the correction of the mismatch [26], [27]. MMR improves fidelity of DNA replication 10–104-fold depending on the sequence context. Thus, replicative polymerases and MMR are the major factors in high-fidelity DNA replication [28]–[31].
Several reversible histone modifications have been implicated in DNA replication, repair, and damage response [32], [33]. Histone H3 K56 acetylation is one such modification located in the αN-helix that is adjacent to the histone fold domain [34], [35]. When histone H3 acetylated on K56 (H3K56ac) is part of a nucleosome, the acetylation is near the entry and exit sites of DNA and appears to loosen the histone-DNA contacts [34]. Nearly all newly synthesized yeast H3 histones are acetylated on K56 [36] by the histone acetyltransferase Rtt109 and histone chaperone Asf1 in S phase [37]–[40]. The loss of yeast H3K56ac enhances the sensitivity of cells to several DNA damaging drugs [35], [40]–[42] and destabilizes stalled replication forks [43]. During DNA damage response, yeast H3K56ac is required for both restoration of chromatin on repaired DNA and subsequent recovery of the cells from the DNA damage checkpoint [41]. H3K56ac has been identified in human cells where it is also involved in DNA damage response [44], [45].
The NAD-dependent histone deacetylases Hst3 and Hst4 erase H3K56ac marks from the newly generated chromatin in G2/M [36], [46], [47]. Like H3 K56 acetylation, H3 K56 deacetylation by Hst3 and Hst4 is important for DNA damage response. In the presence of DNA damage in G2/M in wild-type strains, H3 K56 deacetylation is delayed to allow DNA repair to take place [35]. Furthermore, it is known that H3 K56 acetylation and deacetylation are critical for selecting sister chromatid as the template for repair of replication-born double strand breaks by homologous recombination (HR) [48]. About 92% of chromatin histone H3 molecules are continuously acetylated on K56 residues in hst3Δ hst4Δ strains [36]. Strains lacking both HST3 and HST4 display spontaneous DNA damage, a strong sensitivity to genotoxic agents, a five-fold increase in mitotic homologous recombination, and an elevated level of chromosome loss in mitosis [36], [47], [49], [50]. Hst3 and Hst4 are members of the conserved sirtuin family also containing Hst1 and Hst2 [49], [51]. The targets of Hst1 and Hst2 enzymes are not well defined. A recent study reported that Hst1 is important for histone H3 K4 deacetylation in euchromatin [52].
Nuclear DNA is part of chromatin, but little is known about the relationship between chromatin and mutation avoidance. Previous studies have demonstrated that the yeast chromatin factors Caf1, Asf1, Hst3, and Rtt109 are involved in the suppression of GCRs [39], [53]–[55]. Furthermore, human CAF-1 has been shown to interact functionally and physically with the mismatch recognition factor MutSα and modulate MMR in cell-free extracts and reconstituted systems [56], [57]. A recent report has described that a depletion of the histone methyltransferase SETD2 triggers microsattelite instability and an increased mutation frequency at HPRT [58]. Because microsattelite instability is a hallmark of MMR defects and the MSH6 subunit of MutSα recognizes H3K36me3, these findings suggest that SET2D-dependent H3K36me3 is required for the action of human MMR in vivo [58].
In this work, we analyzed the impacts of both H3 K56 deacetylation and acetylation on spontaneous mutagenesis in S. cerevisiae. We found that H3 K56 deacetylation by the combined action of Hst3 and Hst4 plays a major role in the defense against GCRs, base substitutions, 1-bp insertions/deletions, and complex mutations. Our analysis also showed that in addition to being part of the protection from GCRs [54], H3 K56 acetylation is involved in the prevention of small insertions/deletions and complex mutations. Furthermore, our results revealed that both the acetylation and deacetylation of H3 K56 are important for genetic stabilization mechanisms that act in concert with MMR and the proofreading activities of replicative DNA polymerases to suppress spontaneous mutagenesis.
We started this work to investigate whether chromatin is involved in the defense against spontaneous point mutations in the haploid yeast S. cerevisiae. Many of our experiments relied on CAN1 and his7-2 reporters for scoring mutations. CAN1 is a counter-selectable marker that allows the selection of any mutation that inactivates the gene including base substitutions, small insertions/deletions and complex mutations. In addition, CAN1 can be inactivated by GCRs involving the 43-kb CAN1-containing region of chromosome V [3]. The his7-2 reporter permits the selection of net +1 frameshift mutations causing a reversion to HIS7 [59]. As shown in Table 1, analysis of several histone deacetylase and acetyltransferase mutants revealed that the CAN1 and his7-2 mutation rates for three different hst3Δ hst4Δ strains are about 25 times as high as those for isogenic wild-type strains. However, deletion of HST3 or HST4 alone causes little or no mutator phenotype (Table 1). We also established that the CAN1 mutation rates in hst3Δ hst4Δ are very similar to those in the MMR-deficient msh2Δ and mlh1Δ strains (Figure 1C). Collectively, these data demonstrated that the loss of HST3 and HST4 strongly promotes spontaneous mutagenesis.
Hst3 and Hst4 remove acetylations on H3 K56 residues that are introduced by Rtt109 [36]–[40], [46]. No other enzymatic activity has been assigned to Hst3 and Hst4. Based on this information, we hypothesized that Hst3 and Hst4 participate in the suppression of spontaneous mutations (Table 1) by deacetylating chromatin histones H3 on K56. If this hypothesis is correct, the loss of H3K56ac by deletion of RTT109 or introduction of H3K56R should make the H3 K56 deacetylation activities of Hst3 and Hst4 unnecessary, and therefore suppress the mutator phenotype of hst3Δ hst4Δ. (H3K56R variant mimics histone H3 that is not acetylated on lysine 56 [35], [60].) However, if H3 K56 deacetylation by Hst3 and Hst4 is not involved in the protection of yeast genome from spontaneous mutations, the loss of H3K56ac should not affect the mutator phenotype of hst3Δ hst4Δ. We found that deletion of RTT109 or introduction of H3K56R suppresses the mutator phenotype of hst3Δ hst4Δ to the level observed in rtt109Δ and H3K56R (Table 1). We concluded from these data that H3 K56 deacetylation by Hst3 and Hst4 is required for the prevention of spontaneous mutations.
H3K56ac is weakly mimicked by H3K56Q [35], [46], [60], [61]. Consistent with H3K56Q being a weak mimic of H3K56ac [46], [60], [61], we observed that the CAN1 and his7-2 mutation rates for H3K56Q are increased, but 6 and 2 times lower, respectively, than those for hst3Δ hst4Δ (Table 1). Furthermore, we found that the mutation rates for H3K56Q and hst3Δ hst4Δ H3K56Q do not differ from each other (Table 1). Therefore, these data further support the conclusion that H3 K56 deacetylation by Hst3 and Hst4 is required for mutation avoidance.
Nicotinamide (NAM) is a potent inhibitor of Hst3, Hst4, and other NAD-dependent sirtuins [36], [50], [62]. Yeast strains grown in 25-mM NAM-containing media accumulate an abnormally high level of H3K56ac [36]. We studied whether the presence of NAM in the culture medium promotes spontaneous mutagenesis of several yeast strains. We found that exposure to 25-mM or 50-mM NAM increases the mutation rates in wild type (Figures 1A and 1B). For example, the CAN1 mutation rate for wild type treated with 50-mM NAM increases 30-fold compared to that for untreated wild type. Importantly, our control experiments established that exposing H3K56R and rtt109Δ to 25-mM or 50-mM NAM has no effect on their mutation rates (Figures 1A and 1B). Together, these results provided independent evidence that H3 K56 deacetylation is important for the suppression of spontaneous mutagenesis.
We also found that the CAN1 and his7-2 mutation rates in the hst3Δ hst4Δ strain grown in the presence of 25-mM NAM are twice and five times higher, respectively, than those in untreated strain (Figures 1A and 1B). This finding suggested that an NAD-dependent histone deacetylase activity participates in the defense against spontaneous mutations in hst3Δ hst4Δ.
Hst1 and Hst2 are homologous to Hst3 and Hst4, but their biological functions remain enigmatic [36], [49]. In light of our evidence that the mutation rates in hst3Δ hst4Δ are increased in the presence of 25-mM NAM (Figures 1A and 1B), we sought to determine whether Hst1 and Hst2 are involved in the suppression of spontaneous mutations. We found that the CAN1 and his7-2 mutation rates in the hst1Δ, hst2Δ, hst1Δ hst3Δ, hst1Δ hst2Δ hst3Δ, and hst1Δ hst2Δ hst4Δ strains are not significantly different from those in wild type (Table 1). Furthermore, deletion of HST2 in the hst3Δ hst4Δ and hst3Δ hst4Δ hst1Δ strains does not increase spontaneous mutagenesis above the existing levels. However, the mutation rates in hst3Δ hst4Δ hst1Δ are twice higher than those in hst3Δ hst4Δ (Table 1). Together, these findings suggested that Hst1, but not Hst2, contributes to maintaining genome integrity in strains lacking Hst3 and Hst4.
The histone acetyltransferase Rtt109 produces H3K56ac in the presence of the histone chaperone Asf1 [37]–[40]. We inquired whether H3 K56 acetylation plays a role in mutation avoidance. We determined that deletion of RTT109 causes 9- and 2-fold increases of the his7-2 and CAN1 mutation rates, respectively (Table 2). The mutation rates for asf1Δ and H3K56R are nearly identical to those for rtt109Δ (Table 2). Importantly, we found that there is epistasis between H3K56R and rtt109Δ for CAN1 and his7-2 mutations (Table 2). The simplest interpretation of these results is that H3 K56 acetylation by Rtt109 is involved in a mutation avoidance mechanism that suppresses spontaneous mutations in his7-2 and CAN1.
We considered the possibility that the absence of H3K56ac causes a defect in replication-coupled nucleosome assembly, which in turn increases spontaneous mutagenesis. The current view suggests that normal replication-coupled chromatin assembly in yeast depends on histone chaperones Caf1 (Cac1-Cac2-Cac3 heterotrimer) and Rtt106 [33]. However, as shown in Table S1, deletions of the replication histone chaperone genes CAC2 and RTT106 have nearly no effect on the CAN1 and his7-2 mutation rates. These results suggested that the increased mutagenesis in strains lacking H3K56ac is not caused by defects in the Caf1- and Rtt106-dependent chromatin assembly.
H3K56ac is required for conferring cellular resistance to several DNA-damaging drugs [35]. In this pathway, H3K56ac acts through the ubiquitin ligase containing Rtt101, Mms1, and Mms22 subunits [63], [64]. Our data revealed that deletion of RTT101, MMS1, or MMS22 causes an ∼7-fold increase in his7-2 frameshifts (Table S2). Furthermore, we established that rtt101Δ and rtt109Δ are epistatic for his7-2 frameshifts (Table S2). These results suggested that the Rtt101 cullin-containing ubiquitin ligase is part of an H3 K56 acetylation-dependent mutation avoidance mechanism.
H3K56ac is involved in the regulation of budding yeast transcription [65]. In one mechanism of transcriptional regulation, the presence of H3K56ac leads to the SWR-C-dependent removal of the histone variant H2A.Z from promoter-proximal nucleosomes [65], [66]. In strains deficient in H2A.Z, transcription of ∼320 genes is upregulated while transcription of ∼480 genes is repressed [65]. To test whether the H3K56ac-dependent transcription regulation plays a role in the control of spontaneous mutagenesis, we measured CAN1 and his7-2 mutation rates in htz1Δ and swr1Δ strains. (HTZ1 is the only gene for the histone variant H2A.Z and SWR1 encodes the catalytic subunit of the SWR-C chromatin remodeling complex.) As shown in Table 2, the CAN1 and his7-2 mutation rates in the htz1Δ and swr1Δ strains are nearly identical to those in wild type. These findings indicated that the defects in the SWR-C- and H2A.Z-dependent transcription regulation do not increase the levels of can1 and HIS7 mutations in strains proficient in both the acetylation and deacetylation of H3 K56.
Analysis of genetic interactions has been critical for understanding the functions of numerous proteins involved in mutation avoidance. Previous studies have defined the existence of multiplicative, synergistic, and additive relationships between mutants that inactivate different mutation avoidance mechanisms [29], [67]–[71]. In a synergistic relationship, the relative mutation rate for a double mutant is greater than the sum of those for the single mutants [29]. A multiplicative relationship is a form of synergistic relationship in which the relative mutation rate in a double mutant is equal to the product of those in the single mutants [29]. The presence of a synergistic or multiplicative relationship indicates that one of the mutants is deficient in one mechanism and the other mutant in a different mechanism, and that the two mechanisms act in concert to suppress the same pool of DNA lesions [29]. On the other hand, the existence of an additive relationship indicates that either mechanism suppresses a different pool of DNA lesions [29]. In an additive relationship, the relative mutation rate for a double mutant is equal to the sum of those for the single mutants [29], [71].
In hst3Δ hst4Δ strains, nearly all H3 histones are acetylated on K56 at replication forks [36], [47]. We thought that the presence of excess H3K56ac might interfere with high-fidelity DNA replication. Therefore, we decided to test whether histone H3 K56 deacetylation by Hst3 and Hst4 contributes to DNA replication fidelity. In these experiments we used four replication fidelity mutants: msh2Δ and mlh1Δ completely inactivate MMR [59], [72], pol2-4 disables the proofreading activity of DNA polymerase ε [73], and pol3-5DV eliminates the proofreading activity of DNA polymerase δ [74]. Based on the results of the previous research [29], [67]–[71] described above, we predicted that if a histone H3 K56 deacetylation-dependent mutation avoidance mechanism cooperates with MMR and the proofreading activities of replicative polymerases in promoting replication fidelity, each of triple mutant combinations containing hst3Δ hst4Δ and one of the replication fidelity mutants (msh2Δ, mlh1Δ, pol2-4, or pol3-5DV) should display synergistic or multiplicative, but not additive, increases in the relative CAN1 and his7-2 mutation rates. We found that the hst3Δ hst4Δ msh2Δ, hst3Δ hst4Δ mlh1Δ, hst3Δ hst4Δ pol2-4, and hst3Δ hst4Δ pol3-5DV triple mutants indeed show synergistic increases in the relative CAN1 and his7-2 mutation rates (Figure 1C and Table S3). In addition, weak synergies were observed when hst3Δ, but not hst4Δ, was combined with msh2Δ, mlh1Δ, pol2-4, or pol3-5DV (Table S3). Taken together, these findings suggested that an H3 K56 deacetylation-dependent mutation avoidance mechanism act in concert with MMR and the proofreading activities of replicative polymerases δ and ε to maintain high-fidelity DNA replication.
Because replication-coupled nucleosome assembly incorporates H3K56ac in chromatin in S phase [35], we tested whether this modification is important for maintaining high-fidelity DNA replication. We found a multiplicative increase in the relative CAN1 mutation rate when an H3 K56 acetylation mutant (rtt109Δ, H3K56R, or asf1Δ) was combined with a replication fidelity mutant (msh2Δ, pol2-4, or pol3-5DV) (Table 2). Furthermore, we established that all these double mutant combinations display synergistic increases in his7-2 mutation rates (Table 2). Collectively, these data suggested that an H3K56ac-dependent mutation avoidance mechanism cooperates with MMR and the proofreading activities of DNA polymerases to promote replication fidelity.
In addition to H3 K56, Rtt109 acetylates other targets [39], [75], [76]. Analysis of data in Table 2 indicated that the synergies between rtt109Δ and replication fidelity mutants for his7-2 mutations are often weaker than those between asf1Δ or H3K56R and msh2Δ, pol2-4, or pol3-5DV. Therefore, acetylation of a different target by Rtt109 may compromise replication fidelity.
To characterize spontaneous mutagenesis caused by the deficiency in H3 K56 deacetylation (Figure 1 and Table 1), we determined mutations that occurred within the 1.77-kb CAN1 ORF in the wild-type and hst3Δ hst4Δ strains by PCRs and DNA sequencing (Figures 2, S1, and S2). Consistent with a previous report [77], we observed that in the wild-type strain 79% of can1 mutations are base substitutions (Figure 2A). Genetic alterations detected in the hst3Δ hst4Δ strain include base substitutions, 1-bp deletions, 1-bp insertions, complex mutations, and deletions of CAN1 gene (Figures 2A and S2). Of those, deletions of CAN1 gene are the most common mutations generated at a rate of 190×10−8. This unexpected finding suggested that strains defective in H3 K56 deacetylaton are very susceptible to GCRs and we confirmed this idea in experiments described in the next subsection. We also found that base substitutions in the hst3Δ hst4Δ strain accumulate at a high rate of 160×10−8. Strikingly, the rate of base substitutions in hst3Δ hst4Δ is comparable with that in MMR-deficient msh2Δ. This finding suggested that H3 K56 deacetylation is a major player in the protection of S. cerevisiae from base substitutions. The most common base substitution in the spectrum of hst3Δ hst4Δ is a G→T transversion produced at a rate of 50×10−8 (Figures 2B and S2B). Analysis of the spectrum also suggested that C→G transversions and C→T transitions are formed at high rates in the H3 K56 deacetylation-deficient strain. Among other mutations detected in hst3Δ hst4Δ are six medium-size deletions ranging from 40-bp to 1,036-bp. Examination of the end points of the deletions revealed that five out of the six deletions occurred between perfect or nearly perfect direct repeats (Figure S2D).
Deletion of HST1 in hst3Δ hst4Δ promotes spontaneous mutagenesis (Table 1). To obtain additional insight into the interaction between hst1Δ and hst3Δ hst4Δ, we determined can1 mutation spectrum of hst3Δ hst4Δ hst1Δ. Analysis of the can1 mutation spectrum showed that the rates of base substitutions, 1-bp deletions, 1-bp insertions, and deletions of CAN1 gene for hst3Δ hst4Δ hst1Δ are 2–8 times higher than those for hst3Δ hst4Δ (Figures 2A). We also found that the rate of base substitutions for hst3Δ hst4Δ hst1Δ exceeds that for msh2Δ by 3-fold. Complex mutations comprising 2 or more mutations within an ∼10-bp DNA are a signature of the action of DNA polymerase ζ [5]. Surprisingly, the mutation spectrum of hst3Δ hst4Δ hst1Δ does not contain even a single complex mutation whereas six complex mutations are present in the spectrum of hst3Δ hst4Δ (Figures 2A and S2C). Collectively, these findings suggested that deletion of HST1 in hst3Δ hst4Δ significantly affects the dynamics of DNA metabolism.
To characterize the synergy between hst3Δ hst4Δ and msh2Δ (Figure 1C and Table S3), we determined can1 mutation spectra of the msh2Δ and hst3Δ hst4Δ msh2Δ strains. As expected from the results of an earlier work [72], 71% and 19% of mutations in the msh2Δ spectrum are 1-bp deletions and base substitutions, respectively (Figure 2A). Analysis of the data indicated that the rate of CAN1 gene deletions in hst3Δ hst4Δ msh2Δ is 5 times lower than that in hst3Δ hst4Δ (Figure 2A). This result provided us with the first clue that MMR might be involved in the formation of a large fraction of CAN1 deletions in H3 K56 deacetylation-defective strains. The can1 mutation spectrum of hst3Δ hst4Δ msh2Δ is dominated by base substitutions and 1-bp deletions accumulating at the rates of 1,100×10−8 and 1,600×10−8, respectively (Figure 2A). Among different base substitutions observed in the spectrum of hst3Δ hst4Δ msh2Δ, G→A changes are the most frequent (Figures 2B and S2B). Further analysis of the data revealed that there is a synergistic relationship [29] between hst3Δ hst4Δ and msh2Δ for base substitutions, 1-bp deletions, and 1-bp insertions (Figures S2A and S2B). For example, the relative rate of G→A substitutions for hst3Δ hst4Δ msh2Δ is 8 times as high as the sum of those for msh2Δ and hst3Δ hst4Δ (Figure S2B). Taken together, these findings established that H3 K56 deacetylation cooperates with MMR to prevent base substitutions, 1-bp deletions, and 1-bp insertions.
To better understand the H3 K56 acetylation-dependent suppression of spontaneous mutations, we determined spectra of mutations of the rtt109Δ and rtt109Δ msh2Δ strains (Figure 2A, 2C, and S2A). Compared to wild type, rtt109Δ displays higher rates of 1-bp insertions, complex mutations, and deletions of CAN1 (Figure 2A). The most common mutation in the his7-2 reporter of the rtt109Δ mutant was an A insertion that extended the A7 into an A8 run (Figure 2C). In addition, the HIS7 spectrum contains other net 1-bp insertions, small deletions, and complex mutations consisting of a 1-bp insertion and an adjacent base substitution. Noticeably, the rate of complex mutations in his7-2 for rtt109Δ is 20 times as high as that for wild type. Collectively, these results demonstrated that H3 K56 acetylation is important for the protection from 1-bp insertions, small deletions, and complex mutations. Comparison of can1 mutation spectra of the rtt109Δ, msh2Δ, and rtt109Δ msh2Δ strains revealed a synergy between rtt109Δ and msh2Δ for both base substitutions and 1-bp insertions/deletions (Figure S2A). Therefore, these data established that an H3 K56 acetylation-dependent mutation avoidance mechanism acts synergistically with MMR to prevent 1-bp deletions, base substitutions, and 1-bp insertions.
The genomic DNAs of 40% of our can1 hst3Δ hst4Δ isolates did not support PCR amplification of can1, but templated the expected POL2 PCR product (Figure S1). This finding implied that the hst3Δ hst4Δ mutant loses all or part of CAN1 due to GCRs (Figure 2A). To further investigate this phenomenon, we carried out experiments that took advantage of contour-clamped homogenous electric field (CHEF) electrophoresis coupled with Southern blot hybridization. The data revealed the presence of a rearranged chromosome V in can1 hst3Δ hst4Δ isolates that did not support PCR-based amplification of can1 (Figure 3A). Some of the isolates appear to carry fusions of the chromosome V arm with a different chromosome (Figure 3A, lanes 2, 4, and 6), while the other isolates contain deletions within chromosome V (Figure 3A, lanes 7–16). Such chromosomal rearrangements have been detected in previous studies [78], [79].
To provide further evidence that the defect in H3 K56 deacetylation triggers GCRs, we measured the rate of GCRs in the wild-type and hst3Δ hst4Δ strains using an approach developed by Richard Kolodner and coworkers [4]. URA3 was inserted 2.1-kb telomeric to CAN1 and the simultaneous loss of the two markers occurring as a result of a GCR was measured (Figures 3B and 3C). The rate of GCRs in the hst3Δ hst4Δ strain is 15,600-fold as high as that in wild type (Figure 3C). This finding demonstrated that the lack of H3 K56 deacetylation causes a dramatic increase in the rate of GCRs. Combining hst3Δ hst4Δ with pol3-5DV does not significantly change the rate of GCRs. Strikingly, the hst3Δ hst4Δ msh2Δ and hst3Δ hst4Δ mlh1Δ strains display GCR rates that are 15 times lower than that in isogenic hst3Δ hst4Δ (Figure 3C). Furthermore, we found that deletion of MSH3 or MSH6 in the hst3Δ hst4Δ mutant decreases the rate of GCRs (Figure 3C). These results suggested that the formation of the majority of GCRs in hst3Δ hst4Δ strains involves MMR action dependent on both MutSα and MutSβ. A model shown in Figure 3D outlines a possible mechanism of this phenomenon and is described in the Discussion section.
Next, we investigated whether several DNA repair proteins contribute to the high mutation rates in the hst3Δ hst4Δ strains (Table 3). As described above, the spectrum of hst3Δ hst4Δ contains complex mutations (Figures 2A and S2C). Rev3 is the catalytic subunit of DNA polymerase ζ [80] that produces complex and other mutations during replication of damaged and undamaged DNA and during double-strand break repair [71], [81]. The CAN1 and his7-2 mutation rates for rev3Δ hst3Δ hst4Δ are nearly identical to those for hst3Δ hst4Δ. Therefore, these results implied that DNA polymerase ζ is not responsible for the majority of mutations occurring in the H3 K56 deacetylation-deficient strains.
Deletion of RTT101 or CTF18 suppresses a strong defect of hst3Δ hst4Δ strains for growth at 37°C [47], [82]. Rtt101 is required for the progression of replication forks through pause sites and damaged DNA template [83], and is part of the H3K56ac-dependent resistance to genotoxic stress [82]. Ctf18 is the largest subunit of the Ctf18-RFC complex, which is essential for sister chromatid cohesion [84] and unloads PCNA from DNA [85]. We analyzed the involvement of both RTT101 and CTF18 in the formation of spontaneous mutations in hst3Δ hst4Δ. As shown in Table 3, the CAN1 and his7-2 mutation rates in rtt101Δ hst3Δ hst4Δ are 12 and 6 times lower, respectively, than those in hst3Δ hst4Δ. This finding implicated Rtt101 in the formation of the majority of can1 and HIS7 mutations in H3 K56 deacetylation-deficient strains. Furthermore, we found that the mutation rates in ctf18Δ hst3Δ hst4Δ are lower than those in hst3Δ hst4Δ. This result indicated that Ctf18-RFC is involved in promoting spontaneous mutagenesis in H3 K56 deacetylation-deficient strains.
The presence of complex mutations in the mutation spectra of rtt109Δ (Figures 2A and 2C) suggested that DNA polymerase ζ might contribute to spontaneous mutagenesis in the H3 K56 acetylation-deficient strains. We determined that deletion of REV3 in rtt109Δ completely suppresses the CAN1 mutation rate and decreases the his7-2 mutation rate two-fold (Table 3). These results support the idea that DNA polymerase ζ is involved in the formation of mutations in H3 K56 acetylation-deficient strains.
Rad52 and Rad51 are key components of HR [12]. To characterize the genetic interactions between H3 K56 acetylation and HR, we measured the CAN1 and his7-2 mutation rates in the rad51Δ, rad51Δ rtt109Δ, rad52Δ, and rad52Δ rtt109Δ strains (Table 3). Unfortunately, the relative CAN1 mutation rates in the single and double mutants do not allow us to distinguish between epistasis and additivity in the genetic interactions of rtt109Δ with the HR alleles (Table 3). However, we found that rtt109Δ displays epistatic relationships with rad51Δ and rad52Δ for his7-2 mutations (Table 3). This finding suggested that the recombination proteins and H3K56 acetylation act in the same pathway to promote the integrity of replication fidelity.
Because our results provided evidence that H3K56 acetylation acts synergistically with MMR (Table 2) and epistatically with HR (Table 3) to control spontaneous mutagenesis, we hypothesized that HR might contribute to fidelity of DNA replication. To test this hypothesis, we studied the genetic interactions between rad52Δ and msh2Δ (Figures 4A and 4B). We found the presence of a synergistic relationship between rad52Δ and msh2Δ for both CAN1 and his7-2 mutations. Rev3 produces the majority of mutations in rad52 strains by acting on ssDNA generated by the resection of double-strand breaks [86]. We established that in the rev3Δ background, the relationship between rad52Δ and msh2Δ is nearly multiplicative for both CAN1 and his7-2 mutations (Figures 4A and 4B). Next, we compared the can1 mutation spectra of the rad52Δ msh2Δ and msh2Δ strains (Figure 2A). The mutation spectrum of rad52Δ msh2Δ is similar to that of msh2Δ, but the rates of the most common classes of mutations (1-bp deletions, base substitutions, and 1-bp deletions) for msh2Δ are 2–3 times lower than those for rad52Δ msh2Δ. Taken together, these results suggested that Rad52-dependent HR contributes to fidelity of DNA replication.
Chromatin controls many critical aspects of metabolism in eukaryotes. Besides being a major regulator of transcription, chromatin profoundly affects DNA damage response, replication, and repair [32], [33], [87]. However, our knowledge about the relationship between chromatin and spontaneous mutagenesis is very limited. Previous research has identified that yeast Asf1, Caf1, Hst3, and Rtt109-dependent H3 K56 acetylation are involved in the control of GCRs [39], [53]–[55]. Additionally, a recent study reported that SET2D-dependent H3K36me3 regulates the mismatch correction function of human MMR [58]. Up to date, no information has been available about the involvement of either histone acetylation or deacetylation in the protection from point and complex mutations. In yeast S. cerevisiae, H3K56ac is an abundant posttranslational modification introduced in and removed from chromatin in a cell cycle-dependent manner [34]–[36], [46]. In this work, we investigated the impact of both the deacetylation and acetylation of H3 K56 on spontaneous mutagenesis in S. cerevisiae. We demonstrated that H3 K56 deacetylation by Hst3 and Hst4 plays a critical role in the suppression of GCRs, base substitutions, small insertions/deletions, and complex mutations (Figures 2, 3, and 5). Remarkably, a strain deficient in Hst3- and Hst4-dependent H3 K56 deacetylation forms GCRs at a rate that is 15,600-fold as high as that in isogenic wild type (Figure 3C). Furthermore, we showed that the rates of base substitutions in the hst3Δ hst4Δ and msh2Δ strains are similar to one another (Figure 2A). This finding suggests that H3 K56 deacetylation is as important for the prevention of base substitutions as MMR. We also showed that H3 K56 acetylation by Rtt109 and Asf1 is involved in the protection of DNA from 1-bp insertions, small deletions, and complex mutations (Figures 2A and 2C); however the effects of H3 K56 acetylation are weaker than those of H3 K56 deacetylation. Therefore, our findings indicate that in addition to controlling gene transcription and GCRs [39], [53]–[55], [87], histone acetylation and deacetylation are required for the defense against point and complex mutations.
This study was greatly facilitated by the availability of the H3K56R and H3K56Q alleles [35], [46], [60] (Figure 1 and Tables 1 and 2). In our experiments, H3K56R mimicked well H3 unacetylated on K56, but H3K56Q behaved as a weak mimic of H3K56ac. The latter conclusion is based on the observation that the mutation rates for hst3Δ hst4Δ exceeded those for H3K56Q and hst3Δ hst4Δ H3K56Q by 2–6 fold (Table 1). Previous studies have also found that H3K56Q mimics weakly H3K56ac [46], [60], [61].
Exposure of the hst3Δ hst4Δ strain to 25–50-mM NAM increases the mutation rates 2–5-fold (Figures 1A and 1B). This finding suggested that another NAD-dependent histone deacetylase is involved in mutation avoidance. Consistent with this, we found that the CAN1 and his7-2 mutation rates in hst1Δ hst3Δ hst4Δ are twice higher than those in hst3Δ hst4Δ (Table 1). Given that 8% of histone H3 is still unacetylated on K56 in hst3Δ hst4Δ strains [36] and that the CAN1 and his7-2 mutation rates for hst1Δ H3K56Q and H3K56Q do not differ from each other (Table 1), we hypothesize that Hst1 is involved in the suppression of spontaneous mutagenesis in hst3Δ hst4Δ cells by weakly deacetylating H3 on K56. Alternatively, Hst1 may promote genetic stability by acting on a different target. Since the removal of euchromatic H3K4ac mainly depends on Hst1 [52], H3K4ac may be this target.
Mutagens, in general, produce a lesion by directly acting upon DNA, which is later converted into a mutation. One of the few deviations from this rule is the demonstration that cadmium cations trigger genetic instability in yeast strains by inhibiting an enzymatic system, MMR [88]. We showed that exposure of yeast strains to 50-mM NAM, a specific inhibitor of the NAD-dependent histone deacetylases [62], produces a strong 30-fold increase in the CAN1 mutation rate (Figure 1A). This effect depends on the presence of both H3 K56 and RTT109. To the best of our knowledge, these data have provided the first example of a small molecule that inhibits chromatin-modifying enzymes and by doing so strongly promotes spontaneous mutagenesis.
MMR and the proofreading activities of DNA polymerases δ and ε are critical for maintaining high-fidelity DNA replication [2], [18]. We found the presence of synergistic increases in CAN1 and his7-2 mutation rates when hst3Δ hst4Δ is combined with msh2Δ, mlh1Δ, pol2-4, or pol3-5DV (Figures 1C, 2A, and 2B and Table S3). Furthermore, we established the existence of a synergy between hst3Δ hst4Δ and msh2Δ for base substitutions, 1-bp insertions, and 1-bp deletions (Figure 2A, 2B, and S2A). It is also evident that the relationships of rtt109Δ, asf1Δ, and H3K56R with msh2Δ, pol2-4, and pol3-5DV are synergistic for his7-2 frameshifts and multiplicative for CAN1 mutations (Table 2). The presence of synergistic and multiplicative relationships supports the view that both the deacetylation and acetylation of H3 K56 are involved in mutation avoidance pathways that act in concert with MMR and the proofreading activities of the replicative polymerases to promote DNA replication fidelity. The absolute CAN1 mutation rates for hst3Δ hst4Δ msh2Δ (Table S3) and msh2Δ rtt109Δ (Table 2) are 4.2 times and 7.6 times lower, respectively, than that for the pol2-4 msh2Δ mutant [69]. Therefore, this comparison suggests that the contributions of the acetylation and deacetylation of H3 K56 to replication fidelity are not as strong as that of the proofreading activity of DNA polymerase ε.
GCRs have been implicated in triggering many different cancers [9]. S. cerevisiae has been instrumental for dissecting the mechanisms of GCRs [4], [78], [89], [90]. A study that used a URA3-CAN1 cassette containing the two genes 7.5-kb apart from each other described that deletion of HST3 or RTT109 increases the rate of GCRs four-fold [55]. To analyze GCRs, we utilized a URA3-CAN1 cassette in which the distance between URA3 and CAN1 is 2.1-kb (Figures 3B and 3C). We determined that the deletions of HST3 and RTT109 cause 45- and 6-fold increases of the GCR rate, respectively (Figure 3C). Surprisingly, the rate of GCRs in hst3Δ hst4Δ exceeds those in the corresponding single mutants and rtt109Δ by at least 350-fold. Therefore, our findings (Figure 3C) are in good accord with and extend the previous observations that identified that Hst3 and Rtt109-dependent H3 K56 acetylation play roles in the control of GCRs [39], [53], [89]. In addition, our data (Figure 3C) support the view that Hst3 is the principal enzyme for H3 K56 deacetylation [36], [46].
Msh2 and Mlh1 are the key components of yeast MMR [2], [15]. Msh2 is a subunit of the mismatch recognition factors MutSα and MutSβ, whereas Mlh1 forms MutLα endonuclease by dimerizing with Pms1. Strikingly, deletion of MSH2 or MLH1 in hst3Δ hst4Δ reduces the rate of GCRs by 15-fold (Figure 3C). On the other hand, the rate of GCRs in the hst3Δ hst4Δ pol3-5DV mutant is nearly identical to that in hst3Δ hst4Δ. Therefore, these data demonstrate that MMR, but not the proofreading activity of DNA polymerase δ, is required for the generation of the majority of GCRs in the H3 K56 deacetylation-defective strains. We infer from these results that histone H3 K56 deacetylation is necessary to suppress malfunction of MMR. It is possible that in addition to promoting GCRs, MMR malfunction may result in the formation of some point mutations in hst3Δ hst4Δ (Figures 2A and 5). We speculate that MMR malfunction triggered by a defective environment may be responsible for the formation of a subset of cancer-initiating GCRs and point mutations. It has been known that MMR initiates several neurodegenerative diseases by destabilizing a number of DNA triplet repeats [2]. Thus, the idea that MMR can cause pathogenic consequences has already gained significant experimental support. We also analyzed the importance of the Msh6 subunit of MutSα and the Msh3 subunit of MutSβ for GCR formation (Figure 3C). The results of this analysis suggested that both MutSα and MutSβ contribute to the high rate of GCRs in hst3Δ hst4Δ, but the impact of the latter complex is somewhat weaker compared to that of the former.
How does MMR contribute to the formation of GCRs in hst3Δ hst4Δ cells? Our data permit us to suggest a speculative model shown in Figure 3D. It is known that in the absence of nucleosomes or concomitant nucleosome assembly the MutSα-dependent endonuclease activity of MutLα causes excessive degradation of mismatch-containing DNA in cell-free extracts and defined systems [22], [23], [56]. Therefore, we hypothesize that excessive and persistent nicking of DNA by MutLα may occur in the presence of the defective H3 K56 deacetylation. Such strand breaks can be converted into double-strand breaks in the next round of replication. If DNA flanking an end of one double-strand break carries a sequence that is a direct repeat of DNA flanking an end of another double-strand break, the MutSβ-dependent single strand annealing (SSA) mechanism [91]–[93] can join these two ends producing a GCR. Previous studies already demonstrated the importance of the MutSβ-dependent SSA mechanism for the repair of double-strand breaks flanked by direct repeats [91]–[93]. In this mechanism, MutSβ stabilizes the annealed DNA ends permitting the Rad1-Rad10 nuclease to cleave nonhomologous DNA tails [94]. The involvement of MutSβ and SSA, which is a major mechanism for repairing double-strand breaks flanked by direct repeats [95], in the formation of GCRs in hst3Δ hst4Δ cells is consistent with the following findings. First, the loss of MutSβ in hst3Δ hst4Δ decreases the GCR rate four-fold (Figure 3C). Second, five out of six identified medium-size deletions in CAN1 of the hst3Δ hst4Δ mutant were between direct repeats (Figure S2D).
In addition, a different mechanism may lead to GCRs in the hst3Δ hst4Δ strains. In this mechanism, H3 K56 hyperacetylation, MutSα or MutSβ, and a mismatch activate MutLα endonuclease to initiate the excision of the mismatch on opposite strands. Such aberrant excision may produce a double-strand break. When two double-strand breaks arise in the same hst3Δ hst4Δ cell, they may be repaired by the SSA mechanism causing a GCR. This mechanism is somewhat related to the one that has been proposed to explain the mismatch repair-dependent killing of E. coli dam recA mutants [96].
Though there are strong synergistic relationships between the H3 K56 acetylation mutants (rtt109Δ, asf1Δ, and H3K56R) and the replication fidelity defects (msh2Δ, pol2-4, and pol3-5DV) for CAN1 mutations, the double mutants show weaker synergistic increases in his7-2 mutations (Table 2). Furthermore, hst3Δ hst4Δ displays weak synergistic relationships with the MMR-deficient and proofreading mutants for CAN1 and his7-2 mutations (Figure 1C and Table S3). These findings suggest that a large fraction of mutations in both hst3Δ hst4Δ and rtt109Δ strains is produced from DNA lesions/mismatches formed outside S phase. In wild-type strains, H3K56ac appears in S phase and is removed in G2/M [36], [46]. Wild-type, hst3Δ, and hst4Δ strains do not have H3K56ac in G1, unlike hst3Δ hst4Δ [36]. Therefore, it is likely that a significant fraction of pre-mutagenic lesions/mismatches in hst3Δ hst4Δ is formed in G1 as a result of the presence of H3K56ac. Compared to wild type, rtt109Δ does not have H3K56ac in S phase and a part of G2/M. Hence, it is possible that S phase-independent pre-mutagenic lesions/mismatches in rtt109Δ arise in G2/M as a consequence of the lack of H3K56ac. In addition, the indicated variations in the presence or absence of H3K56ac in the different stages of the cell cycle provide a good explanation of why hst3Δ hst4Δ and rtt109Δ impact spontaneous mutagenesis differently (Tables 1–3 and Figures 2A and 3C).
What are the mechanisms that could be responsible for the generation of pre-mutagenic lesions in hst3Δ hst4Δ during G1 and in rtt109Δ during G2/M? Studies of gene transcription have identified many factors that recognize/read the presence or absence of histone modifications including histone acetylations [87], [97]. In addition, some factors read a specific combination of modifications [87]. After forming a complex with a modified/unmodified residue(s), the factor alters transcription of the affected gene. Thus, it is plausible that factors that read the absence or presence of H3K56ac alone or in combination with different modifications in the different stages of the cell cycle change transcription, DNA repair, and/or other mechanisms in a way that results in spontaneous mutagenesis. For example, transcription-dependent variations in the levels of some DNA repair proteins triggered by the defects in the deacetylation or acetylation of H3 K56 may shift the dynamics of DNA metabolism towards increased formation of spontaneous mutations. Consistent with this, it is known that H3K56ac is involved in several mechanisms that regulate gene transcription in yeast [34], [37], [65], and that transcription of ∼370 genes is deregulated in H3K56Q cells [65].
One of the mechanisms of transcriptional regulation that involves H3K56ac uses this modification to facilitate SWR-C-dependent removal of the H2A.Z variant from promoter-proximal nucleosomes [65]. Transcription of ∼900 genes is upregulated or downregulated in mutants lacking H2A.Z [65]. We tested whether defects in this H3K56ac-dependent transcriptional regulation affect spontaneous mutagenesis. However, we found that strains lacking H2A.Z or Swr1 do not have increased levels of CAN1 and his7-2 mutations (Table 2). Thus, the H3K56ac-dependent transcriptional regulation does not contribute to spontaneous mutagenesis in strains with the intact control of both H3 K56 acetylation and H3 K56 deacetylation. Nevertheless, it is still possible that this mechanism of transcriptional regulation contributes to the formation of mutations in strains that are deficient in the acetylation or deacetylation of H3 K56. Furthermore, the defects in the acetylation and deacetylation of H3 K56 may promote the formation of spontaneous mutations via a different mechanism of transcriptional regulation.
The abundant and persistent H3K56ac in hst3Δ hst4Δ mutants impairs DNA replication [47]. Replication forks in hst3Δ hst4Δ are able to adapt to the high level of H3K56ac when RFC1 is overexpressed or CTF18 is deleted [47]. The overexpression of RFC1 or deletion of CTF18 also suppresses the temperature-sensitive phenotype of the cells. Additionally, we observed that deletion of CTF18 in hst3Δ hst4Δ strongly reduces the mutation rates (Table 3). Because RFC loads PCNA onto DNA and CTF18-RFC unloads the clamp from DNA [85], these results suggest that a higher concentration of PCNA at replication forks allows the cells to adapt to the abundant and persistent H3K56ac. Intriguingly, PCNA is also required for MMR [22], [98], [99]. Thus, we speculate that inadequate concentrations of PCNA at replication forks in hst3Δ hst4Δ strains may impair both the replicative proofreading and MMR and this compromises replication fidelity. Nevertheless, it is also feasible that factors that recognize unacetylated H3 K56 and promote mutation avoidance cannot be recruited to replication forks in hst3Δ hst4Δ cells. Alternatively, the presence of the excessive H3K56ac may cause recruitment of mutagenic factors to the replication forks.
What is the mechanism that causes spontaneous mutagenesis in S phase in rtt109Δ cells? Our results suggest that H3K56ac is involved in a yet unknown HR mechanism that promotes replication fidelity (Figure 4 and Tables 2 and 3). The progression of replication forks is often impeded by spontaneous DNA damage. Therefore, it is possible that H3K56ac is important for an error-free bypass of spontaneous lesions by the HR machinery during DNA replication. In this mechanism, the presence of H3K56ac may be necessary for recruiting an interacting complex that promotes efficient chromatin remodeling around the lesions and by doing so facilitates an error-free bypass. Understanding the mechanisms that depend on the acetylation and deacetylation of H3 K56 to prevent spontaneous mutagenesis will require further experimentation.
In summary, our findings revealed that the cell cycle-regulated acetylation and deacetylation of chromatin on H3 K56 are critical for suppressing spontaneous mutagenesis (Figure 5). The acetylation and deacetylation of H3 K56 are involved in mutation avoidance mechanisms that act in concert with MMR and replicative polymerases to maintain genome stability. The lack of H3K56ac appears to compromise an HR mechanism that promotes replication fidelity. Defective H3 K56 deacetylation causes spontaneous mutagenesis involving Rtt101 and Ctf18, and results in the formation of MMR-dependent GCRs.
The S. cerevisiae wild-type strains used in this work are E134 (MATα ade5-1 lys2::InsE-A14 trp1-289 his7-2 leu2-3,112 ura3-52) [68], E35 (MATα ade5-1 lys2::InsE-A8 trp1-289 his7-2 leu2-3,112 ura3-52) [68], BY4742 (MATα his3Δ1 leu2Δ0 lys2Δ0 ura3Δ0), and 1B-D770 (MATa ade5-1 lys2::Tn5-13 trp1-289 his7-2 leu2-3,112 ura3-4) [59]. Strain SY579 and plasmids pPK588 and pPK589 have previously been described [60], [100]. If not indicated, the mutant strains are derivatives of E134. All strains used in this work are listed in Table S4.
To create gene replacements, disruption cassettes with homologous or heterologous markers [101] were amplified in PCRs and introduced into yeast cells by the lithium acetate/PEG-based transformation method [102]. Yeast genomic DNAs were isolated from recombinant isolates with MasterPure Yeast DNA purification kits (Epicentre) and all gene replacements were verified by PCR. The pol2-4 [67] and pol3-5DV [74] alleles were introduced by the integration-excision method. To analyze GCRs, S. cerevisiae URA3 gene amplified from the pFL34 plasmid with primers (5′- ATACATGCACATATAGCTACTACATAGTCAAGAACATATCATAACATTTGTCTGGCTTTTCAATTCATC-3′) and (5′- GTCGGTAGAGCCAGCATCAGATGCAAAGCCATGCAAAGACTGATATAAAGACTGTTATACAGATCTGAGCTTTTTCTTTCC-3′) was inserted at position 29617 of chromosome V, which is 2,077 bp telomeric to CAN1 and 3′ adjacent to SIT1.
Spontaneous mutation rates were measured using fluctuation analysis carried out according to a previously described procedure [59]. On average 15 cultures (no less than 9 cultures), started from single colonies of two-four freshly prepared independent isolates of the same genotype, were used to determine spontaneous mutation rate for this genotype. The cultures were grown to saturation in 3–50 ml YPD medium (1% yeast extract, 2% bacto-peptone, 2% dextrose) supplemented with 60 mg/L adenine and 60 mg/L uracil at 30°C for 20–48 h. When indicated, nicotinamide was added to the supplemented YPD medium to the final concentration of 25 mM or 50 mM. Each saturated culture was plated, after dilution, on a synthetic complete (SC) medium for scoring the total number of cells. The cultures were also plated on SC medium lacking histidine for scoring His+ revertants, SC medium lacking arginine and supplemented with 60 mg/L L-canavanine for scoring can1 mutants, and/or SC medium lacking arginine and supplemented with 60 mg/L L-canavanine and 1 g/L 5-FOA for scoring GCRs. The plates were incubated for 3–5 days at 30°C, and the colonies were counted.
The CAN1 and his7-2 mutation rates were calculated from the total numbers of cells and mutants in the cultures with the Drake formula and are presented as median values with 95% confidence intervals [7], [59]. Where indicated, the significance of observed differences in the CAN1 and his7-2 mutation rates was analyzed with Mann-Whitney U two-tailed test (GraphPad Prism 6 software), where the null hypothesis is that there is no difference between the two data sets. The rates of GCRs were calculated with the Ma-Sandri-Sarkar maximum likelihood method [103], [104] using the web tool FALCOR at http://www.keshavsingh.org/protocols/FALCOR.html [105], [106].
Mutation spectra in CAN1 gene were determined essentially as described [71]. Patches were started from single colonies on YPD plates and then replica plated on SC plates supplemented with 60 mg/L L-canavanine and lacking arginine. A single CanR clone from each patch was randomly selected, purified on the selective medium, and then propagated by patching on a YPD plate. Genomic DNAs were isolated from the patched cultures with a MasterPure Yeast DNA purification kit (Epicentre). 2,057-bp fragments containing the entire length of can1 ORF were amplified with primers 1 (5′-GCAGAAAGAAGAGTGGTTGCGAAC-3′) and 2 (5′-GAGAATGCGAAATGGCGTGGAAATG-3′) in PCR reactions. The amplified fragments were purified with a PCR purification kit (Qiagen) and sequenced such that the entire DNA sequence of the mutant ORF in each clone was determined.
To generate HIS7 mutation spectra for the wild-type and rtt109Δ strains, 1.4-ml or 2.8-ml saturated cultures started from single colonies were concentrated and plated on SC medium lacking histidine. One His+ clone from each plate was randomly selected and processed as above. 2005-bp fragments spanning HIS7 ORF were PCR-amplified with primers 3 (5′-CTCCACGGCTAATTAGGTGATCATG-3′) and 4 (5′-CCTACTGACACCACCAATAATACAAC-3′). The PCR fragments were purified as described above and part of HIS7 ORF, corresponding to chromosome II coordinates 716234 – 715434, was sequenced. his7-2 reverts to HIS7 by acquiring a +1-net frameshift in a 51-bp region (chromosome II coordinates 716023 - 715973) containing an A7 run [59].
Yeast cells were embedded into 0.8% agarose plugs at a concentration of 6×108 cells/ml, and chromosomal DNA was separated by a contour-clamped homogeneous electric field (CHEF) gel electrophoresis in a 1.2% agarose gel/0.5×TBE for 40 hours at 6 V/cm and at 14°C, using the CHEF Mapper XA system (Bio-Rad). The included angle was 120 degrees. The initial and final switch times were 36.63 sec and 2 min 6.67 sec, respectively. The separated yeast chromosomal DNAs were transferred onto a nylon membrane and probed with a 32P-labeled MET6-specific probe. (The probe was generated by a random prime labeling of a MET6 PCR fragment amplified with primers 5′-GACGCCATCAAGGGCTTGCCAG-3′ and 5′-CGTTAGCTTCTAGGGCAGCAGC-3′.) The indirectly labeled yeast chromosomal DNAs were visualized with a Kodak BioMax film.
|
10.1371/journal.ppat.1007851 | Enteropathogenic Escherichia coli remodels host endosomes to promote endocytic turnover and breakdown of surface polarity | Enteropathogenic E. coli (EPEC) is an extracellular diarrheagenic human pathogen which infects the apical plasma membrane of the small intestinal enterocytes. EPEC utilizes a type III secretion system to translocate bacterial effector proteins into its epithelial hosts. This activity, which subverts numerous signaling and membrane trafficking pathways in the infected cells, is thought to contribute to pathogen virulence. The molecular and cellular mechanisms underlying these events are not well understood. We investigated the mode by which EPEC effectors hijack endosomes to modulate endocytosis, recycling and transcytosis in epithelial host cells. To this end, we developed a flow cytometry-based assay and imaging techniques to track endosomal dynamics and membrane cargo trafficking in the infected cells. We show that type-III secreted components prompt the recruitment of clathrin (clathrin and AP2), early (Rab5a and EEA1) and recycling (Rab4a, Rab11a, Rab11b, FIP2, Myo5b) endocytic machineries to peripheral plasma membrane infection sites. Protein cargoes, e.g. transferrin receptors, β1 integrins and aquaporins, which exploit the endocytic pathways mediated by these machineries, were also found to be recruited to these sites. Moreover, the endosomes and cargo recruitment to infection sites correlated with an increase in cargo endocytic turnover (i.e. endocytosis and recycling) and transcytosis to the infected plasma membrane. The hijacking of endosomes and associated endocytic activities depended on the translocated EspF and Map effectors in non-polarized epithelial cells, and mostly on EspF in polarized epithelial cells. These data suggest a model whereby EPEC effectors hijack endosomal recycling mechanisms to mislocalize and concentrate host plasma membrane proteins in endosomes and in the apically infected plasma membrane. We hypothesize that these activities contribute to bacterial colonization and virulence.
| Enteropathogenic Escherichia coli (EPEC) are pathogenic bacteria that cause infantile diarrhea. Upon ingestion, EPEC reaches the small intestine, where an injection device termed the type III secretion system is utilized to inject a set of effector proteins from the bacteria into the host cell. These proteins manipulate the localization and functions of host proteins, lipids and organelles and contribute to the emergence of the EPEC disease. The molecular mechanisms underlying the functions of the EPEC effector proteins are not completely understood. Here we show that early upon infection, two such effector proteins, EspF and Map, hijack host endosomes at bacterial adherence sites to facilitate endocytosis and recycling of plasma membrane proteins at these sites. The consequence of this event is the enrichment and mislocalization of host plasma membrane proteins at infection sites. One such protein is the transferrin receptor, which is a carrier for transferrin, whose function is to mediate cellular uptake of iron. Iron is a critical nutrient for bacterial growth and survival. We postulate that the unique manipulation of transferrin receptor endocytic membrane trafficking by EPEC plays an important role in its survival on the luminal surface of the intestinal epithelium.
| Enteropathogenic Escherichia coli (EPEC) and enterohemorrhagic Escherichia coli (EHEC) are diarrheagenic extracellular pathogens affecting humans worldwide [1]. EPEC utilizes a molecular syringe termed type III secretion system (T3SS) to deliver a set of effector proteins from the bacterial cytoplasm into the host cell. The coordinated action of these effectors in space and time is critical for remodeling diverse host cell organelles and processes, e.g. the cytoskeleton, signaling pathways and intracellular trafficking, to support successful bacterial colonization and survival in the intestinal mucosa [2,3,4]. A prominent hallmark of EPEC infection is the induction of “attaching and effacing” (AE) lesions of the mucosal tissue, characterized by intimate microbial attachment to the apical plasma membrane of the infected epithelial cells, local elimination of brush border microvilli and the formation of a filamentous actin (F-actin)-rich membrane protrusion (often called pedestal) located immediately beneath the attached bacterium [5,6].
EPEC (strain O127:H6 E2348/69) translocates via its T3SS at least 21 ‘effector’ proteins into its host. These effectors are encoded by genes in the locus of enterocyte effacement (LEE) pathogenicity island, and by outside the LEE (non-LEE) genomic sites [7,8]. The first translocated effector is the intimin receptor (Tir) [9]. Upon translocation, Tir is incorporated into the host cell plasma membrane and its exposed extracellular domain binds the bacterial outer membrane protein, intimin. Tir-intimin interactions allow intimate attachment of the bacterium to its host plasma membrane and the formation of the F-actin rich pedestal [3]. Other injected effectors manipulate various processes in the host, including signaling pathways and immune response processes, cytoskeletal remodeling, the subversion of epithelial junctional components and membrane trafficking pathways [10,11,12]. Notably, however, limited information exists about the capacity of such effectors to manipulate membrane trafficking pathways of their host and, in particular, endosomal trafficking. Exploring the capacity of these effectors to hijack and manipulate endocytic trafficking pathways in a way that benefits the pathogen is important for better understanding the mechanisms by which EPEC colonizes its host.
Studies have shown that regulators of clathrin-mediated endocytosis (CME) are recruited to sites of EPEC infection. These include phosphatidylinositol 4,5-bisphosphate [PI(4,5)P2] [13], dynamin [14], the adaptor proteins Eps15 and epsin1 [15], Dab2 and Hip1R [16], CD2AP [17] and clathrin [18]. The recruitment of these components has been linked to the process of actin-rich pedestal formation. However, whether these recruitments are directly caused by injected effectors or a result of a broad host response to the bacterial assault is not known. The endocytic Rab5 [19], and the recycling VAMP3,Rab11 and Rab35 [20,21] protein markers have also been shown to be subverted by EPEC and EHEC. Nonetheless, neither study addressed the question of whether these endocytic regulators are recruited to infection sites and by what mechanism.
Data linking EPEC type III secreted effectors to endocytic trafficking have shown that the protein effector EspF transiently interacts with pre-existing clathrin-coated pits, and binds neuronal Wiskott-Aldrich syndrome protein (N-WASP) as well as the dynamin-associated PX-BAR domain sorting nexin 9 (SNX9) protein [22,23]. Recently, Tapia et al. have shown that EspF is involved in facilitating basolateral endocytosis and apical redistribution of the Crumbs polarity complex and Na+/K+ ATPase [19]. However, the mode by which EspF regulates these endocytic activities remains unknown. Here we show that early upon EPEC microcolony attachment to non-polarized host cells, EspF and Map prompt the recruitment of transferrin receptors (TfnR) and Rab11a-positive recycling endosomes to peripheral infection sites. This event may contribute to an increase in TfnR endocytosis and recycling (i.e. endocytic turnover) at these sites. We identified three novel EspF binding proteins, WIPF1, SNX18 and SNX33, which together with N-WASP and SNX9 could contribute to endocytic remodeling of the host. In polarized cells, we show that EspF is sufficient to promote basolateral-to-apical transcytosis of TfnR and increase in apical endocytic turnover. Our results tie for the first time the activity of an EPEC effector protein to endosomal dynamics and breakdown of surface polarity of infected cells. This event could lead to enrichment of the apical infected plasma membrane and endosomes with basolateral membrane proteins that contribute to bacterial colonization and the EPEC disease.
Previous studies have shown that EPEC and EHEC subvert various elements of the clathrin-coated pits endocytic and recycling machineries [13,14,15,17,18,19,20,21,24]. However, the extent to which T3SS elements are involved in this event has been sparsely studied. We therefore investigated the T3SS dependence of the distribution of clathrin endocytic (CHC, AP2-α, Rab5a and EEA1) and recycling [Rab11a, Rab11b, Rab25, Myosin 5b (Myo5b)] components upon EPEC infection of polarized MDCK cells. Results show that all these markers accumulated at EPEC-wt, and not at EPEC-escV infection sites (S1 Fig). These data suggest the involvement of type III secreted elements in the recruitment of early and recycling endocytic components to bacterial infection sites at the apical plasma membrane.
Transferrin receptors (TfnR) are basolateral recycling membrane proteins that typically do not reach the sub-apical Rab11a/Rab25 endosomal recycling compartment (reviewed in [25]) unless they are enforced to undergo transcytosis [26]. We analyzed whether the distribution of TfnR and their Tfn cargo is altered in response to EPEC infection of polarized MDCK-PTR9 (S2A Fig), Caco2-BBe (S2B Fig), or MDCK-GFP-TfnR (S3A Fig) cells. Fluorescently-tagged Tfn was internalized either from the apical, or from the basolateral (basal) plasma membrane of the polarized cells concomitant to apical exposure to EPEC-wt or EPEC-escV. The ligand distribution along the basal-apical axis of the cells was analyzed in fixed (MDCK-PTR9 and Caco2-BBe), and live (MDCK-GFP-TfnR) cells by confocal microscopy. Tfn internalized from either cell pole was found to be clustered at EPEC-wt, but not at EPEC-escV, infection sites. Surface immunostained TfnR (s-TfnR) also showed significant clustering at EPEC-wt, compared to EPEC-escV infected cells. Interestingly, SH3BP4 and ACAP1, which have been shown to promote TfnR endocytosis and recycling by binding to the receptor [27,28], were also recruited to infection sites in a T3SS-dependent manner (S3B Fig). These data suggest that T3SS elements prompt the clustering of Tfn-TfnR complexes and proteins that directly modulate their endocytic/recycling pathways at apical infection sites.
The observation that early endocytic and recycling machineries are recruited to infection sites has raised the hypothesis that EPEC modulates them to control membrane trafficking at the infected plasma membrane. To address this hypothesis, we examined whether EPEC infected cells display altered Tfn internalization and recycling capacities. To this end, MDCK-PTR9 cells were infected and concomitantly exposed to Tfn administered at either the apical or the basolateral poles of the cells. In the case of apically internalized Tfn, one set of EPEC-wt infected cells was treated with the dynamin inhibitor Dynasore. Cells were then washed, lysed and the amount of cell-associated Tfn was analyzed by Western blotting. Data showed an apparent increase in cell-associated Tfn only when the ligand was administered to the apical surface of EPEC-wt infected cells. The apical intake of Tfn by the Dynasore treated cells was significantly lower compared to EPEC-wt infected cells and uninfected cells (Fig 2A). These data suggest that type III secreted elements stimulate dynamin-dependent apical endocytosis of Tfn.
To further validate this conclusion, we applied a flow-cytometry-based assay for monitoring Tfn endocytosis and recycling. Polarized MDCK-PTR9 cells were infected with EPEC-wt, or EPEC-escV, or left uninfected, in the presence of fluorescently-tagged Tfn applied at either the apical, or the basolateral (basal) poles of the cells. Thereafter, adherent cells were detached from their substrate and cell-associated Tfn was determined by flow cytometry. Results showed an apparent increase in cell-associated Tfn only in EPEC-wt infected cells that were exposed to apical Tfn (Fig 2B). The EPEC-wt driven increase in Tfn endocytosis was also observed in polarized MDCK-GFP-TfnR cells (Fig 2C), and in Caco2-BBe cells, which express endogenously the hTfnR (Fig 2D). The intake of fluorescently labeled Tfn by EPEC-wt infected cells was diminished when an excess of unlabeled Tfn was co-administered with Tfn-AF647 to the apical plasma membrane of the cells (Fig 2C). In contrast, the addition of unlabeled Tfn to the basolateral compartment had no effect on the EPEC-induced intake of Tfn-AF647 added to the apical plasma membrane (Fig 2C). These data further argue that Tfn is specifically internalized from the infected apical plasma membrane. Finally, we used flow cytometry to monitor the effects that EPEC infection may have on the release (i.e. apical recycling) of apically internalized Tfn. Results showed enhanced Tfn release in response to EPEC-wt infection, compared to EPEC-escV and uninfected cells (Fig 2E). Together, these data suggest that type III secreted components stimulate apical endocytosis and recycling of Tfn, i.e. increase the apical endocytic turnover of the ligand.
Next, we asked whether the effects observed in polarized cells can be extended to non-polarized cells. Indeed, we found that early and recycling endocytic markers (S4A & S4B Fig), as well as internalized Tfn and s-TfnR (S4C Fig), were recruited at EPEC-wt, but not at EPEC-escV infection sites. Monitoring the recruitment dynamics of some of these markers using live cell imaging revealed that the process, which is T3SS-dependent, is initiated early upon EPEC microcolony landing on the host (S5 Fig; S1–S6 Movies). Unlike Tfn, fluid-phase uptake (70 kDa Dextran), lysosomal (LysoTracker) and late endosomal (Rab7a) markers were not recruited at EPEC-wt infection sites (S6 Fig). Additionally, EPEC also prompts the recruitment of the Rab11a/Myo5b-dependent recycling cargoes, β1-integrins, in non-polarized and polarized cells (S7 Fig [21,33]). Collectively, these data suggest that the recruitment effects were specific for cargoes utilizing CME and Rab11 recycling endosomes. Flow cytometry analysis showed increased levels of Tfn endocytosis and recycling, as well as increased surface bound Tfn in EPEC-wt, compared to EPEC-escV and uninfected cells (S8A Fig). The EPEC-induced increase in Tfn endocytosis was inhibited by the dynamin inhibitors Dynasore and Dyngo (S8B Fig). These data, together with data presented in Fig 3, suggest that infected plasma membranes of polarized and non-polarized cells promote T3SS-dependent increase in Tfn endocytic turnover.
In non-polarized cells, internalized TfnR utilizes either a fast constitutive recycling pathway i.e. shuttling to the cell surface from early recycling endosomes or a slow pathway that involves the transport via perinuclear Rab11-positive recycling endosomes [34]. The latter is thought to play a key role in directing recycling proteins to specific and specialized cellular locations on the cell surface, such as the leading edges of motile cells [35,36]. On the basis of these paradigms, we hypothesized that T3SS elements mediate the trafficking of TfnR from the slow perinuclear/pericentriolar recycling endosomal compartment to infection sites confined to the host periphery. To address this hypothesis, endosomes of uninfected, EPEC-escV, or EPEC-wt infected HeLa cells were loaded with Tfn-AF647 and imaged by confocal microscopy. The area of perinuclear recycling endosomes was determined. As predicted, internalized Tfn accumulated at the perinuclear recycling endosomal compartment of uninfected cells (Fig 3A, red arrow). The area occupied by this compartment was not significantly altered in EPEC-escV infected cells. In contrast, the area of the perinuclear endosomal compartment was considerably diminished in EPEC-wt infected cells (Fig 3A). The majority of Tfn positive endosomes apparently redistributed mainly to the periphery of the cells and concentrated at the infection sites (Fig 3A; yellow and green arrows). We applied single particle image analysis to explore whether Tfn is enriched in individual peripheral endosomes of the infected cells. As expected, data showed significant enrichment of Tfn in TfnR (Fig 3B) and in EEA1 (Fig 3C)-positive peripheral endosomes in EPEC-wt, but not in EPEC-escV or uninfected cells. EEA1 acts downstream of Rab4 (fast recycling) and Rab5 (early) endosomes [37,38,39]. Thus, the fact that Tfn was enriched in EEA1-positive endosomes suggests that EPEC type III secreted elements shifted the pathway of Tfn from slow to fast recycling endosomes. The increased abundance of Tfn-positive endosomes at peripheral infection sites in EPEC-wt compared to EPEC-escV infected cells could also be observed at the ultrastructural level (Fig 3D).
We asked whether vectorial shuttling of Tfn/Rab11a-positive endosomes takes place from perinuclear to peripheral infection sites. To this end, HeLa cells were transfected with Rab11a fused to photoconvertible tdEos encoding plasmid (tdEos-Rab11a). These cells were loaded with Tfn-AF647 concomitant to EPEC-wt or EPEC-escV exposure. The Tfn-AF647 and tdEos-Rab11a labeled perinuclear endosomes (Fig 4; Pre-Conversion; red arrow) were selected and irreversibly photoconverted (Fig 4; Post-Conversion; red arrow), and time-lapse imaging was applied to track the distribution of the photoconverted Rab and internalized Tfn-AF647 over time (S7 and S8 Movies). A representative image of EPEC-wt infected cells (Fig 4; t = 15; yellow arrow) and quantitative analysis of the time-dependent accumulation of photoconverted Rab11a at EPEC-wt and EPEC-escV infection sites are shown (Fig 4). Data clearly show T3SS-dependent recruitment of the photoconverted Rab11a at bacterial infection sites, suggesting that type III secreted components elicit the shuttling of Rab11a/Tfn-positive endosomes from perinuclear recycling endosomes to peripheral infection sites.
To further elucidate this notion, we used Dynasore and Dyngo, which inhibit endocytosis but not recycling of clathrin-dependent cargoes [40]. HeLa cells were first treated with Dynasore or Dyngo and then exposed to EPEC and Tfn-AF647. Cells treated with DMSO served as controls. The localization of Tfn, Rab5a, EEA1 (endocytic markers) and Rab11a (recycling marker) in these cells was analyzed by confocal microscopy. Interestingly, while Tfn, Rab5a and EEA1 showed very low levels of clustering at infection sites, Rab11a displayed significantly higher clustering at these sites (S9A Fig). If Tfn was internalized into the cells prior to exposure to Dyngo along with EPEC-wt, both Tfn and Rab11a clustered efficiently at infection sites (S9B Fig). Clustering of F-actin was not affected in any of the experiments. These results suggest again that Tfn/Rab11a recruited at peripheral infection sites are contributed by recycling endosomes.
As the movement of TfnR-Rab11 positive recycling endosomes is controlled by Myo5b motors [32], we asked whether Myo5b is involved in their recruitment at infection sites. To address this question, the following constructs were ectopically expressed in HeLa cells: GFP-Myo5b-FL, the GFP-Myo5b-tail mutant (Myo5b-tail), which harbors the Rab8/Rab11 C-terminal binding motif but lacks the ability to mobilize Rab11-dependent cargo due to deleted N-terminal motor domain, the GFP-Myo5b tail-QLYC mutant, which binds Rab11 but not Rab8a, and the GFP-Myo5b-tail-YEQR mutant, which does not bind Rab11 but binds Rab8a [41] (see S3 Table). Consistent with previous data, over expression of Myo5b-tail and Myo5b-tail-QLYC, but not of Myo5b-FL or Myo5b-tail-YEQR, resulted in the sequestration of Rab11a and Tfn into large intracellular puncta (S10A Fig). These data further corroborate the existence of a correlation between the ability of these Myo5b mutants to sequester Rab11a-Tfn endosomes and to reduce s-TfnR levels (S10B Fig). These effects are contributed by the ability of Myo5b-tail mutants to inhibit Tfn recycling, resulting in the retention and accumulation of internalized Tfn within the cells (S10C Fig). Infection of these cells with EPEC-wt resulted in significant Tfn (Fig 5A) and Rab11a (Fig 5B) recruitment in Myo5b-FL or Myo5b-tail-YEQR, but not in Myo5b-tail or Myo5b-tail-QLYC expressing cells. Infection with EPEC-wt induced Tfn endocytosis in Myo5b-FL or Myo5b-tail-YEQR, but not in Myo5b-tail or Myo5b-tail-QLYC expressing cells (Fig 5C). These data suggest that EPEC requires accessible Rab11 and functional Myo5b motors for recruiting TfnR and Rab11a-positive endosomes to peripheral plasma membrane infection sites, and for stimulating Tfn endocytosis.
The involvement of Rab11 in endosomal subversion by EPEC was further examined using selective gene silencing by siRNA. The expression level of Rab11a and Rab11b was either individually or simultaneously knocked-down using siRNA, as demonstrated in Fig 6A. Consistent with previous studies [42], internalized Tfn localized mainly to the cell periphery, rather than to the perinuclear recycling endosome, in the siRab11a+b depleted cells (S11A Fig). Additionally, a significant diminishment in perinuclear recycling endosomal area size (S11B Fig), and an increase in Tfn endocytic turnover (S11C Fig) were observed in these cells. Interestingly, Tfn clustering at EPEC-wt infection sites was apparent in cells treated with siRab11a, or siRab11b, but not in cells treated with siRab11a+b (Fig 6B), suggesting that the two Rab11 variants share structural information that supports the EPEC-mediated recruitment of Tfn to infection sites. Our flow cytometry data showed that EPEC-wt failed to increase Tfn endocytosis in the Rab11a+b depleted cells (Fig 6C). The inability of EPEC to further enhance endocytosis in these cells could be attributed to lack of capacity of the pathogen to further redistribute the already peripherally distributed recycling endosomes, and to concentrate them at infection sites.
Next, we aimed at identifying type III secreted effectors that might mediate the clustering of Rab11a and Tfn-positive endosomes at bacterial infection sites. HeLa cells were infected with a series of EPEC strains mutated in their LEE, or non-LEE effector encoding genes, and concomitantly exposed to Tfn-AF647. Cells were then fixed, subjected to F-actin staining and confocal imaging. Results show that, compared to EPEC-wt infected cells, infection with EPEC-espF, EPEC-map, EPEC-cesT, or EPEC-escV resulted in low levels of Tfn/TfnR clustering at infection sites (Fig 7A). F-actin recruitment to EPEC-espF and EPEC-map infection sites was not significantly altered compared to EPEC-wt, suggesting that the reduced Tfn clustering is not contributed by the capacity of EPEC to recruit F-actin. Inefficient Tfn/TfnR clustering in the EPEC-cesT infected cells is likely contributed by inefficient Map translocation that is partially mediated by the CesT chaperone [9,43]. The involvement of CesT-dependent effectors other than Map in the clustering effect cannot be excluded at this point. Tfn clustering at infection sites was restored upon cell infection with EPEC-espF+EspF or EPEC-map+Map (Fig 7B, S12A Fig). Infection with these strains also prompted the recruitment of Rab11a and Myo5b at infection sites (Fig 7C, S12B Fig), where EspF and Map staining partially overlapped with Rab11a and Myo5b (S12C Fig), suggesting that the effector proteins reside in close proximity to the host proteins. Simultaneous deletion of espF and map from the bacterial genome resulted in a complete loss of EPEC-stimulated recruitment of all indicated endocytic markers to the infection sites (Fig 7D, S12D Fig). Notably, both complemented EPEC strains were capable of translocating the expressed effectors into the infected cells (S13A Fig) and a fraction of translocated EspF and Map has been located in mitochondria-free regions at these sites (S13B Fig), suggesting that the translocated protein effectors may exert their activities prior to their targeting to mitochondria.
Infection with EPEC-espF+EspF and EPEC-map+Map evoked a sharp decrease in perinuclear Tfn-positive endosomal area (Fig 8A and 8B), a concomitant increase in Tfn/TfnR and Tfn/EEA1 intensity in the peripheral endosomes (Fig 8C and 8D), and an increase in Tfn endocytic turnover (Fig 8E and 8F). Together, these results suggest that upon translocation, EspF and Map promote the redistribution and clustering of Rab11a/Tfn positive endosomes at peripheral infection sites to increase the endocytic turnover at the infected plasma membrane. EspF contains three proline-rich modules that can bind the N-WASP/WASL and sorting nexin9 (SNX9) [22,23] proteins. We examined whether EPEC-espF complemented with plasmids encoding EspF mutants deficient in the interactions with N-WASP, or SNX9 (EspF mod-LA and EspF mod-RD, respectively; S3 Table) affects the ability of EspF to alter Tfn endocytosis. Results showed that infection with either bacterial strain resulted in diminished Tfn endocytosis compared to EPEC-espF + EspF or EPEC-espF+EspF mod-wt, and that is similar to the level obtained by EPEC-espF (Fig 8G). Map activates Cdc42 by a WXXXE guanine nucleotide exchange (GEF) motif [44,45,46]. Map has also been reported to possess a C-terminal TRL PDZ class I binding motif [47]. EPEC-map strains complemented with Map encoding plasmids mutated in either one of the aforementioned two motifs (S3 Table) failed to elicit increased endocytic activity (Fig 8H). Together, these data suggest a role for EspF and Map binding host proteins in eliciting Tfn endocytic turnover in response to EPEC infection.
Next, we asked whether EspF or Map can affect endosomal recruitment and traffic in polarized epithelial cells. Polarized MDCK cells co-expressing mRFP-LifeAct and GFP-Rab11a were infected with EPEC-wt, EPEC-escV, EPEC-espF, EPEC-map, or left uninfected. Cells were fixed, and the recruitment of the expressed proteins at infection sites was analyzed by confocal microscopy (Fig 9A). Data showed that compared to EPEC-wt, the recruitment of GFP-Rab11a to infection sites of EPEC-escV and EPEC-espF, but not of EPEC-map, was significantly reduced. Neither EPEC-espF nor EPEC-map had an effect on mRFP-LifeAct recruitment to these sites, suggesting again that the effects obtained in EPEC-espF infected cells were not contributed by alterations in F-actin. We then infected polarized MDCK cells co-expressing mRFP-LifeAct and GFP-Rab11a, or mRFP-LifeAct and GFP-TfnR with EPEC-espF, or EPEC-espF+EspF. The results showed diminished GFP-Rab11a and GFP-TfnR clustering in EPEC-espF compared to EPEC-espF+EspF infected cells, suggesting that the recruitment of these proteins to apical infection sites is EspF dependent (Fig 9B, upper). The clustering of Tfn internalized from the basolateral pole of the GFP-TfnR expressing cells at apical infection sites also showed EspF dependence (Fig 9B, lower). Based on previous observations, these data suggest that EspF increases Tfn-endocytic turnover and Tfn basolateral-to-apical transcytosis. Indeed, data obtained by flow cytometry (Fig 9C) and transcytosis assays (Fig 9D) concur with this notion.
Because of the significance of EspF, its interactions with host cell proteins upon translocation were examined. Using co-immunoprecipitation followed by mass-spectrometry analysis we show that aside of the previously reported N-WASP and SNX9, the proteins SNX18, SNX33 and WIPF1 were specifically co-immunoprecipitated with translocated EspF (Fig 10A and S6 Table). STRING analysis identified the possible protein-protein interaction network among these proteins (Fig 10B). The interactions between EspF and SNX, which were more efficient than the interactions with WIPF1 and N-WASP (see “Fold change” S6 Table), could also be confirmed by Western blotting (Fig 10C). Our data also showed a partial, but consistent reduction in the co-immunoprecipitation levels of the three SNX proteins with the EspF mod-RD mutant (Fig 10D), suggesting that these proteins interact with the SNX9 binding motif of EspF.
Our results show that EPEC type III secreted components stimulate the recruitment of clathrin-dependent early (Rab5a, EEA1) and recycling (e.g. Rab11a, Rab11b, Rab4a, Rab25, Myo5b, FIP2) endocytic machineries to plasma membrane infection sites of polarized and non-polarized epithelial cells (S1, S4A and S4B Figs). These rearrangements, which occur early upon EPEC microcolony contacting the host (S5 Fig), coincided with enhanced endocytosis and recycling activities, i.e. increased endocytic turnover, at that host cell surface (Fig 2 and S8 Fig), and with basolateral-to-apical transcytosis of basolaterally recycled cargoes, e.g. TfnR (Fig 1C). The consequence of these events is the enrichment of the infected plasma membrane with specific membrane proteins.
The recruitment of endosomes to infection sites is mediated by several mechanisms. From the ‘host cell perspective’, Myo5b motors are required to mobilize Rab11-positive recycling endosomes and their cargoes to host infection sites. In non-polarized cells, this process involves Myo5b-dependent shuttling of Rab11 endosomes and their cargoes from perinuclear slow recycling endosomes to peripheral infection sites (Figs 3 and 4). In polarized cells, the process involves Myo5b-dependent shuttling of basolaterally internalized cargoes to apical infection sites (Fig 1B). In both cases, the enrichment of peripheral infection sites with recycling endosomes could play an important role in facilitating endocytic turnover near these sites. This hypothesis is supported by previous studies suggesting that the shift in cargo localization from slow perinuclear to fast peripheral recycling endosomes results in an increased endocytic turnover [48,49,50]. In our studies, depletion of Rab11 in non-polarized cells caused loss of perinuclear endosomes, and increased the abundance of peripheral endosomes and endocytic turnover (S11 Fig). Infection with EPEC-wt, did not affect the endocytic turnover in these cells (Fig 6), probably because it could not further redistribute the already peripherally redistributed endosomes.
From the ‘pathogen’s perspective', we identified EspF and Map as the protein effectors which mediate endosomal recruitment to infection sites and the increase in endocytic turnover in non-polarized cells (Figs 7 and 8). Interestingly, EspF, but not Map, has been identified as mediating these activities in polarized cells (Fig 9), emphasizing the role of EspF in the process. This conclusion may coincide with recent data suggesting that EspF, but not Map, is involved in the disruption of epithelial cell polarity [19]. EspF and Map may perform these activities immediately upon translocation into the host and prior to reaching mitochondria (S13B Fig). Studies have shown that EspF interacts with high specificity and strong affinity with the eukaryotic SNX9 and N-WASP to mediate endocytic membrane remodeling coupled to Arp2/3 actin nucleation [23]. Using co-immunoprecipitation followed by proteomics and Western blotting analyses, we confirmed these interactions (Fig 10A and 10C and S6 Table). SNX9, SNX18 and SNX33, which represent a subgroup of SNX-BAR (Bin, Amphiphysin, Rvs) sorting nexins are predicted to function as modulators of endocytosis and endosomal sorting [51]. Studies have shown that SNX18 interacts with the Rab11 family interacting protein (FIP) 5, and that these interactions are required for membrane tubulation and polarized transport of apical proteins [52]. We identified SNX18, SNX33 and WIPF1 as novel interactors of EspF (Fig 10A and 10C and S6 Table). Thus, while the interactions of EspF with SNX9 and with WASP/WIPF1 may be linked to the ability of the effector protein to activate endocytosis (Figs 8G and 9C), the interactions with SNX18 could play a role in Rab11/FIP5 dependent cargo shuttling to the infected apical plasma membrane. Interestingly, EspF seems to interact with SNX18 and SNX33 through the previously identified SNX9 binding motif (Fig 10D). These SNX-EspF interactions may mediate the EspF-dependent increase in endocytic activity.
Our data also indicate that the Map’s GEF and PDZ binding domain play a role in eliciting endocytic turnover (Fig 8H). These observations are consistent with previous studies implicating Cdc42 in modulating a functional connection between the actin cytoskeleton and endocytic traffic [53,54], and in the induction of endocytic membrane turnover [55]. The significance of the PDZ binding motif of Map, which binds EBP50, could be explained by a role attributed to EBP50 in endocytic recycling [56]. Notably, Clements et al have reported that translocated EspG imposed inhibition of endosomal recycling [21] by targeting the TfnR/Rab11/Rab35 positive recycling endosomes [20]. Thus, it is possible that EspG counterbalances the endocytic and recycling activities elicited by EspF and Map.
The functional significance of these events with respect to the pathogen and host cell physiology could be broad. It would be reasonable to assume that the lumen of the infected gut, particularly following experiencing acute diarrhea, is depleted of nutrients. We suggest that upon landing on the apical cell surface of small intestinal enterocytes, the pathogen generates a local niche that enables nutrient acquisition from the serosal environment. One way of achieving this is by apical missorting of basolateral plasma membrane proteins which carry micronutrients from the blood, such as TfnR bound to iron-loaded Tfn. Indeed, bacterial pathogens have evolved remarkably efficient strategies to hijack iron from their host, a critical micronutrient for their homeostasis and growth, [57], including from transferrin [58,59]. Here we show increased EPEC colonization of host cells exposed to iron-loaded Tfn that has been introduced either to the basolateral or the apical poles of polarized MDCK-PTR9 cells. A similar phenomenon was seen when these cells were exposed to free iron. However, such an increase was not observed in parental MDCK cells, which express very low levels of the native canine receptors (S14 Fig). These data suggest that apically infecting EPEC is capable of hijacking and accessing iron bound Tfn even when the ligand is internalized from the basolateral side of the cells, likely by importing the cargo to the apically infected plasma membrane via transcytosis (Fig 1). This process seems to be vital for promoting bacterial colonization of the infected cell surface. Another possible way is the weakening of the tight junction barrier functions, which may allow the infiltration of nutrients from the blood to bacteria adhered to apical cell-cell junctions [5]. Interestingly, tight junctions disruption has been attributed to several EPEC effectors, including effectors investigated in this study, EspF and Map [10]. We suggest that through both these activities the pathogen gains a survival advantage over other bacteria (e.g. commensals), resulting in successful colonization and infection of the gut.
Previous studies have shown that infection with the murine A/E pathogen C. rodentium resulted in mislocalization of the water channels aquaporin 2 and 3 (AQP2 and AQP3) from the host cell membranes to the cytoplasm. The process, which was partially dependent on EspF and EspG, correlated with a diarrhea-like phenotype [60]. More recently, transiently expressed eGFP-AQP3 was observed at EPEC infection sites [61]. Studies have shown that similar to the TfnR, AQP2 utilizes the clathrin-dynamin dependent endocytic pathway [62,63], as well as the Myo5b and Rab11-FIP2 recycling route [64]. Moreover, AQP2 has been localized to the apical or basolateral surfaces of polarized epithelial cells [63,65], while AQP3 was localized exclusively to the basolateral surface of these cells [66,67]. Here we found that native AQP2 and AQP3 are recruited to EPEC infection sites in a T3SS-dependent manner and that both water transporters colocalized with Myo5b (S15 Fig). Thus, it is possible that similarly to TfnR, EPEC utilizes type III secreted effectors to mislocalize aquaporins to Myo5b/Rab11 recycling endosomes, a process that could alter the water homeostasis in the infected cells and thereby lead to the diarrheal effect.
To summarize, our data suggest the following model. In non-polarized cells (Fig 11A), translocated EspF and Map promote the biogenesis and nucleation of early and recycling endosomes in proximity to plasma membrane infection sites. Actin has shown to be important for endosomal membrane remodeling, endosomal dynamics and plasma membrane protein (cargo) recycling (reviewed in [68]). Thus, EspF and Map may achieve their subversive effect through interactions with actin modulators, e.g. Map-mediated activation of Cdc42 [69] and EspF interactions with N-Wasp and cortactin [23,70,71]. The end result of these activities is a local increase in endocytic turnover and enrichment of recycling plasma membrane proteins at the infection sites. In polarized epithelial cells, another layer of complexity is added (Fig 11B). Basolateral recycling plasma membrane cargoes, e.g. TfnR and AQP3 [67], are sorted, likely by EspF and yet unidentified other effector proteins, to Rab11/Myo5b-positive apical recycling endosomes hijacked to the apical infection sites. These endosomes then utilize their apical recycling capacity to target these cargoes to the infected apical plasma membrane. The consequence of this event is misrouting of basolateral plasma membrane proteins to apical recycling endosomes and plasma membrane infection sites. Some of these proteins may give a survival advantage to the pathogen, while others may disturb the host physiology. A future challenge would be to explore the molecular mechanisms by which EspF (and other protein effectors) target the polarized endocytic sorting machinery of the host, and to elucidate how this contributes to bacterial colonization and virulence.
Bacterial strains, antibodies, plasmids and reagents used in this study are listed in S1–S4 Tables, respectively. Mutation of map in the EPEC espF::kan strain was done using the λ Red system [72]. The upstream and downstream recombination sequences were PCR amplified from genomic DNA of EPEC wild-type using primer sets 1371–4197 and 1495–4198, respectively (S5 Table). A chloramphenicol cassette was amplified from pKD3 (S3 Table) using primers 1354 and 1355. A DNA fragment of Δmap::cam allele was prepared by isothermal assembly of these three PCR fragments [73] followed by electroporation into EPEC espF::kan strains containing pKD46 harboring λ Red genes (γ, β and exo). Then, the desired mutants were selected and pKD46 was cured at 42°C. The mutation was verified using PCR with flanking primers and sequencing.
Bacterial growth and pre-activation of their T3SS was performed as described [74]. HeLa cells were infected with activated bacteria (multiplicity of infection ~100) for 30 min at 37 °C in plain DMEM. Polarized epithelial cells (see below) were infected with non-activated bacteria [i.e. bacteria grown in Luria-Bertani mixed 1:50 (v/v) with plain minimal essential medium (MEM)], for 180 min at 37 °C. For HeLa cells infected with EPEC-espF+EspF or EPEC-map+Map, expression of the protein effectors was induced by adding 0.2mM isopropyl-β-D-thiogalactopyranoside (IPTG) for the last 30 min of the activation step. For polarized MDCK cells infected with these EPEC strains, expression of the protein effectors was induced by introducing 0.2mM IPTG into the medium during the last 60 min of the infection. Deviations from these conditions are indicated in the text. All infections were performed in a CO2 incubator (37°C, 5% CO2, 90% humidity).
HeLa (J. Orly; The Hebrew University of Jerusalem), Madin-Darby canine kidney (MDCK, type II; K. Mostov; University of California, San Francisco), and Caco2-BBe cells (Tet-off; J. Turner; University of Chicago; Harvard Medical School [75]) were cultured as described [13,74]. MDCK-PTR9 cells (K. Mostov, University of California, San Francisco), which stably express the human Tfn and the rabbit polymeric immunoglobulin receptors, have been described [76,77]. MDCK cells stably expressing a green fluorescent protein (GFP)-human TfnR chimera (MDCK-GFP-TfnR; The Hebrew University of Jerusalem) were generated by transfecting MDCK cells with a human GFP-TfnR encoding plasmid (S3 Table) followed by G418 selection. It was determined that ~85% of the transfected GFP-TfnR is expressed on the basolateral surface of the cells. Cell polarity was obtained by seeding the cells on Transwell filters (12-mm, 0.4 μm, #3401; Corning, Acton, MA), as described [13,74]. Transient transfections of HeLa and MDCK cells with plasmids were performed using the TransIT-X2 6000, or the Lipofectamine 2000 systems, following the manufacturers’ instructions. HeLa cells were analyzed 16 hrs after transfection. MDCK cells were seeded on Transwell filters 16 hrs after transfection and analyzed 96 hrs later.
Unless otherwise indicated, polarized epithelial cells were incubated with fluorescently tagged transferrin (Tfn; S2 Table; 5 μg/ml) administered to either the apical [Tfn (apical)] or the basolateral [Tfn (Basal)] medium during the last 60 min of the infection time. HeLa cells were exposed to the fluorescently tagged Tfn (5 μg/ml) during the 30 min infection period.
The basolateral surface of polarized MDCK-PTR9 cells was initially exposed to Tfn-AF488 (10 μg/ml) for 90 min at 37°C. Cells were then washed extensively with plain DMEM lacking phenol red (37°C). The apical surface of these cells was infected with pre-activated EPEC (in DMEM lacking phenol red) for 120 min, or left uninfected. The basolateral and apical media of the cells were collected and stored at 4°C. Cells were washed and Texas-Red (TR)-Dextran [70 KDa; 0.5mg/ml in cold (4 °C) DMEM lacking phenol red] was added to their apical surface, while the basolateral surface was kept in plain DMEM lacking phenol red. Cells were incubated for 60 min at 4°C, and the apical and basal media were collected. The fluorescence levels of Tfn-AF488 released to the apical (i.e. transcytosed) or to the basal (recycled) medium and the fluorescence levels of TR-Dextran present in the basolateral medium (cell monolayer leakiness) were determined by the Synergy H1, Hybrid Multi-mode Microplate Reader (BioTek, Winooski, VT), and presented as % change of uninfected cells.
Polarized MDCK-PTR9 cells were washed and incubated in warm (37°C) MEM containing 1% BSA for 60 min to remove cell-associated Tfn. Cells were then infected for 180 min with EPEC, or left uninfected. Unlabeled human holo-Tfn was added to the apical or basal media of the cells for the last 60 min of the infection period. The surface-bound ligand was stripped off by extensive washes in ice-cold PBS followed by cell incubation in stripping buffer [50mM 2-(N-morpholino) ethanesulfonic acid (MES; pH = 5.0), 200 mM NaCl, 100 mM deferoxamine mesylate salt] for 30 min at 17°C. Cells were lysed in ice-cold lysis buffer [10 mM Tris-HCl pH8.0, 1% v/v Triton-X100, 150 mM NaCl, 5mM EDTA, protease inhibitors cocktail], and shacked by vortex for 30 min at 4°C. Detergent-insoluble materials were removed by centrifugation (16,000 g, 10 min, 4°C). Cell lysates were analyzed for the presence of Tfn by SDS-PAGE followed by Western blotting, using rabbit anti-human Tfn antibody (S4 Table). In some experiments, the dynamin inhibitor Dynasore (80 μM) was applied apically for 60 min prior to cell exposure to Tfn, and together with Tfn during the infection period.
HeLa cells (20,000 cells) were seeded on ibiTreat μ-slide 8 well plates (Ibidi, Martinsried, Germany). One day after seeding, cells were transfected with tdEos-Rab11a (S3 Table). Sixteen hrs post-transfection cells were washed with warm PBS, and exposed to a 1:1:1 mixture of DMEM, activated EPEC and Tfn-AF647 (100 μl each) (S2 Table; 5μg/ml). Perinuclear Rab11a/Tfn positive puncta were selected and irreversibly photo-converted by exposure to a 405nm laser beam for 30 sec. Cells were then subjected to time-lapse confocal live cell imaging. Quantitative analysis of fluorescence levels at infection sites was measured relative to uninfected areas, and termed “% change of uninfected cells”.
HeLa cells were infected with EPEC in the presence of Tfn-HRP (5 μg/ml; S2 Table). Cells were rinsed in DPBS and fixed for 30 min in freshly prepared fixative containing 4% paraformaldehyde and 0.25% glutaraldehyde (Electron Microscopy Sciences, Fort Washington, PA, USA) in 0.1 M phosphate buffer (pH 7.4). After thorough washings, samples were treated with 3,3' diaminobenzidine tetrahydrochloride (DAB, 5 mg/20 ml PBS supplemented with 4 μl H2O2) for 10 min to visualize the HRP reaction product. The DAB product was further enhanced and substituted with silver/gold particles, as described [80]. Finally, the samples were postfixed in a mixture of 1% osmium tetroxide and 1.5% potassium ferricyanide in 0.1 M cacodylate buffer pH 7.0, dehydrated in ascending concentrations of ethanol and embedded in EM-BED812 (Electron Microscopy Sciences, Hatfield, PA). Ultrathin sections were lightly stained with uranyl acetate and lead citrate and examined with a Tecnai-12 TEM 100kV (Phillips, Eindhoven, the Netherlands) electron microscope equipped with MegaView II CCD camera and Analysis version 3.0 software (SoftImaging System GmbH, Münstar, Germany).
HeLa cells (~ 50% confluence) were transfected with si-Rab11a (50 nM; S2 Table), or si-Rab11b (50 nM; S2 Table), or Rab11a+b (25 nM each) for 72 hours, using the TransIT-X2 6000 system. Non-Targeting siRNA Pool #2 (50 nM; S2 Table) was used as control. Rab11 silencing was confirmed by Western blotting, using anti-Rab11a and anti-Rab11b antibodies (S4 Table).
Results are presented as means ± standard error (SE). Statistical significance was determined by two-tailed Student’s t-test. A p-value < 0.05 indicates a statistically significant difference. ***p<0.0005, ** p<0.005, * p<0.05; ns = statistically not significant, p>0.05.
|
10.1371/journal.pcbi.1003292 | Reconstructing the Genomic Content of Microbiome Taxa through Shotgun Metagenomic Deconvolution | Metagenomics has transformed our understanding of the microbial world, allowing researchers to bypass the need to isolate and culture individual taxa and to directly characterize both the taxonomic and gene compositions of environmental samples. However, associating the genes found in a metagenomic sample with the specific taxa of origin remains a critical challenge. Existing binning methods, based on nucleotide composition or alignment to reference genomes allow only a coarse-grained classification and rely heavily on the availability of sequenced genomes from closely related taxa. Here, we introduce a novel computational framework, integrating variation in gene abundances across multiple samples with taxonomic abundance data to deconvolve metagenomic samples into taxa-specific gene profiles and to reconstruct the genomic content of community members. This assembly-free method is not bounded by various factors limiting previously described methods of metagenomic binning or metagenomic assembly and represents a fundamentally different approach to metagenomic-based genome reconstruction. An implementation of this framework is available at http://elbo.gs.washington.edu/software.html. We first describe the mathematical foundations of our framework and discuss considerations for implementing its various components. We demonstrate the ability of this framework to accurately deconvolve a set of metagenomic samples and to recover the gene content of individual taxa using synthetic metagenomic samples. We specifically characterize determinants of prediction accuracy and examine the impact of annotation errors on the reconstructed genomes. We finally apply metagenomic deconvolution to samples from the Human Microbiome Project, successfully reconstructing genus-level genomic content of various microbial genera, based solely on variation in gene count. These reconstructed genera are shown to correctly capture genus-specific properties. With the accumulation of metagenomic data, this deconvolution framework provides an essential tool for characterizing microbial taxa never before seen, laying the foundation for addressing fundamental questions concerning the taxa comprising diverse microbial communities.
| Most microorganisms inhabit complex, diverse, and largely uncharacterized communities. Metagenomic technologies allow us to determine the taxonomic and gene compositions of these communities and to obtain insights into their function as a whole but usually do not enable the characterization of individual member taxa. Here, we introduce a novel computational framework for decomposing metagenomic community-level gene content data into taxa-specific gene profiles. Specifically, by analyzing the way taxonomic and gene abundances co-vary across a set of metagenomic samples, we are able to associate genes with their taxa of origin. We first demonstrate the ability of this approach to decompose metagenomes and to reconstruct the genomes of member taxa using simulated datasets. We further identify the factors that contribute to the accuracy of our method. We then apply our framework to samples from the human microbiome – the set of microorganisms that inhabit the human body – and show that it can be used to successfully reconstruct the typical genomes of various microbiome genera. Notably, our framework is based solely on variation in gene composition and does not rely on sequence composition signatures, assembly, or available reference genomes. It is therefore especially suited to studying the many microbial habitats yet to be extensively characterized.
| Microbes are the most abundant and diverse life form on the planet. Recent advances in high-throughput sequencing and metagenomics have made it possible to study microbes in their natural environments and to characterize microbial communities in unprecedented detail. Such metagenomic techniques have been used to study communities inhabiting numerous environments, ranging from the bottom of the ocean [1], [2] and the roots of plants [3], [4] to the guts of mammals [5]. In particular, human-associated microbial communities have attracted tremendous attention, with several large-scale initiatives aiming to characterize the composition and variation of the human microbiome in health and disease [6]–[8]. Such studies have demonstrated a strong link between the microbiome and the health of the host, identifying marked compositional shifts in the microbiome that are associated with a variety of diseases [9]–[12].
Using a variety of experimental techniques and bioinformatic protocols [13], [14], metagenomics-based surveys can now characterize both the taxonomic and gene composition of the microbiome. Specifically, amplicon sequencing of conserved genes, such as the 16S ribosomal RNA gene, can be used to determine the relative abundance of each taxon [13], [15]. Obtained 16S sequences are clustered into Operational Taxonomic Units (OTUs), providing a proxy for the set of taxa found in the community [14], [16]. Alternatively, shotgun metagenomic sequencing can be used to derive short sequences (reads) directly from the community without amplification [17], [18]. These reads can then be mapped to a set of reference genes or orthologous groups (e.g., those defined by KEGG [19] or COG [20]) to translate read count data into relative abundances of functional elements, representing the collective set of genes found in the microbiome.
One of the key challenges in metagenomic research is the identification of the taxonomic origin for each shotgun metagenomic read or gene and, ultimately, the reconstruction of the genomes of member taxa directly from these reads. A diverse set of methods have been developed to parse shotgun metagenomic data and to obtain insights into the underlying taxa. These methods can be largely partitioned into several distinct categories, including: alignment to reference genomes, taxonomic classification, assembly, and binning (Figure S1). For ecosystems that are well-covered by reference genomes, such as the human microbiome [6], [21], alignment to reference genomes provides a way to determine the abundance of the various strains, species, or clades in the community [6], [8], [22] and can be used to assess strain variation within and between samples [23] (Figure S1A). Taxonomic classification methods, also referred to as taxonomic or phylogenetic binning, provide a less-specific phylogenetic label to each read, usually through a more permissive alignment to known sequences in nucleotide or peptide space [24], [25] (Figure S1B). These methods can be useful for determining the abundance of specific clades and for assisting with assembly efforts. These techniques, however, are limited by the set of reference genomes available and are only useful when relatively many community members have been previously sequenced. Considering the vast diversity of microbial communities and the challenges involved in isolation and culturing efforts [17], this approach can be applied on a large scale to only very few microbial communities.
When reference genomes are not available, assembly methods can be used to link reads into contigs and scaffolds that are easier to annotate [26] (Figure S1C). Such methods have been used to reconstruct full species [27], coexisting strains [28], [29], and more generally to construct catalogues of genes specific to particular or general ecosystems [6], [8], [12], [30]. Assembly is generally limited by the fraction of reads that can be mapped due to the complexity of most communities and the low coverage of each individual genome. Consequently, de-novo assembly of complete genomes from shotgun metagenomic samples is feasible only in extreme cases of low-complexity communities, very deep sequencing, or in combination with sample filtration techniques [27], [29], [31], [32]. Binning methods similarly aim to cluster reads into distinct groups, but do not necessarily require sequences to overlap (Figure S1D). Binning methods typically partition reads based on frequencies of nucleotide patterns (k-mers) [33]–[42], but can also use abundance and similarity metrics [12], [30], [41]–[43]. While these methods utilize every read in a metagenomic sample, they have several significant shortcomings. K-mer methods, for example, are constrained by the short length of each read, the low resolution of nucleotide usage profiling, assume homogeneity of coding bias both across genomes and locally, and may not accurately discriminate highly-related organisms. Furthermore, methods based on clustering do not usually allow reads, genes, or assemblies to be assigned to more than one group, which is problematic for highly conserved regions of a genome and for mapping reads from gene catalogs that use a low threshold on sequence identity [8].
Finally, in addition to the above well-established categories, yet another category of methods for parsing metagenomic data can be defined, which we refer to here as deconvolution. Deconvolution-based methods aim to determine the genomic element contributions of a set of taxa or groups to a metagenomic sample (Figure S1E). These methods profoundly differ from the binning methods described above as a single genomic element, such as a read, a contig, or a gene, can be assigned to multiple groups. An example of such a method is the non-negative matrix factorization (NMF) approach [44]–[46], a data discovery technique that determines the abundance and genomic element content of a sparse set of groups that can explain the genomic element abundances found in a set of metagenomic samples.
In this manuscript, we present a novel deconvolution framework for associating genomic elements found in shotgun metagenomic samples with their taxa of origin and for reconstructing the genomic content of the various taxa comprising the community. This metagenomic deconvolution framework (MetaDecon) is based on the simple observation that the abundance of each gene (or any other genomic element) in the community is a product of the abundances of the various member taxa in this community and their genomic contents. Given a large set of samples that vary in composition, it is therefore possible to formulate the expected relationships between gene and taxonomic compositions as a set of linear equations and to estimate the most likely genomic content of each taxa under these constraints. The metagenomic deconvolution framework is fundamentally different from existing binning and deconvolution methods since the number and identity of the groupings are determined based on taxonomic profile data, and the quantities calculated have a direct, physical interpretation. A comparison of the metagenomic deconvolution framework with existing binning and deconvolution methods can be found in Supporting Text S1.
We begin by introducing the mathematical basis for our framework and the context in which we demonstrate its use. We then use two simulated metagenomic datasets to explore the strengths and limitations of this framework on various synthetic data. The first dataset is generated with a simple error-free model of metagenomic sequencing that allows us to characterize the performances of our framework without the complications of sequencing and annotation error. The second dataset is generated using simulated metagenomic sequencing of model microbial communities composed of bacterial reference genomes and allows us to study specifically the effects of sequencing and annotation error on the accuracy of the framework's genome reconstructions. We finally apply the metagenomic deconvolution framework to analyze metagenomic samples from the Human Microbiome Project (HMP) [6] and demonstrate its practical application to environmental and host-associated microbial communities.
Consider a microbial community composed of some set of microbial taxa. From a functional perspective, the genome of each taxon can be viewed as a simple collection of genomic elements, such as k-mers, genes, or operons. The metagenome of the community can accordingly be viewed as the union of these genomic elements, wherein the abundance of each element in the metagenome reflects the prevalence of this element in the various genomes and the relative abundance of each genome in the community. Specifically, if some genomic element is prevalent (or at least present) in a certain taxon, we may expect that the abundance of this element across multiple metagenomic samples will be correlated with the abundance of the taxon across the samples. If the abundances of both genomic elements and taxa are known, such correlations can be used to associate genomic elements with the various taxa composing the microbial community [47], [48]. In Supporting Text S1, we evaluate the use of a simple correlation-based heuristic for predicting the genomic content of microbiome taxa and find that such simple correlation-based associations are limited in accuracy and are extremely sensitive to parameter selection. This limited utility is mostly due to the fact that associations between genomic elements and taxa are made for each taxon independently of other taxa, even though multiple taxa can encode each genomic element and may contribute to the overall abundance of each element in the various samples. We therefore present here a statistical deconvolution framework, improving upon the simple correlation metric and developing a mathematical model of shotgun metagenomic sequencing. This model quantifies the associations between a genomic element found in a set of samples and all the taxa in the community simultaneously, providing an estimate for the prevalence of this element in the genome of each taxon. Such statistical approaches have proven successful in analyzing gene expression data, allowing, for example, to deconvolve microarray data from mixed tissue samples into cell type-specific expression profiles [49].
Formally, if denotes the abundance of genome k in the community and denotes the prevalence of an element j in genome k (e.g., in terms of copy number or length in nucleotides), the total abundance of this element in the community can be represented as:(1)Note that similar models have been used as the basis for simulating shotgun metagenomic sequencing [50]–[53], and the total abundance of the element in the community is independent of the individual genome sizes. Now, assume that the total abundances of genomic elements, , can be determined through shotgun metagenomic sequencing, and that the abundances of the various genomes, , can be obtained using 16S sequencing or from marker genes in the shotgun metagenomic data [54], [55]. Accordingly, in Eq. (1) above, the only terms that are unknown are the prevalence of each genomic element in each genome, , and these are the specific quantities required to functionally characterize each taxon in the community.
Clearly, if only one metagenomic sample is available, Eq. (1) cannot be used to calculate the prevalence of the genomic elements . However, assume M different metagenomic samples have been obtained, each representing a microbial community with a somewhat different taxonomic composition. For each genomic element, we can now write a system of linear equations of the form:(2)…or more compactly as(3)where the subscript i denotes the sample and is a normalization coefficient (see below). Given enough samples, , the prevalence of a given genomic element j in each taxon, can be analytically solved by linear regression. Repeating this process for all genomic elements found in the community, we can therefore obtain an estimate of the prevalence of each element (e.g. each gene) in each taxon, effectively reconstructing the genomic content of all community members.
The normalization constant represents, technically, the total amount of genomic material in the community. Clearly, is not known a priori and in most cases cannot be measured directly. Assume, however, that some genomic element is known to be present with relatively consistent prevalence across all taxa in the community. Such an element can represent, for example, certain ribosomal genes that have nearly identical abundances in every sequenced bacterial and archaeal genome (see Methods). We can then rewrite Eq. (3) in terms of this constant genomic element, with a total abundance in sample i, :(4)Assuming that the taxonomic abundances have been normalized to sum to 1, this simplifies to(5)We can accordingly substitute in Eq. (3) with this term, obtaining a simple set of linear equations where the only unknown terms are the prevalence of each genomic element in each taxon, .
Metagenomic deconvolution is a general framework for calculating taxa-specific information from metagenomic data. Notably, this framework is modular, comprising four distinct components: (i) determination of taxonomic composition in each sample, ; (ii) determination of the abundances of genomic elements in each sample, ; (iii) selection of a constant genomic element, ; and (iv) calculation of the taxa-specific genomic element abundances, , by solving Eq. (3). Each of these components can be implemented in various ways. For example, different metagenomic techniques, sequence mapping methods, and annotation pipelines can be used to determine the abundance of various genomic elements in each sample. Genomic elements can represent k-mers, motifs, genes, or other elements that can be measured in the samples and whose taxonomic origin are unknown. Similarly, there are multiple regression methods that can be applied to solve the set of equations obtained and to estimate , including least squares regression, non-negative least squares regression, and least squares regression with L1-regularization (e.g., lasso [56]). Finally, the taxonomic abundances need not be derived necessarily from 16S sequencing but can rather be determined directly from metagenomic samples [54], [55].
In this study, we used gene orthology groups (which we will mostly refer to simply as genes), specifically KEGG orthology groups (KOs) [19], as the genomic elements of interest in Eq. (3) above. In this context, we defined the abundance of a KO in a metagenomic sample, , as the number of reads mapped to this KO, and the prevalence of a KO in a genome, , as the number of nucleotides encoding it in the genome. We accordingly applied our deconvolution framework to predict the length of each KO in each genome, ultimately obtaining ‘reconstructed’ genomes in the form of a list of all the KOs present in a genome and their predicted lengths.
We used the 16S rRNA gene as the constant genomic element, , to calculate the normalization coefficient . The length of the 16S gene is largely consistent across all sequenced archaeal and bacterial strains. When the abundances of the 16S gene across shotgun metagenomic samples, , are not available, other genes or groups of genes with a consistent length across the various taxa can also be used. Specifically, in applying our framework to metagenomic samples from the HMP below, we used a set of bacterial and archaeal ribosomal genes to estimate (Methods). Finally, we used least squares and non-negative least squares regression to solve Eq. (3) and to estimate (Methods). Notably, such regression techniques require that there are at least as many samples as taxa in order for there to be a solution. However, if there are fewer samples than taxa, regularized regression techniques, such as the lasso [56], can be used. For each dataset presented in this manuscript, we have evaluated the solutions presented by these regression methods and compared their accuracies across the different datasets in Supporting Text S1.
Notably, in many cases, our key goal is to determine which genes are present in (or absent from) a given genome, rather than their exact length (e.g., in nucleotides) in this genome. To predict the presence or absence of a gene in a genome, we used a simple threshold-based method. Specifically, we compared the predicted length of each gene to the average length of this gene across sequenced genomes. Genes for which the ratio between these two values exceeded a certain threshold were predicted to be present. For example, we could predict that a gene is present in a genome if it is predicted to have a length greater than 0.5 the average length of all sequenced orthologs of this gene. This method will also allow us to correct for inaccuracies in length predictions. In the results reported below, we further demonstrate the robustness of reconstructed genomes to threshold value selection.
We first use a simple model of metagenomic sampling to characterize metagenomic deconvolution in the absence of sequencing and annotation errors. To this end, we simulated microbial communities composed of 60 “species” of varying abundances (see Methods). In this model, each species was defined as a collection of “genes” assigned randomly from a total set of 100 gene orthology groups. These genes had no sequence, were assumed to vary in length, and could be present in multiple copies in each genome. 100 model microbial communities were generated with different, but correlated, abundances for each member species (Figure S2). The relative abundances of each species in the communities were assumed to be known (e.g. from targeted 16S sequencing). Metagenomic samples consisting of 5M reads were generated, simulating shotgun sequencing through a random sampling process weighted by the relative abundance of each gene in the community. Reads were assumed to map without error to the appropriate orthology group, counting towards the observed relative abundance of each gene orthology group in the sample. Full details of the model are given in the Methods.
We applied the deconvolution framework described above to predict the length of each gene in each species. Examining the predicted length of a typical gene across all species, we found that we successfully predicted the actual genomic length of this gene among the different species (Figure 1A). Similarly, comparing the predicted lengths of all genes in a typical species to the species' actual genome, we find that our framework accurately reconstructed the genomic content of the species, successfully identifying absent genes and correctly estimating a wide range of gene lengths (Figure 1B). Furthermore, analysis of the predictions obtained for all genes and for all species in the community clearly demonstrates that the metagenomic deconvolution framework can effectively reconstruct gene lengths across all genomes, orthology groups, and copy numbers (Figure 1C).
Clearly, the predicted gene lengths described above, while accurate, are not perfect, and may be affected by various sources of noise in the data. Moreover, as noted above, in many cases, we are primarily interested in predicting whether a gene is present in a certain genome rather than in determining its exact length. Converting the predicted gene lengths to gene presence/absence predictions using a threshold of 0.5 of the gene length, we find that we are able to correctly predict the presence and absence of all genes in all species with 100% accuracy. We further confirmed that this result is robust to the specific threshold used, with all thresholds values between 0.2 and 0.8 yielding perfect predictions (Figure S3).
Predictions of a given gene's length across the species vary in accuracy from gene to gene, with some genes having a noticeably higher overall error than others (Figure 1C). By examining the distribution of genes among samples and species, we find that prediction accuracy for a gene is significantly correlated with its level of variation across samples (Figure 2A) and across species (Figure 2B), with more variable genes having lower prediction error on average. These patterns in prediction accuracy are not surprising. Since our framework is based on detecting species and gene abundances that co-vary, highly variable genes or species carry a stronger signal and lead to more accurate predictions.
Interestingly, however, this seemingly limiting link between prediction accuracy and variation is one of the strengths of our framework, as it provides better accuracy for predicting exactly the genes that are of most interest. Specifically, genes that vary from species to species are those that confer species-specific functional capacity and are those that are most crucial for characterizing novel genomes. Similarly, genes that vary most from sample to sample are those endowing each community with specific metabolic potential and are therefore often of clinical interest. In contrast, genes with little variation from species to species and from sample to sample are likely to include many housekeeping genes, whose presence in each genome is not surprising and can mostly be assumed a priori.
Clearly, many microbial communities exhibit high species diversity and are inhabited by an extremely large number of species, challenging deconvolution efforts. Moreover, the abundances of species across samples are not independent: In a given environment, some species may dominate all samples, while other species may tend to be rare across all samples. Interactions between species may also introduce correlations between the abundances of various species. These inter-sample and inter-species correlations might also affect our ability to correctly deconvolve each member species, as they in effect reduce the level of variation in the data. For example, species with highly correlated abundances (e.g., the set of dominant species across all samples) will contribute similarly to the abundances of genes in the various samples and will be hard to discriminate. To explore the impact of the number of species in the community and of correlations between species abundances on metagenomic deconvolution, we used an additional set of simulated communities. Specifically, metagenomic samples were generated with a varying number of species and a varying level of inter-sample correlation in species abundances (Methods; Figure S4). We find, as expected, that the accuracy (Figure S5A) and recall (Figure S5B) of deconvolution decreases as the number of species increases (assuming a constant sampling depth). Furthermore, increasing the level of correlation between species abundances across samples similarly results in reduced accuracy and recall (Figure S5).
The simple model presented above allowed us to explore the metagenomic deconvolution framework in ideal settings where reads are assumed to be error free and to unambiguously map to genes. We next set out to examine the application of our framework to synthetic metagenomic samples that incorporate both next-generation sequencing error and a typical metagenomic functional annotation pipeline. To this end, we simulated metagenomic sampling of microbial communities composed of three reference genomes (Methods). We specifically focused on strains that represent the most abundant phyla in the human gut, as determined by the MetaHIT project [8], and for which full genome sequences were available. Furthermore, these strains represented different levels of coverage by the KEGG database (which we used for annotation), ranging from a strain for which another strain of the same species exists in the database, to a strain with no member of the same genus in the database (Methods).
Ten communities with random relative strain abundances were simulated. The relative abundances in each community were assumed be to known through targeted 16S sequencing. For the analysis below, relative abundances ranged over a thousand-fold, but using markedly different relative abundance ratios had little effect on the results (see Supporting Text S1). Shotgun metagenomic sequencing was simulated using Metasim [50], with 1M 80-base reads for each sample and an Illumina sequencing error model (Methods). The abundances of genes in each metagenomic sample were then determined using an annotation pipeline modeled after the HMP protocol [47], with reads annotated through a translated BLAST search against the KEGG database [19]. To assess the accuracy of this annotation process and its potential impact on downstream deconvolution analysis, we first compared the obtained annotations to the actual genes from which reads were derived. Overall, obtained annotation counts were strongly correlated with expected counts (0.83, P<10−324; Pearson correlation test; Figure S6). Of the reads that were annotated with a KO, 82% were annotated correctly. Notably, however, only 62% of the reads originating from genes associated with KOs were correctly identified and consequently the read count for most KOs was attenuated. Highly conserved genes, such as the 16S rRNA gene, were easily recognized and had relatively accurate read count (Figure S6). Full details of this synthetic community model and of the sequencing simulations are provided in Methods.
We deconvolved each KO using the obtained abundances to predict the length of each KO in each genome. We found that the predicted lengths were strongly correlated with the actual lengths (rho 0.84, P<10−324; Pearson correlation test), although for most KOs predicted lengths were shorter than expected (Figure 3). This under-prediction of KO lengths can be attributed to the normalization process. Specifically, as noted above, the detected abundances of conserved genes used for normalization tended to be less attenuated by the annotation pipeline than the abundances of other genes, which were therefore computed to be shorter than they actually were. Notably, some KOs that are in fact entirely absent from the genomes under study were erroneously detected by the annotation pipeline and consequently predicted to have non-negligible lengths in the reconstructed genomes (Figure 3). To discriminate the error stemming from the annotation pipeline from error stemming directly from the deconvolution process, we reanalyzed the data assuming that each read was correctly annotated. We found that with the correct annotations, predicted KO lengths accurately reflected the actual length of each KO in each genome (rho 0.997, P<10−324; Pearson correlation test; Figure S7).
Importantly, while the error introduced by the annotation pipeline significantly affects the accuracy of predicted KO lengths, the presence (or absence) of each KO in each genome can still be successfully predicted by the threshold approach described above (Figure 4A). Specifically, using a threshold of 0.1 of the average length of each KO, metagenomic deconvolution reached an accuracy of 89% (correctly predicting both KO presence and absence) and a recall of 98% across the various genomes. Figure 4B further illustrates the actual and predicted genomic content of each strain, demonstrating that the method can accurately predict the presence of the same KO in multiple strains, highlighting the difference between the metagenomic deconvolution framework and existing binning methods (see also Discussion). We compared these predictions to a naïve ‘convoluted’ prediction (see Methods), confirming that deconvolution-based predictions were significantly more accurate than such a convoluted null model regardless of the threshold used (P<10−324, bootstrap; Figure 4A). For example, using a threshold of 0.1 as above, convoluted genomes were only 54% accurate. Considering the determinants of prediction accuracy described above, we further confirmed that prediction accuracy markedly increased for highly variable and taxa-specific genes (Supporting Text S1).
Given the noisy annotation process, we again set out to quantify the contribution of annotation inaccuracies to erroneous presence/absence predictions in the reconstructed genomes. As demonstrated in Figure 4B, most KO prediction errors were false positives – KOs wrongly predicted to be present in a strain from which they were in fact absent. Examining such KOs and the annotation of reads in each genome, we found that 99% of the false positive KOs were associated with mis-annotated reads, suggesting that deconvolution inaccuracies in these settings could be attributed almost entirely to erroneous annotation rather than to the deconvolution process itself. We again confirmed that when correct annotations are assumed, both accuracy and recall increase to more than 99%.
The analysis above was used to evaluate the impact of sequencing and annotation error on the metagenomic deconvolution framework using simulated metagenomic datasets generated from simple 3-strain communities. In Supporting Text S1, we further present a similar analysis, using simulated metagenomic samples generated from 20-strain communities and based on the HMP Mock Communities. We show that our framework obtains similar reconstruction accuracies for these more complex communities (Figure S8).
Finally, we considered human-associated metagenomic samples to demonstrate the application of the metagenomic deconvolution framework to real metagenomic data from highly complex microbial communities. These datasets further represent an opportunity to evaluate genome reconstructions obtained by our framework owing to the high-coverage of the human microbiome by reference genomes [6], [21] that can be used for evaluation. The Human Microbiome Project [6], [14] has recently released a collection of targeted 16S and shotgun metagenomic samples from 242 individuals taken from 18 different body sites in an effort to comprehensively characterize the healthy human microbiome. These human-associated microbial communities are diverse, with several hundred to several thousand 16S-based OTUs (operational taxonomical units clustered at 97% similarity) per sample and a total of more than 45,000 unique OTUs across all HMP samples. These OTUs represent bacteria and archaea from across the tree of life, including many novel taxa [57], and their diversity is in agreement with shotgun metagenomics-based measures [6]. Clearly, the high number of unique OTUs in each sample does not permit deconvolution and genome reconstruction at the OTU level. Moreover, these OTUs do not represent individual species, but rather distinct sequences accurate to only a genus-level phylogenetic classification [6]. Examining the phylogenetic distribution of the taxa comprising the microbiome suggests that certain body sites, such as the tongue dorsum, are dominated by relatively few genera. This allows us to use metagenomic deconvolution at the genus level, predicting the most likely genomic content of the various genera found in the microbiome. Reconstructed genus-level genomes can be viewed as the average genomic content across all present strains in the genus, providing insight into the capacities of the various genera. Moreover, while many species inhabiting the human microbiome have not yet been characterized or sequenced, most human-associated genera include at least a few fully sequenced genomes, allowing us to assess the success of our framework and the accuracy at which reconstructed genera capture known genus-level properties. Notably, however, microbial communities from other environments or from other mammalian hosts often harbor many uncharacterized taxa, even at levels higher than genera [58], [59], making a genus-level deconvolution a still biologically relevant goal.
We accordingly applied our deconvolution framework to HMP tongue dorsum metagenomic samples (Methods). OTU abundances and taxonomic classification were obtained from the HMP QIIME 16S pipeline [14]. KO abundances were obtained from the HMP HUMAnN shotgun pipeline [19]. In total, 97 tongue dorsum samples had both OTU and KO data available. OTUs were pooled to calculate the relative abundance of each genus in each sample. After pooling, we identified 14 genera that dominated the tongue dorsum. We deconvolved these samples to obtain reconstructed genera and computed KO presence/absence in each reconstructed genus using a threshold of 0.25 copies.
To evaluate our predictions, we calculated the similarity between the 14 reconstructed genera and every sequenced genome from these genera (Methods). We find that 12 of the 14 reconstructed genera are most similar to genomes from the correct genus (Figure 5A). Interestingly, Capnocytophaga, one of the two reconstructed genera that did not most closely resemble genomes from its own genus, was the least abundant genus and appeared to be most similar to genomes from the Fusobacterium genus, with which it significantly co-occurs in the tongue dorsum [60]. This potentially reflects the sensitivity of deconvolution to highly correlated taxonomic abundances (see Discussion). Furthermore, overall, the observed similarities between each reconstructed genus and sequenced genomes from other genera (Figure 5A) largely reflect inter-genus similarities between the genomes from the various genera (Figure 5B). For example, although the reconstructed Prevotella is most similar to sequenced genomes from the Prevotella genus, it also exhibits high similarity to genomes from Porphyromonas and Capnocytophaga, two other genera from the Bacteroidetes phylum with relatively similar genomic content. These findings suggest that our deconvolution framework was able to accurately capture the similarities and the differences between the various genera based solely on variation in KO and OTU abundances across samples.
To further study the capacity of genus-level deconvolution to reconstruct and characterize the various genera in the microbiome, we next focused on the set of genes that best distinguish one genus from the other. Clearly, even within a genus, the set of genes present in a genome varies greatly from species to species and from strain to strain. Yet, for each genus, a small number of genes that are present in almost every genome from that genus and that are absent from most other genomes can be found. These genus-specific genes best typify the genus, potentially encoding unique genus-specific capacities. Moreover, since such genes are consistently present or consistently absent within each genus, genus-level deconvolution is not complicated by the genus-level pooling of genomes. We defined genus-specific KOs as those present in 80% of the genomes from a given genus and in less than 20% of all others HMP reference genomes. We found in total 99 such KOs across 4 genera. Examining the reconstructed genera, we found that our framework successfully predicted the presence or absence of these genus-specific KOs (90% accuracy and 82% recall; Figure 6). Increasing the stringency for our definition and focusing on the 63 KOs that appeared in 90% of the genomes from a certain genus and in less than 10% of all others genomes further increased the accuracy (92%) and recall (94%) of our reconstructed genera. Predictions obtained using alternative regression methods were similarly accurate (see Supporting Text S1; Figures 6, S9).
The metagenomic deconvolution framework introduced in this manuscript is a technique for associating genomic elements found in shotgun metagenomic samples with their taxa of origin and for reconstructing the genomic content of the various taxa comprising the community. Many different approaches have been developed to create such groupings of metagenomic features. Broadly, these methods fall into one of two categories, “binning” or “deconvolution”, depending on whether the genomic elements can be assigned to more than one group or not. As demonstrated in Supporting Text S1 (and see also Table S2), the differences between the metagenomic deconvolution framework and these existing methods originate primarily from the different mathematical frameworks employed by the various methods.
Binning methods, such as metagenomic linkage analysis [12], metagenomic clustering analysis [30], and MetaBin [43], are designed to cluster genomic elements that can only exist in one taxon (or group). Specifically, metagenomic linkage analysis clusters genes into groups based on their abundances and phylogeny across sets of metagenomic samples using the CHAMELEON algorithm [61]. Similarly, metagenomic clustering analysis clusters genes into groups based on their abundances across sets of metagenomic samples using the Markov clustering algorithm [62]. MetaBin, on the other hand, clusters individual reads based on their sequence similarities and abundances across sets of metagenomic samples using k-medoids clustering. As these methods all cluster genomic elements into distinct groups, they cannot correctly distribute elements that exist in multiple taxa (or groups), making them less appropriate for addressing questions of core vs. shared genome content (and see, for example, refs [54], [63], [64]). As we demonstrate in Supporting Text S1, these methods accordingly could not be used to reconstruct the genomic content of the three strains present in the simulated metagenomic samples incorporating sequencing and annotation error in terms of the gene orthology groups identified in the samples.
In contrast, deconvolution methods, such as non-negative matrix factorization (NMF) [44]–[46] and the proposed metagenomic deconvolution framework, are designed to assign genomic elements to multiple taxa. Specifically, NMF is a data discovery and compressed sensing tool that is designed to create a set number of groupings of elements that best fits the observed samples by factoring the feature matrix (here, the genomic elements found across a set of metagenomic samples) into two matrices. One matrix represents the abundance of the set of groups in each sample, and the other represents the distribution of genomic elements among these groups. The optimal number of groups can be determined from the fit of the matrix factorization to the original matrix [45], [46] or the stability of the solutions for a given number of groups [44]. Importantly, while NMF utilizes a mathematically similar approach to the metagenomic deconvolution framework, and can thus theoretically obtain comparable accuracies (see also Supporting Text S1), the two represent fundamentally different techniques. First, the groups identified by NMF are unlabeled, while those used by the metagenomic deconvolution framework by definition have a distinct taxonomic identity. Furthermore, the optimal number of groups detected in a set of samples by NMF does not necessarily correspond to any phylogenetic groupings present in the set of samples. Indeed, NMF does not group the gene orthology groups present in the simulated metagenomic samples incorporating sequencing and annotation error into strain-specific groupings (Supporting Text S1). Second, in the metagenomic deconvolution framework, the calculated quantities of genomic elements in each group have a direct physical interpretation (i.e. gene length or copy number), while NMF calculates coefficients without assigning a clearly interpretable meaning. Lastly, NMF functions on the entire set of genomic elements present in a set of samples (the feature matrix) as a whole, whereas the metagenomic deconvolution framework solves for the distribution of each genomic element among the various groups independently. This separability allows for custom regression techniques to be used for each genomic element (for example, regularized regression like lasso can be used for those genomic elements that are sparsely distributed) and the option to target only those genomic elements of interest.
In this study, we presented a novel framework for deconvolving shotgun metagenomic samples and for reconstructing the genomic content of the member microbial taxa. This metagenomic deconvolution framework utilizes the magnitude by which abundances of taxa and of genomic elements co-vary across a set of metagenomic samples to identify the most likely genomic content of each taxon. Above, we have described the mathematical formulation of this framework, detailed computational considerations for implementing it, characterized its performance and properties on synthetic metagenomic datasets, and demonstrated its practical use on metagenomic samples from the Human Microbiome Project.
The metagenomic deconvolution framework represents a fundamentally different approach to associating genomic elements found in shotgun metagenomic samples with the taxa present than the approaches employed by previously introduced methods. For example, methods relying on alignment to reference genomes [6], [8], [22], [24], [25] are heavily dependent on the availability of sequenced genomes from community members or from closely related species. As metagenomics research expands and researchers set out to characterize new environments inhabited by many novel, diverse, and never before seen species, such methods may be challenged by the scarcity of reference genomes and by the low phylogenetic coverage of many genera across genomic databases. In contrast, our method does not require reference genomes (see also below). Moreover, metagenomic deconvolution uses a mathematical model of shotgun sequencing to directly calculate the desired quantities of genomic elements (such as gene lengths or copy numbers) in specific taxa (such as a strain or genus), rather than to create groupings of elements that best fit the measured distribution.
Metagenomic deconvolution associates genomic elements with genomes of present taxa by identifying genomic elements that co-vary in abundance with organisms. As demonstrated above, this approach brings about an important advantage: The more variation of a given genomic element across samples and organisms, the more accurately it will be assigned to the various taxa. The deconvolution framework can accordingly be thought to be tuned to best identify those elements that make a taxon or a set of samples unique and that are therefore of most biological interest. Moreover, to a large extent, in analyzing the way gene and taxonomic abundances co-vary across the set of samples under study, it utilizes orthogonal, self-constrained information. Notably, the specific implementation presented in this study utilizes functional read annotation and therefore required a set of annotated reference genes. However, functional annotation is markedly less sensitive to the specific set of reference genomes available than the methods discussed above, since any gene with detectable homology will suffice. Moreover, one can easily imagine a different implementation that clusters the reads contained in the samples themselves without identifying specific orthology groups, making this approach entirely independent from any exogenous genomic data (see also below). These properties of metagenomic deconvolution make it an ideal framework for analyzing metagenomic samples from the many microbial habitats yet to be extensively characterized.
A deconvolution-based framework also has some obvious limitations. First, it requires multiple metagenomic samples and information on both taxonomic and gene abundances. While this may have been a significantly limiting factor in the past, with the ever decreasing cost of sequencing technologies and the recently introduced advances in molecular and computational profiling of taxonomic and gene compositions, current studies in metagenomics often generate such data regardless of planned downstream analyses (e.g., [6], [7]). Furthermore, if a genomic element is known to be sparsely distributed among the taxa in a collection of samples, then regularized regression techniques, such as the lasso [56], can be used to predict the presence and absence of the genomic element among the taxa, even if the number of samples is much smaller than the number of taxa. Additionally, as demonstrated above, strong correlations between taxa abundances reduce the amount of variation, decreasing the signal and potentially hindering the accuracy of the deconvolution process. Improved understanding of the assembly rules that give rise to such correlations may help alleviate this problem. Finally, our framework relies on accurate estimations of gene and taxonomic abundances. These estimations may be skewed by annotation errors or by the specific method used to evaluate relative taxonomic abundances. Specifically, 16S copy number variation between taxa in a sample (even between strains of the same species [65]) may markedly bias abundance estimates, although this can largely be resolved by estimating the 16S copy number in each taxon using measured copy numbers in sequenced strains [66]. No such correction was performed in this study, as we sought to present a generic implementation of the metagenomic deconvolution framework applicable to analyzing sets of metagenomic samples without the need for coverage by reference genomes.
The deconvolution framework presented in this study can serve as a basis for many exciting extensions and can be integrated with other analysis methods. It is easy, for example, to redefine the scale at which both genomic elements and taxa are defined. In analyzing the HMP samples, we partition genes among genera, rather than into individual OTUs. A similar approach can be used to deconvolve higher or lower (e.g., strain) phylogenetic levels or even to deconvolve different taxa at different phylogenetic levels. One can, for example, target particular species for genome reconstruction while resolving others only on the genus level. Similarly, deconvolution can be performed for other genomic elements such as k-mers or other discrete sequence motifs. Deconvolution can also be carried out incrementally, first deconvolving highly abundant taxa or taxa for which partial genomic information is available. The expected contribution of each deconvolved taxon to the overall gene count in the metagenome can then be calculated and subtracted computationally from each sample, effectively generating lower complexity samples and facilitating the deconvolution of additional taxa. A similar approach can also be used to subtract the contribution of fully sequenced strains whose genomic content is known. Notably, in implementing and characterizing the deconvolution framework here, we did not utilize any information about known strains' genomes. Such information can be used in principal to calibrate various parameters and to normalize the obtained results. Most importantly, this metagenomic deconvolution framework can be naturally combined with other binning methods or metagenomic assembly efforts [67]. For example, by treating contigs, or groups of contigs (such as those generated by metagenomic linkage groups [12] or metagenomic clusters [30]) as individual genomic elements ( and , Eq. 3), deconvolution can be used to assign these larger-scale genomic fragments to individual taxa and aid in assembly. Such a process would be especially useful in the case of time-series data where the abundances of strains change with time. Finally, the metagenomic deconvolution framework facilitates novel analysis approaches for studying microbial communities. Samples taken from a community can be post-processed in multiple ways to preferentially select for certain taxa (e.g. filter microbes by size or nutrient requirements), essentially creating different views of the same community. Deconvolution can then be used to recombine these views and to reconstruct the genomic content of each taxon.
To truly take advantage of the data being produced by metagenomic studies and by forthcoming studies of the metatranscriptome and metametabolome of many microbial communities, tools that can reliably determine the taxonomic origins of each “meta'omic” element are crucial. Metagenomic deconvolution represents both a novel strategy for the analysis of such meta'omic data and a framework for future developments in genome reconstruction and annotation.
Simple models of metagenomic samples were created from collections of model “microbial species” by simulating genomes and shotgun sequencing without the complexities of actual genome sequences or sequencing error. Microbial species were modeled as a set of “genes”, taken from a global set of 100 gene orthology groups (simply referred to as genes). These genes had no sequence; their only property was length, which was chosen at random between 400 and 500 bases, and was fixed across all homologs. Simulations with species-specific variations in gene length showed qualitatively similar results (see Supporting Text S1). Each of the 100 genes was randomly assigned to between 20 to 80% of the species, with each species containing a minimum of 10 genes. Within a given species, each gene had a 5% chance of duplication, with the rates for higher copy number decreasing exponentially. Each species included a single copy of a “constant gene” with a length of 1500 bases (see Results).
Sets of model “microbial communities” were created as a linear combination of model microbial species. Each microbial community in a set had a different, but correlated, species abundance profile, with the abundance of a species j in sample i, determined by the function, , where represents the typical abundance of species j, v is a parameter that governs the amount of inter-sample correlation in the abundance profiles and rij is a Gaussian-distributed random number with mean of 0 and standard deviation of 1. To examine the robustness of deconvolution to the number of species and the level of inter-sample correlation, 30 different sets of related communities were created, with the number of species ranging from 20 to 100 in steps of 20, and the correlation parameter v logarithmically distributed, (Figure S4). The set of communities analyzed in the main text was modeled with 60 species and a correlation parameter of v = 0.10.
Model metagenomic samples were generated from each microbial community by simulating a shotgun sequencing sampling: Sequencing reads were created by randomly selecting a gene in the community, weighted by the relative abundance of each gene in the community (Eq. 3). 5M sequencing reads were generated for each community. Due to the finite sequencing depth and the exponentially distributed species abundances, species whose abundances were below 0.5% of the most abundant species in the sample were considered absent from the set of shotgun metagenomic reads and excluded from our analysis. These samples and the related data can be found in Supporting Dataset S1 and on our website (http://elbo.gs.washington.edu/download.html).
Deconvolution was performed for species that were present in at least half the samples using least squares, non-negative least squares, and lasso regression using the solvers implemented in MATLAB. The computation times for these deconvolution runs on a four-core 3.10 GHz Intel Xeon CPU were 2±1×10−4 s/gene, 4.6±0.8×10−3 s/gene, and 1.63±0.05 s/gene for least squares, non-negative least squares, and lasso regression respectively. Adding additional samples required 8×10−7, 7×10−7, and 0.9 s/gene/sample for least squares, non-negative least squares, and lasso (for underdetermined systems) regression, respectively; for overdetermined systems, lasso had a performance increase of 1.7×10−2 s/gene/sample. Adding additional species required 2×10−6, 7×10−5, and 4×10−2 s/gene/species for least squares, non-negative least squares, and lasso regression, respectively.
Simple models of metagenomic samples were created from the fully sequenced genomes of microbial reference organisms to introduce the complexities associated with actual genome sequences and annotation error. 10 model communities were composed as linear combinations of the reference organisms Alistipes shahii WAL 8301, Ruminococcus champanellensis sp. nov., and Bifidobacterium longum longum F8. These strains were chosen because they each had a different level of coverage by the KEGG database used in this study (see below): B. Longum had a different strain of the same species present in the database; R. Champanellensis had only a member of the same genus present; and A. Shahii had no relatives within the same genus present. Complete species genomes were obtained from the Integrated Microbial Genomes database [68]. These communities had species relative abundances assigned randomly, ranging over a thousand-fold; however, the magnitude of the range of relative abundances was shown to have little impact on our results (Supporting Text S1). Model metagenomic samples were created from each community by simulating 1M shotgun metagenomic sequencing reads with Metasim [50], using 80-base reads with an Illumina sequencing error model.
The abundances of gene orthology groups present in each model metagenomic sample were determined from the set of reads by annotating each read with KEGG orthology groups (KOs) through a translated BLAST search against the KEGG Orthology v60 [19]. Reads were annotated with the KO of the best hit with an E-value<1, similar to the method employed by the HMP [6]. Reads with a best-hit match to a KEGG gene without a KO annotation were not assigned a KO. In cases of e-value ties, the read was assigned the annotations of all the tied matches, with each annotation receiving a fractional count. Reads containing an ambiguous base were not annotated. The abundance of the 16S rRNA KO was determined through a nucleotide BLAST search against a custom database containing the sequences of all 16S rRNA genes in the KEGG database. These samples (as well as the 20 strain community samples) and the related data can be found in Supporting Dataset S2, Supporting Dataset S3, and on our website (http://elbo.gs.washington.edu/download.html).
Deconvolution was performed using least squares, non-negative least squares, and lasso regression for KOs whose average count was greater than 0.1% of the most abundant KO using the solvers implemented in MATLAB. The computation times for these deconvolution runs on a four-core 3.10 GHz Intel Xeon CPU were s/KO, s/KO, and s/KO for least squares, non-negative least squares, and lasso regression respectively.
To evaluate the presence/absence prediction made by our framework, we used a null model in which community members are all assumed to have an identical (‘convoluted’) genome, directly derived from the set of metagenomic samples. Specifically, the KO lengths in this model corresponded to the average relative abundance of each KO across all samples, normalized by the length and abundance of the 16S KO. Formally, the length of KO j, , was calculated as , where is the average relative abundance of KO j across all metagenomic samples, is the average length of the 16S KO, and is the average relative abundance of the 16S KO.
HMP data was downloaded from the HMP Data Analysis and Coordination Center (DACC) (http://www.hmpdacc.org/). OTU abundances and taxonomy were based on the QIIME 16S pipeline [14]. The abundance of each genus was calculated by adding the abundances of all OTUs in that genus. Only genera with relative abundance >5% in at least one sample were considered. KO abundances were based on the HUMAnN pipeline [19]. For samples with technical replicates, the replicate with the greater sequencing depth was used. To reduce annotation error, only KOs present in at least 80% of the tongue dorsum samples were used in the analysis. Since HMP KO abundance data included only proteins, we used a set of 15 ribosomal proteins ubiquitous across Bacteria and Archaea instead of the 16S RNA gene as the constant genomic element in Eq. 5 (see below). Deconvolution was performed for KOs that were present in at least half the samples using least squares, non-negative least squares, and lasso regression using the solvers implemented in MATLAB. The computation times for these deconvolution runs on a four-core 3.10 GHz Intel Xeon CPU were s/KO, s/KO, and s/KO for least squares, non-negative least squares, and lasso regression respectively.
Genomes for the HMP Reference Organisms were obtained from the Integrated Microbial Genomes – Human Microbiome Project (IMG/HMP) database on 5/7/2012 (http://www.hmpdacc-resources.org/cgi-bin/imgm_hmp/main.cgi). In order for the annotations to be compatible with the version of the database used in this study, each organism was annotated through a BLAST search of each ORF against the KEGG genes database with a protocol similar to that used by the IMG [68]. Each ORF was annotated with the KO of the best match gene with an e-value <1×10−5. In cases of ties, the ORF was annotated with all corresponding KOs, with a proportionally fractional count. ORFs that best matched a KEGG gene with no KO annotation were not annotated. KOs were considered to be present in a genome if this annotation procedure resulted in a copy number ≥0.1. For species with more than one sequenced strain, the average annotation across strains was used. KOs present in at least 75% of HMP reference organisms were considered core KOs and were removed from the analysis. Similarly, KOs present in fewer than 1% of HMP reference genomes were assumed to be spurious annotations and were excluded.
One of the components required to deconvolve metagenomic samples is a constant genomic element or gene that can be used as a normalization coefficient for inferring the length (or copy number) of all other genomic elements. Ideally, genes used for normalization should be present in all the species in the community, have the same copy number in each genome, and have a consistent length across all species. The 16S rRNA gene is a natural candidate, but other gene orthology groups can be used as well. Specifically, in the main text, we deconvolved tongue dorsum samples from the Human Microbiome Project using a combination of ribosomal protein-coding genes. Ribosomal genes are generally good candidates for normalization since the ribosome is a highly-conserved construct. Using the combined abundances of multiple genes can reduce the potentially deleterious effect of read annotation errors in any one gene. Starting with 31 ribosomal protein-coding KOs present in both bacteria and archaea, we first considered those that were present in at least 1445 (98%) of the 1475 bacteria and archaea in KEGG v60 [19]. Of these KOs, we selected a subset of 15 KOs that had a lower variation in length across all genomes than the 16S gene (Table S1). These 15 KOs were used jointly as our constant genomic element for normalization, using the sum of the abundances as the constant genomic element abundance and sum of the lengths as the constant genomic element length in Eq. 5.
|
10.1371/journal.pcbi.1005195 | PoCos: Population Covering Locus Sets for Risk Assessment in Complex Diseases | Susceptibility loci identified by GWAS generally account for a limited fraction of heritability. Predictive models based on identified loci also have modest success in risk assessment and therefore are of limited practical use. Many methods have been developed to overcome these limitations by incorporating prior biological knowledge. However, most of the information utilized by these methods is at the level of genes, limiting analyses to variants that are in or proximate to coding regions. We propose a new method that integrates protein protein interaction (PPI) as well as expression quantitative trait loci (eQTL) data to identify sets of functionally related loci that are collectively associated with a trait of interest. We call such sets of loci “population covering locus sets” (PoCos). The contributions of the proposed approach are three-fold: 1) We consider all possible genotype models for each locus, thereby enabling identification of combinatorial relationships between multiple loci. 2) We develop a framework for the integration of PPI and eQTL into a heterogenous network model, enabling efficient identification of functionally related variants that are associated with the disease. 3) We develop a novel method to integrate the genotypes of multiple loci in a PoCo into a representative genotype to be used in risk assessment. We test the proposed framework in the context of risk assessment for seven complex diseases, type 1 diabetes (T1D), type 2 diabetes (T2D), psoriasis (PS), bipolar disorder (BD), coronary artery disease (CAD), hypertension (HT), and multiple sclerosis (MS). Our results show that the proposed method significantly outperforms individual variant based risk assessment models as well as the state-of-the-art polygenic score. We also show that incorporation of eQTL data improves the performance of identified POCOs in risk assessment. We also assess the biological relevance of PoCos for three diseases that have similar biological mechanisms and identify novel candidate genes. The resulting software is publicly available at http://compbio.case.edu/pocos/.
| Several studies try to predict the individual disease risk using genetic data obtained from genome wide association studies (GWAS). Earlier studies only focus on individual genetic variants. However, studies on disease mechanisms suggest the aggregation of genomic variants may contribute to diseases. For this reason, researchers commonly use prior biological knowledge to identify genetic variants that are functionally related. However, these approaches are often limited to variants that are in the coding regions of genes. However, several risk variants are in the regulatory region. Here, we incorporate known regulatory and functional interactions to find sets of genetic variants which are informative features for risk assessment. Our result on seven complex diseases show that our method outperforms individual variant based risk assessment models, as well as other methods that integrate multiple genetic variants.
| Genome-wide association studies (GWAS) have a transformative effect on the search for genetic variants that are associated with complex traits, since they enable screening of hundreds of thousands of genomic variants for their association with traits of interest [1]. Recently published GWAS lead to the discovery of susceptibility loci for many complex diseases, including type 2 diabetes [2], psoriasis [3], multiple sclerosis [4], and prostate cancer [5]. For improved identification of risk variants, researchers draw information from clinical, microarray, copy number, and single nucleotide polymorphism (SNP) data to build disease risk models, which are then used to predict an individual’s susceptibility to the disease of interest [6, 7]. Several companies, such as deCODE genetics (http://www.decode.com) and 23andme (https://www.23andme.com) have started using SNPs identified by GWAS, to provide personal genomic test services in the United States and health related genomic test services in Canada and the United Kingdom.
An important problem with GWAS is that the identified variants account for little heritability [8, 9]. However, empirical evidence from model organisms [10] and human studies [11] suggests that the interplay among multiple genetic variants contribute to complex traits. Epistasis among pairs of loci, i.e., significantly improved association with the phenotype when two loci are considered together, is also shown to provide provide further insights into disease mechanisms [12–14]. Therefore, recent studies focus on identifying the interactions among pairs of genomic loci, as well as among multiple genomic loci [15–17]. These studies suggest that consideration of more than one locus together can better capture the relationship between genotype and phenotype. For this reason, genetic markers that aggregate multiple genomic loci can be used to design effective strategies for risk assessment and guide treatment decisions [18].
The Polygenic score is a commonly used method to identify the joint association of a large mass of the loci to predict disease risk [19]. The first application of polygenic score on GWAS data shows that the genetic risk for schizophrenia is a predictor of bipolar disorder [20]. There are also several studies demonstrating that polygenic risk score is a powerful tool in risk prediction [20–22]. However, polygenic score does not make use of prior biological knowledge, which may be useful in generating more robust features by incorporating the functional relationships among individual variants. Furthermore, according to a recent comparative assessment of various classification algorithms, there are no statistically significant differences between state-of-the-art classification algorithms in terms of performance in risk assessment [23]. This observation suggests that research on construction of features for risk assessment can be useful in improving the classification performance of these algorithms.
Since detection of epistasis and higher order interactions is computationally expensive, many methods first assess the disease association of individual loci and then use functional knowledge to integrate these associations [24–26]. The key idea behind these methods is that functionally related variants, e.g., those that induce dense subnetworks in protein-protein interaction (PPI) networks, can provide stronger statistical signals when they are considered together [27]. Based on similar insights, some researchers integrate GWAS with pathway information to identify statistically significant pathways that are associated with the disease [28, 29]. Recently, Azencott et al. propose a method to discover sets of genomic loci that are associated with a phenotype while being connected in an underlying biological network [30]. They use an additive model to integrate the genotypes of loci and use connectivity patterns in the network to select a functionally coherent set of disease associated SNPs. While this method works on a network of genomic loci, the network is constructed based on the interactions among genes and mapping of loci to genes. For this reason, the application of these methods is limited to the variants in coding regions or in regions that are in close proximity to genes. However, 88 percent of genotyped variants in GWAS fall outside of coding regions [31]. Several risk variants are found in non-coding regions of the genome and it is shown that the functional effects of these variants are regulatory (e.g., mRNA expression, microRNA expression) as opposed to directly influencing protein structure or function [32].
In this paper, we propose a new algorithm for the identification of multiple functionally related genomic variants that are collectively associated with a phenotype. The proposed method builds on the concept of “Population Covering Locus Sets” (PoCos) [33, 34]. A PoCo is a set of loci that harbor at least one susceptibility allele in samples with the phenotype of interest. Here, we extend the notion of PoCos to enable adaptive identification of “susceptibility genotype” (as opposed to susceptibility allele) for each locus. We also develop a method for aggregating the genotypes of multiple loci in a PoCo to compute representative genotypes for use in risk assessment. Finally, in order to capture the functional relationship between genomic loci, we integrate GWAS data with human protein-protein interaction (PPI) network and regulatory interactions identified via expression quantitative trait loci (eQTL).
We use the PoCos identified by the proposed framework to construct features that can be used in risk assessment. We evaluate the performance of PoCos in risk assessment via cross-validation on seven GWAS case-control data sets obtained from the Wellcome Trust Case-Control Consortium (WTCCC). We compare the risk assessment performance of models built using PoCos to that of models built using individual loci and polygenic score. Our experimental results show that PoCos significantly outperform individual loci and polygenic score in risk assessment. Furthermore, we assess the information added by the incorporation of PPI and eQTL and observe that inclusion of these data leads to more parsimonious models for risk assessment.
In the next section, we describe the proposed procedure for modeling the genotypes and identifying PoCos. Then we describe how we use PoCos to develop a model for risk assessment. Subsequently, we present comprehensive experimental results on GWAS data sets for Type 2 Diabetes (T2D), Psoriasis (PS), Type 1 Diabetes (T1D), Hypertension (HT), Bipolar Disorder (BD), Multiple Sclerosis (MS) and Coronary Artery Disease (CAD). Our results show that the proposed method significantly outperforms individual variant based risk assessment model as well as the state-of-the-art polygenic score. We also observe that integrating prior biological information leads to more parsimonious models for risk assessment.
In this section, we first present the set-up for genome-wide association studies. We then define “Population Covering Locus Sets” (PoCos) and describe the algorithm we use to identify PoCos. Finally, we describe our feature selection framework for the selection of PoCos to be used for risk assessment. The workflow of the proposed method is presented in Fig 1.
The input to the problem is a genome-wide association (GWA) dataset D = (C, S, g, f), where C denotes the set of genomic loci that harbor the genetic variants (e.g., single nucleotide polymorphisms or copy number variants) that are assayed, S denotes the set of samples, g(c, s) denotes the genotype of locus c ∈ C in sample s ∈ S, and f(s) denotes the phenotype of sample s ∈ S. Here, we assume that the phenotype variable is dichotomous, i.e., f(s) can take only two values: if sample s is associated with the phenotype of interest (e.g. diagnosed with the disease, responds to a certain drug etc.), s is called a “case” sample (f(s) = 1), otherwise (e.g., was not diagnosed with the disease, does not respond to a certain drug etc.), s is called a “control” sample (f(s) = 0). We denote the set of case samples with S1 and the set of control samples with S0, where S1 ∪ S0 = S. While we focus on qualitative traits here for brevity, the proposed methodology can also be extended to quantitative traits (i.e., when f(s) is a continuous phenotype variable).
The minor allele for a locus is usually defined as the allele that is less frequent in the population. While it is common to focus on the minor allele as the risk allele, specific genotypes can also be associated with a phenotype [35–37]. Different types of encoding may represent different biological assumptions. In an additive model, each genotype is encoded as a single numeric feature that reflects the number of minor alleles (homozygous major, heterozygous, and homozygous minor are respectively encoded as 0, 1 and 2). This model does not capture combinatorial relationships between locus genotypes and phenotype, since the assumption is that one of the alleles quantitatively contributes to risk. In the recessive/dominant model, each genotype is encoded as two binary features (presence of minor allele and presence of major allele). This model does not capture the difference between homozygous and heterozygous genotypes, since it only accounts for the presence of an allele. Here, we argue that considering the effect of all possible genotype combinations can provide more information in distinguishing case samples from control samples. The five models proposed here capture all potential relationships, in that differences in heterozygosity vs. homozygosity, presence vs. absence of a specific risk allele are represented by different genotype models. This notion is particularly useful when the genotypes of multiple loci are being integrated. For example, heterozygosity on one locus can be associated with increased susceptibility to a disease, while homozygous minor allele on another locus may be protective at the presence of heterozygosity in the former locus [38]. In this case, the interaction between the two loci can be detected by considering the association of all possible genotype combinations with the phenotype.
We adaptively binarize the genotypes of each locus by considering all possible allele combinations. Given the genotype of a locus, we consider five different binary genotype models m(i), i ∈ {1, … 5}. Based on each model, we generate a binary genotype profile for each locus. Namely, we consider the following genotype models:
1. Homozygous Minor Allele: This corresponds to the case when the possible effect of the minor allele is “recessive”, i.e., the locus is considered to harbor a genotype of interest if both copies contain the minor allele.
m ( 1 ) ( c , s ) = 1 2 if g ( c , s ) ∈ { a a } 0 otherwise (1)
2. Heterozygous Genotype: The locus is considered to harbor a genotype of interest if the two copies contain different alleles.
m ( 2 ) ( c , s ) = 1 2 if g ( c , s ) ∈ { A a } 0 otherwise (2)
3. Homozygous Major Allele: The locus is considered to harbor a genotype of interest if both copies contain the major allele.
m ( 3 ) ( c , s ) = 1 2 if g ( c , s ) ∈ { A A } 0 otherwise (3)
4. Presence of Minor Allele: This corresponds to the case when the possible effect of the minor allele is “dominant”, i.e., the locus is considered to harbor a genotype of interest if at least one copy contains the minor allele. This is the complement of m(3).
m ( 4 ) ( c , s ) = 1 2 if g ( c , s ) ∈ { A a , a a } 0 otherwise (4)
5. Presence of Major Allele: The locus is considered to harbor a genotype of interest if at least one copy contains the major allele. This is the complement of m(1).
m ( 5 ) ( c , s ) = 1 2 if g ( c , s ) ∈ { A a , A A } 0 otherwise (5)
Note that, although models m4 and m5 are complements of other models, we consider them separately. This is because, as we discuss in the next section, the 1s and 0s in the binary genotype profiles are considered asymmetrically while integrating the genotypes of multiple loci. Also note that “homozygous minor allele or homozygous major allele” is not considered since it is not associated with a specific risk allele.
To select a genotype model for each locus, we separately assess the association of the resulting five genotype profiles with the phenotype of interest. Subsequently, we choose the model that leads to greatest discrimination between cases and controls, and use the respective binary genotype profile as the representative genotype of that locus. This process is illustrated in Fig 2.
For each locus c, binarization according to the five different genotype models produces five |S|-dimensional binary genotype profiles m(i)(c), i ∈ {1, … 5}. For each binary genotype profile m(i)(c), we compute the difference in the fraction of case and control samples that harbor the genotype of interest as follows:
D ( i ) ( c ) = 〈f, m ( i ) ( c ) 〉 | S 1 | - 〈 1 - f , m ( i ) ( c ) 〉 | S 0 | . (6)
where 1 denotes a vector of all 1’s and <.> denotes the inner product of two vectors. We then determine the binary genotype model for each locus as the model that maximizes the difference of relative coverage between case samples and control samples, i.e.:
k ( c ) = argmax i ∈ { 1 ⋯ 5 } { | D ( i ) ( c ) | } . (7)
Based on the selected model for each locus, we compute the binary genotype profile accordingly:
M ( c , s ) = m ( k ( c ) ) ( c , s ) . (8)
Once we compute the binary genotype profiles for all loci, we identify Population Covering Locus Sets (PoCos). In previous work, we define and use PoCos in the context of prioritizing locus pairs for testing epistasis [33]. In this earlier definition, the genotypes of interest are limited to the presence of the minor or major allele; i.e., only the last two models described in the previous section are used to determine the binary genotype profile of each locus. Here, we generalize the concept of PoCo to utilize five different models for determining the genotypes of interest, as described in the previous subsection.
A PoCo is a set of genomic loci that collectively “cover” a larger fraction of case samples while minimally covering control samples. Namely for a given set P ⊆ C of loci, we define the set of case and control samples covered by P respectively as
E ( P ) = ∪ c ∈ P { s ∈ S 1 : M ( c , s ) = 1 } (9)
and
T ( P ) = ∪ c ∈ P { s ∈ S 0 : M ( c , s ) = 1 } . (10)
We define a PoCo as a set P of loci that satisfies |E(P)| = |S1| while minimizing |T(P)|. Note that, since we are interested in finding all sets of loci with potential relationship in their association with phenotype, we do not define an optimization problem that aims to find a single PoCo with minimum |T(P)|. We rather develop an algorithm to heuristically identify all non-overlapping PoCos with minimal |T(P)|.
To identify all non-overlapping PoCos, we use a greedy algorithm that progressively grows a set of loci to maximize the difference of the fraction of case and control samples covered by the loci that are recruited in a PoCo. In another words, we initialize P to ∅ and at each step, add to P the locus that maximizes
δ ( c ) = E ( { c } ) ∩ S ′ | | S 1 | - | T ( { c } ) ∩ S ′ | | S 0 | (11)
where S′ = S\(E(P) ∪ T(P)). The algorithm stops when all case samples are covered. We then record P, remove the loci in P from the dataset, and identify another PoCo. This process continues until it is not possible to find a set of loci that covers all case samples.
We develop two methods to identify two different types of PoCos. The first type of PoCos (named “network-free PoCos”) are identifed using the greedy algorithm described above, without the use of any prior biological information. The second type of PoCos are NetPocos, which are identified by restricting the search space to connected subgraphs of a network of potential functional relationships among genomic loci. As we describe below, this network is constructed by integrating established locus-gene associations from eQTL studies and protein-protein interaction (PPI) data that contains functional relationships among genes.
One potential utility of the PoCos is risk assessment. By construction, PoCos (NetPocos) contain (functionally associated) loci that exhibit improved power in distinguishing cases from control. Consequently, as compared to individual variants, they may provide more robust and reproducible features to be used in predictive models. To investigate the utility of these multi-locus features in risk assessment, we use PoCos to build a model for risk assessment using L1 regularized logistic regression classifier.
To assess the ability of PoCos in producing informative multi-locus features, we evaluate their utility in the context of risk assessment. For this purpose, we use GWAS data from the Welcome Trust Case-Control Consortium (WTCCC), which includes data from studies for seven complex diseases, namely type 1 diabetes (T1D), type 2 diabetes (T2D), psoriasis (PS), bipolar disorder (BD), coronary artery disease (CAD), hypertension (HT), and multiple sclerosis (MS). On each dataset, we first identify PoCos, select features to build a model for risk assessment, and then evaluate the performance of the resulting model. To control for overfitting and to ensure that the performance figures are not biased, we use cross validation.
We first compare the risk assessment performance of the multi-locus features against the standard approach of using individual significant loci. To facilitate fair comparisons, we use the classification and feature selection methods described in the “performance evaluation for risk assessment” section identically for all types of multi-locus and individual-locus based features. We also compare the performance of NetPocos against Polygenic Score, which is a commonly used method for risk assessment. Subsequently, to gain insights into the information provided by network data and specifically eQTL-based regulatory interactions, we also compare the performance of NetPocos, network-free PoCos, and eQTL-free PoCos. Moreover, we investigate the effect of λ in the L1 regularized logistic regression classifier, i.e. the parameter that controls the parsimony of the model. We also assess the biological relevance of some of the selected PoCos using enrichment analysis and a literature-driven list of genes and processes that have been reported to be associated with diseases. Finally, we compare the most frequently recruited genes in PoCos in different diseases to gain insights into shared genetic bases of different diseases. This analysis also suggests novel potential susceptibility genes for these diseases.
For each dataset, we divide the population into 5 groups while preserving the proportion of case and control samples in each group. We reserve one group for testing and we identify NetPocos on the remaining four groups. Then, we use these four groups for feature selection and model building. Finally, we test the performance on the group reserved for testing. All of the reported performance figures are averages across five different cross-validation runs. The number of PoCos identified on each dataset and the size of these PoCos are presented in Table 2. Please note that the variance in number of PoCos does not have a significant effect on the performance (S1 Fig).
In this paper, we propose a novel criterion to assess the collective disease-association of multiple genomic loci (PoCos) and investigate the utility of these multiple-loci features in risk assessment. We also perform extensive experiments to evaluate the effect of using network information to drive the search for multi-locus features on risk assessment. We also investigate the effect of the variants that have regulatory effects (i.e. eQTL data) on performance for risk assessment. Moreover, we compare the proposed method with the polygenic score which has been shown to be successful in different studies. Our result show that our method is significantly more powerful in risk assessment.
Our results show that multi-locus features improve prediction performance as compared to individual locus based features. We also observe that integrating functional information provided by protein-protein interaction data and expression quantitative trait loci (i.e. eQTL) data leads to more parsimonious models for risk assessment. However, inclusion of functional data does not yield significant improvement in prediction performance. This may be indicative of the limitations of genomic data in risk assessment. Furthermore, since PoCos contain loci that are related to each other in the context of a phenotype, PoCos that are discovered without the inclusion of functional information also likely contain functionally related loci. However, utilization of functional information reduces the search space to render the problem computationally feasible, and brings forward PoCos that are more functionally relevant and robust, thereby leading to more parsimonious models.
Based on the success of multi-locus genomic features in risk assessment, we conclude that combining these features with non-genetic risk factors and other biological data may lead to further improvements in risk assessment.
The proposed method is implemented in MATLAB and provided in the public domain (http://compbio.case.edu/pocos/) as open source software.
|
10.1371/journal.pmed.1002225 | Performance and Cost-Effectiveness of Computed Tomography Lung Cancer Screening Scenarios in a Population-Based Setting: A Microsimulation Modeling Analysis in Ontario, Canada | The National Lung Screening Trial (NLST) results indicate that computed tomography (CT) lung cancer screening for current and former smokers with three annual screens can be cost-effective in a trial setting. However, the cost-effectiveness in a population-based setting with >3 screening rounds is uncertain. Therefore, the objective of this study was to estimate the cost-effectiveness of lung cancer screening in a population-based setting in Ontario, Canada, and evaluate the effects of screening eligibility criteria.
This study used microsimulation modeling informed by various data sources, including the Ontario Health Insurance Plan (OHIP), Ontario Cancer Registry, smoking behavior surveys, and the NLST. Persons, born between 1940 and 1969, were examined from a third-party health care payer perspective across a lifetime horizon. Starting in 2015, 576 CT screening scenarios were examined, varying by age to start and end screening, smoking eligibility criteria, and screening interval. Among the examined outcome measures were lung cancer deaths averted, life-years gained, percentage ever screened, costs (in 2015 Canadian dollars), and overdiagnosis. The results of the base-case analysis indicated that annual screening was more cost-effective than biennial screening. Scenarios with eligibility criteria that required as few as 20 pack-years were dominated by scenarios that required higher numbers of accumulated pack-years. In general, scenarios that applied stringent smoking eligibility criteria (i.e., requiring higher levels of accumulated smoking exposure) were more cost-effective than scenarios with less stringent smoking eligibility criteria, with modest differences in life-years gained. Annual screening between ages 55–75 for persons who smoked ≥40 pack-years and who currently smoke or quit ≤10 y ago yielded an incremental cost-effectiveness ratio of $41,136 Canadian dollars ($33,825 in May 1, 2015, United States dollars) per life-year gained (compared to annual screening between ages 60–75 for persons who smoked ≥40 pack-years and who currently smoke or quit ≤10 y ago), which was considered optimal at a cost-effectiveness threshold of $50,000 Canadian dollars ($41,114 May 1, 2015, US dollars). If 50% lower or higher attributable costs were assumed, the incremental cost-effectiveness ratio of this scenario was estimated to be $38,240 ($31,444 May 1, 2015, US dollars) or $48,525 ($39,901 May 1, 2015, US dollars), respectively. If 50% lower or higher costs for CT examinations were assumed, the incremental cost-effectiveness ratio of this scenario was estimated to be $28,630 ($23,542 May 1, 2015, US dollars) or $73,507 ($60,443 May 1, 2015, US dollars), respectively.
This scenario would screen 9.56% (499,261 individuals) of the total population (ever- and never-smokers) at least once, which would require 4,788,523 CT examinations, and reduce lung cancer mortality in the total population by 9.05% (preventing 13,108 lung cancer deaths), while 12.53% of screen-detected cancers would be overdiagnosed (4,282 overdiagnosed cases). Sensitivity analyses indicated that the overall results were most sensitive to variations in CT examination costs. Quality of life was not incorporated in the analyses, and assumptions for follow-up procedures were based on data from the NLST, which may not be generalizable to a population-based setting.
Lung cancer screening with stringent smoking eligibility criteria can be cost-effective in a population-based setting.
| In the US, lung cancer screening is recommended for current and former smokers who have quit within the past 15 y, aged 55 through 80 who smoked at least 30 pack-years; other countries are investigating the feasibility of implementing lung cancer screening policies.
Despite lung cancer screening being recommended by a number of organizations, the cost-effectiveness of lung cancer screening is uncertain; concerns have been raised on the potential costs of implementing lung cancer screening.
Past studies that evaluated the cost-effectiveness of lung cancer screening yielded inconclusive results. However, these studies considered limited numbers of screening policies, providing limited information on how different screening policy characteristics affect the cost-effectiveness of lung cancer screening.
This study investigated how different screening policy characteristics, such as screening starting and stopping ages, screening interval, and different smoking history eligibility criteria, influence the performance and cost-effectiveness of lung cancer screening.
A microsimulation model was used to analyze 576 different lung cancer screening policies for persons born between 1940 and 1969 in Ontario, Canada.
The study found that requiring stringent smoking history eligibility criteria (i.e., requiring higher levels of accumulated smoking exposure) was more cost-effective than less stringent smoking history eligibility criteria.
Limiting screening to individuals with substantial (past) smoking histories may allow lung cancer screening to be implemented in a cost-effective manner.
In contrast to initial expectations, annual screening is suggested to be more cost-effective than biennial screening.
| The National Lung Screening Trial (NLST) showed that screening with low-dose computed tomography (CT) can reduce lung cancer mortality [1]. Although the sensitivity of CT screening in the NLST was reported to be over 90% across the three screening rounds, the reported specificity ranged from 73.4% in the first round to 83.9% in the third round [1]. Overall, 23.3% of the CT screens in the NLST were false positive, which often required additional follow-up CT examinations and, infrequently, invasive procedures (such as a biopsy, bronchoscopy, or thoracotomy) to determine the malignancy of one or more suspicious pulmonary nodules detected by CT screening [1]. Lung cancer screening with three annual screens, as performed in the NLST, was reported to be cost-effective by US standards, yielding estimated cost-effectiveness ratios of US$52,000 per life-year gained and US$81,000 per quality-adjusted life-year gained [2,3]. However, although the cost-effectiveness of lung cancer screening in a population-based setting has been examined previously, it has not been examined extensively [4–9].
To determine the cost-effectiveness of implementing cancer screening programs, microsimulation modeling is invaluable [10,11]. The United States Preventive Services Task Force (USPSTF) recommended lung cancer screening for current and former smokers who have quit within the past 15 y, aged 55 through 80 who smoked at least 30 pack-years [12]. This recommendation was in part based on a comparative modeling study using microsimulation models, as modeling allows one to extrapolate the results of randomized clinical trials and provide information on the long-term benefits and harms for screening programs with different designs and populations than those considered in clinical trials [13]. However, although the modeling study that informed the USPSTF provides an understanding of the trade-offs between the benefits and harms of different screening scenarios, it did not formally consider their cost-effectiveness [13].
In Ontario, Canada, lung cancer is responsible for the largest proportion of cancer deaths (49.9 per 100,000 individuals) in the population of 13.8 million individuals, despite falling smoking rates [15,28,55,56]. The implementation of a lung cancer screening program, in addition to continued efforts in primary prevention of smoking, could reduce lung cancer mortality. However, concerns have been raised about whether and how such a program can be implemented in a cost-effective manner [14,16]. Previous studies on the cost-effectiveness of population-based lung cancer screening have yielded inconclusive results, ranging from US$18,452–US$66,480 per life-year gained and US$27,756–US$243,077 per quality-adjusted life-year gained [4–9]. However, many of these studies reported the average cost-effectiveness ratios (ACER, the ratio of differences in costs to differences in health effects compared to no screening) of the investigated screening scenarios as the incremental cost-effectiveness ratios (ICER, the ratio of incremental costs to incremental health effects of a screening policy relative to its next best alternative), which can give misleading cost-effectiveness estimates [17].
Furthermore, these studies considered limited numbers of screening scenarios, providing little information on the effects of screening eligibility criteria, and may have had insufficient numbers of comparator scenarios to yield correct ICERs [18]. The aim of this study was to investigate the benefits (such as lung cancer mortality reduction and the number of life-years gained), harms (such as the number of false-positive results and occurrence of overdiagnosis), and cost-effectiveness of many different lung cancer screening scenarios for the population of Ontario, overcoming some of the limitations of previous studies.
This study was approved by the Research Ethics Board of Sunnybrook Health Sciences Centre on behalf of the Institute for Clinical Evaluative Sciences (ICES). Individual consent for access to de-identified data was not required.
The MIcrosimulation SCreening ANalysis (MISCAN) Lung model was used for this analysis. Other versions of the MISCAN model have been used to investigate the cost-effectiveness of screening programs for breast, colorectal, cervical, and prostate cancers [19–22]. The MISCAN-Lung model used in these analyses was previously calibrated to individual-level data from the NLST and the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial, from which information on the preclinical duration of lung cancer and the effectiveness of CT screening were derived [23,24]. MISCAN-Lung was one of the models used to inform the USPSTF on their recommendations for lung cancer screening [13]. The structure of the model and its underlying assumptions have been described previously and are detailed in S1 Text; the characteristics of the investigated population are described in the following section and in S2 and S3 Texts [23,24].
In brief, MISCAN-Lung simulates life histories for each individual in the considered population from birth until death in the presence and absence of screening. For each individual, a smoking history is generated based on data on the investigated population. A person’s smoking history influences the probability of developing preclinical lung cancer as well as the probability of dying from other causes. The model considers four histological types of lung cancer: adenocarcinoma/large cell carcinoma/bronchioloalveolar carcinoma, squamous cell carcinoma, other non-small cell carcinoma, and small cell carcinoma. Once preclinical lung cancer has developed, it is assumed to progress through stages IA to IV. During each stage, the cancer may be detected due to symptoms, after which the person is assumed to undergo treatment with associated treatment costs. Lung cancer survival after clinical diagnosis is dependent on the histology and stage of the cancer and the person’s gender. If the screening component is activated, preclinical lung cancers may be detected by screening (at the expense of screening-related costs), which may alter a person’s life history: detection by screening allows treatment at an earlier stage, which may cure the individual, allowing him or her to resume a normal (lung cancer-free) life history. The probability that an individual is cured due to early detection differs by the stage at detection. Screening may also result in serious harms, such as overdiagnosis (the detection of a disease that would never have been detected if screening had not occurred), which may lead to unnecessary (invasive) follow-up procedures, treatments, and anxiety. The effects of screening are derived through utilizing information on the preclinical duration of lung cancer, the screen-detectability of lung cancer, and relevant information on the examined population (such as smoking behavior and other-cause mortality corrected for smoking history) to model the life histories of individuals in the presence and absence of screening [25]. Through comparing the life histories in the presence of screening with the corresponding life histories in the absence of screening, MISCAN-Lung can estimate the effectiveness and costs of screening scenarios.
Three different birth cohorts were investigated: 1940–1949 (ages 66–75 in 2015), 1950–1959 (ages 56–65 in 2015), and 1960–1969 (ages 46–55 in 2015). These cohorts represent approximately 5.2 million individuals in 2016 in the age range for which the USPSTF currently recommends lung cancer screening [12,26]. Birth tables for each cohort were derived from information from Statistics Canada and the Ontario Ministry of Finance [26,27]. Ontario-specific data on smoking behavior were used to model smoking initiation and cessation probabilities and the average number of cigarettes smoked per day (divided into five categories) by age and gender for each cohort [28–32]. Life tables by birth year and gender were extracted from the Canadian Human Mortality Database and adjusted for smoking behavior and lung cancer mortality, as shown graphically for persons born in 1955 in S2 Text Figs C–F [32]. Further information on the methods used to model smoking behavior and adjustment of the life tables for smoking behavior is provided in S2 Text. The age- and gender-specific lung cancer incidence, histology proportions, and stage proportions estimated by MISCAN-Lung were compared to observed data from the Ontario Cancer Registry from 2007–2009, in which screening did not occur, and are detailed in S3 Text [33].
In total, 576 potential screening scenarios were evaluated. The evaluated scenarios considered different combinations of the following characteristics: age to start screening; age to stop screening; screening interval; and screening eligibility regarding cumulative smoking exposure, years since smoking cessation (for former smokers, defined as individuals who have quit smoking permanently), and whether or not former smokers were excluded from further screening after they reach a maximum number of years since cessation (Table 1). Two types of cumulative smoking criteria, of which one is used at a time in an evaluated scenario, were distinguished: the first type was based on the cumulative number of pack-years, used in the NLST (“NLST-like”) [1]; the second type of cumulative smoking criteria was based on the criteria used in the Dutch-Belgian lung cancer screening trial (NELSON), which evaluated a person’s smoking duration and average number of cigarettes per day separately (“NELSON-like”) [34]. Of the 576 screening scenarios, 216 screening scenarios considered “NLST-like” criteria (including the criteria currently recommended by the USPSTF), whereas 360 screening scenarios considered “NELSON-like” criteria. Perfect attendance to screening was assumed for the base-case investigation. Estimations regarding screen-related procedures such as false-positive results (defined as receiving a positive screening test result when lung cancer is not found after diagnostic work-up), follow-up CT examinations, screen-related biopsies/bronchoscopies, and non-lung cancer-related surgeries were derived from individual-level data from the CT-arm of the NLST and are described in S4 Text. Limited information on morbidity and mortality has been reported from randomized controlled trials. In a subgroup analysis of NELSON participants, 1% of participants were found to have incidental findings that required additional work-up procedures or treatment [35]. However, no information on morbidity or mortality was reported. In the NLST, 0.06% of the positive screening tests in the low-dose CT group that did not result in a diagnosis of lung cancer were associated with a major complication after an invasive procedure [1]. Overall, six individuals with a positive screening test in the low-dose CT group that did not result in a diagnosis of lung cancer died within 60 d after an invasive diagnostic procedure (0.04%), but it was unknown whether these deaths were caused by complications of the diagnostic procedures [1]. Thus, given the low occurrence of invasive procedures along with a low frequency of major complications, the occurrence of morbidity or death related to screen-related follow-up procedures is expected to be minor.
The analyses were conducted from a third-party health care payer perspective. Fully allocated costs for lung cancer treatment were estimated by stage, age, and gender from the date of diagnosis until the date of death or last known date of follow-up (by person-month) through data from the Ontario Health Insurance Plan (OHIP), the Canadian Institute for Health Information (CIHI), the Ontario Drug Benefit Plan database, the Ontario Chronic Care database, the Ontario Home Care database, and the Ontario New Drug Funding Program. These datasets were linked using unique encoded identifiers and were analyzed at the ICES. Controls without a lung cancer diagnosis from the Registered Persons Database (a roster of all OHIP beneficiaries) were matched to 12,713 staged cases of lung cancer from the Ontario Cancer Registry (10 controls matched per case), based on age, sex, median household income, and census tract on the date of diagnosis of the case. Fully allocated costs were estimated similarly for controls. The fully allocated costs of controls were subtracted from the fully allocated costs for cases in order to obtain the attributable costs of lung cancer care by phase of care (initial, continuing, and terminal care) [36]. By incorporating the fully allocated costs of lung cancer care by phase of care, it is taken into account that individuals whose lung cancer death is averted will on average incur higher costs over their remaining lifetime. In MISCAN-Lung, the attributable costs for stage I were assumed for the modeled stages IA and IB, whereas the attributable costs of stage III were assumed for the modeled stages IIIA and IIIB.
Each person’s eligibility for lung cancer screening was assumed to be free of misclassification error. Therefore, upon entering the eligible age range for the considered screening scenarios, each ever-smoking individual was assumed to receive an invitation for a lung cancer risk assessment. It was assumed that half of all ever-smokers would accept this invitation; half of the individuals who participated in the risk assessment were assumed to request a consultation with a primary care physician about their risk. Costs for screening-related events were determined using 2013 data from OHIP and CIHI. The costs for screening invitations and fixed costs related to the screening program, such as costs for the screening registry, program infrastructure, communications, and advertising, were derived from those incurred in the recent establishment of the colorectal screening program administered by Cancer Care Ontario. The costs for risk assessments were estimated assuming that screening program staff trained in health communication would administer the assessments. Fixed costs were counted up to the year in which the last individuals in the cohorts are eligible for screening (2045 for screening scenarios that end at age 75, 2050 for screening scenarios that end at age 80).
All costs were expressed in Canadian dollars (using May 2013 levels as a base) and were adjusted to reflect the May 2015 prices for health care services using the Ontario Consumer Price Index derived from Statistics Canada [37]. A lifetime time horizon for the costs and effects of screening was applied to each simulated person. Annual discount rates of 3% were applied to both costs and effects, using 2015 as the reference year [38]. The estimated attributable costs of lung cancer care by phase of care and the estimated costs related to the screening program are presented in Tables 2 and 3, respectively. To reflect the uncertainty in these cost estimates, sensitivity analyses were performed, which varied the costs by 50%, as described in a later section of this manuscript. Although cost-effectiveness thresholds have been proposed in the past, there is no official cost-effectiveness threshold employed in the Canadian health care system [37]. Therefore, a cost-effectiveness threshold of $50,000 Canadian dollars ($41,114 in May 1, 2015, US dollars) per life-year gained was chosen, similar to previous Canadian cost-effectiveness studies [39,40].
For each screening scenario, the number of lung cancer deaths prevented, life-years gained, proportion of individuals ever screened, number of CT examinations, screen-related biopsies/bronchoscopies, false-positive screens, non-lung cancer-related surgeries, overdiagnoses, and costs were compared with a situation in which screening does not occur, from 2015 onward. Screening scenarios that were more costly and less effective (i.e., fewer life-years gained) than other scenarios were ruled out as non-efficient by simple dominance. Scenarios that were more costly and less effective than a combination of other scenarios were also ruled out as non-efficient by extended dominance. The remaining screening scenarios constitute the frontier of efficient screening scenarios, i.e., the efficient frontier. For each efficient screening scenario, the ICER was determined, calculated as the incremental net costs per incremental life-year gained compared to the previous efficient screening scenario.
A number of sensitivity analyses were performed to investigate which groups of cost estimates and attendance assumptions have the greatest influence on the cost-effectiveness estimates by varying the costs for CT examinations by 50% compared with the base-case analyses, varying the attributable costs of lung cancer care by phase of care by 50% compared with the base-case analyses, and imperfect attendance rates for screening: low attendance (33% overall compliance rate), average attendance (55% overall compliance rate), and high attendance (64% overall compliance rate). Sensitivity analyses were performed for all 576 scenarios to investigate the effects of variations in assumptions on the composition of the efficient frontier.
The net discounted costs and life-years gained (from 2015 onwards) for each scenario were used to determine the screening scenarios on the efficient frontier, i.e., the scenarios that provide the highest number of life-years gained for their costs, in the base-case analysis, as shown in Figs 1 and 2. The scenarios that are on the efficient frontier are described in Table 4 and shown in Fig 2. A complete overview of the net discounted costs and life-years gained of all investigated screening scenarios is presented in S1 Appendix Figs A–H. All outcomes were reported per 100,000 individuals alive in 2015.
All scenarios on the efficient frontier consist of annual screening (Table 4), while biennial screening is dominated. Assuming a cost-effectiveness threshold of $50,000 Canadian dollars ($41,114 May 1, 2015, US dollars) per life-year gained as acceptable for the Canadian health care system, Scenario #2 was considered optimal: current and former smokers (who quit ≤ 10 y ago) who smoked ≥40 pack-years would be screened annually between ages 55–75, yielding an ICER of $41,136 Canadian dollars ($33,825 May 1, 2015, US dollars) per life-year gained. If 50% lower or higher attributable costs were assumed, the ICER of this scenario was estimated to be $38,240 ($31,444 May 1, 2015, US dollars) or $48,525 ($39,901 May 1, 2015, US dollars), respectively. If 50% lower or higher costs for CT examinations were assumed, the ICER of this scenario was estimated to be $28,630 ($23,542 May 1, 2015, US dollars) or $73,507 ($60,443 May 1, 2015, US dollars), respectively.
In addition, the benefits and harms of all scenarios on the efficient frontier were examined (Table 5). Scenario #2 would reduce lung cancer mortality in the overall population (which includes non-eligible individuals) by 9.05%, preventing 251 lung cancer deaths and gaining 2,531 life-years (undiscounted) over the lifetime of the program (i.e., on average, 10.08 life-years would be gained for each lung cancer death prevented). However, in Scenario #2, 9.56% of the overall population would receive at least one screen, requiring 91,692 CT screens and follow-up examinations. This scenario would lead to 14,729 false-positive screens and 163 surgeries for potentially benign disease (in persons in whom lung cancer is not detected) and 350 biopsies/bronchoscopies (in persons in whom lung cancer is not detected). Ultimately, 12.53% of all screen-detected cancers would be overdiagnosed, leading to 82 overdiagnosed cases. Based on the estimated number of individuals in the examined cohorts in 2016, Scenario #2 is estimated to screen 499,261 individuals at least once, require 4,788,523 CT examinations, and prevent 13,108 lung cancer deaths, while 4,282 cases of lung cancer would be overdiagnosed [26,27]. The average annual non-discounted costs compared to no screening would be approximately $1,400,000 Canadian dollars ($1,151,178 May 1, 2015, US dollars) over the considered time period; however, the annual costs are higher in the first years compared to later years, due to diminishing numbers of individuals meeting the eligibility criteria. For example, the average non-discounted costs compared to no screening are approximately $5,000,000 Canadian dollars ($4,111,350 May 1, 2015, US dollars) for 2015–2020 compared with approximately $1,600,000 Canadian dollars ($1,315,632 May 1, 2015, US dollars) in 2030–2035.
Scenarios with older starting ages have lower costs compared with scenarios that start at younger ages but also yield a smaller number of life-years gained (S1 Appendix Fig A). Raising the age to stop screening from 75 to 80 increases the costs and the number of life-years gained, but differences are modest (S1 Appendix Fig B). A comparison of scenarios by smoking eligibility criteria indicates that there is little difference between using NLST-like or NELSON-like smoking eligibility criteria (S1 Appendix Fig C).
S1 Appendix Figs D–G demonstrate the importance of cumulative smoking criteria on cost-effectiveness. Each increase in the cumulative smoking requirement for enrollment substantially decreases the costs while modestly decreasing the number of life-years gained in both NLST-like (S1 Appendix Fig D) and NELSON-like (S1 Appendix Fig E) screening scenarios. In other words, scenarios that apply stringent cumulative smoking eligibility criteria are closer to the efficient frontier than those that apply less restrictive cumulative smoking eligibility criteria. In general, scenarios that require only 20 pack-years are dominated by scenarios that apply more stringent (higher) pack-year criteria.
Increasing the maximum number of years since smoking cessation (S1 Appendix Fig F) and not excluding individuals from further screening once they reach the maximum number of years since cessation (S1 Appendix Fig G) both increase the costs and the numbers of life-years gained. However, the effects of these criteria are less pronounced than those related to cumulative smoking requirements.
S1 Appendix Fig H shows the effects of annual screening compared with biennial screening. Although biennial screening scenarios have substantially lower costs compared with annual screening scenarios, the life-years gained are also substantially lower. S1 Appendix Fig H demonstrates that annual screening scenarios dominate biennial screening scenarios.
Altering assumptions about attendance rates, CT examination costs, and attributable costs impacted the scenarios on the efficient frontier in the base-case analysis to varying degrees. Fig 3 provides an overview of the scenarios on the efficient frontier in the base-case analysis along with the discounted life-years gained and costs for these scenarios in the sensitivity analyses.
Altering assumptions also impacted the composition of the efficient frontier, as shown in S5 Text. When the attendance rates were varied, lower attendance rates were observed to shift scenarios with less restrictive criteria, especially with regards to smoking behavior, on the efficient frontier (S5 Text Tables A–C). This may be due to the fixed costs of the screening scenarios, which are independent of the number of screened individuals; at lower levels of participation these costs have a greater influence on the cost-effectiveness than the costs of CT examinations.
When the attributable costs were varied, it was observed that halving the attributable costs had little effect on the scenarios on the efficient frontier (S5 Text Table D). When the attributable costs were doubled, it was observed that scenarios with less restrictive criteria, especially with regards to smoking cessation, were shifted on the efficient frontier (S5 Text Table E).
When the costs of CT examinations were varied, it was observed that halving the costs of CT examinations also shifted scenarios with less restrictive criteria, in particular with regards to smoking cessation, on the efficient frontier (S5 Text Table F). Doubling the costs of CT examinations had the greatest effect of all sensitivity analyses; the scenarios with the most restrictive criteria with regards to age and smoking were shifted on the efficient frontier and the least costly scenarios on this efficient frontier favored biennial screening (S5 Text Table G).
Scenario #2 was on the efficient frontier across all sensitivity analyses, with the exception of assuming the lowest attendance rates (S5 Text Table H). In contrast, although Scenario #5 closely resembles the eligibility criteria that were used in the NLST, it was not on the efficient frontier in any of the sensitivity analyses.
This simulation study indicates that lung cancer screening can be cost-effective in a population-based setting when eligibility is restricted to high-risk groups. In contrast, utilizing loose eligibility criteria yields nonoptimal and potentially cost-ineffective scenarios, as the cost-effectiveness of lung cancer screening is highly dependent on scenario characteristics, primarily the smoking eligibility criteria. Scenarios that utilize stringent smoking eligibility criteria are more cost-effective than scenarios that utilize less restrictive smoking eligibility criteria due to a focus on individuals at higher risk of developing lung cancer. This greatly reduces the number of screening examinations while still screening those at highest risk. Thus, the level of lung cancer risk at which an individual is eligible for lung cancer screening should be considered before implementing lung cancer screening policies. Future research should investigate the cost-effectiveness of lung cancer screening selection based on accurate lung cancer risk prediction models using suitable risk thresholds [41–43].
The results of this study suggest that the greater reduction in lung cancer mortality and number of life-years gained by annual screening outweigh the costs of the additional number of CT examinations compared with biennial screening, which has previously been suggested to be equally or more cost-effective than annual screening [14,44]. However, previous studies indicated that lung cancer may be more difficult to detect in stage IA with biennial screening [24]. As survival in stage IA is considerably higher compared with other stages, the potential for mortality reduction and life-years gained is higher for annual screening compared to biennial screening [45]. This is supported by the modeling study that informed the USPSTF, which showed that annual screening provides substantial benefits over biennial screening at modest diminishing returns [13].
Previous studies that examined the cost-effectiveness of lung cancer screening only considered limited numbers of screening scenarios, which provided limited information on the effects of scenario characteristics [4–8]. The results of this study suggest that scenario characteristics, especially smoking eligibility criteria and screening interval, influence the cost-effectiveness of a scenario and suggest that a large variety of scenarios should be considered. In addition, considering a wide variety of screening scenarios provides sufficient comparator scenarios to yield appropriate ICERs [18]. Previous studies often reported the ACERs of the investigated screening scenarios as the ICERs, which can give misleading cost-effectiveness estimates [17]. This study provides both the ACERs and the ICERs of the scenarios on the efficient frontier, in contrast to previous studies that generally did not report an efficient frontier [4–8]. The robustness of the scenarios on the efficient frontier in this study is demonstrated by the sensitivity analyses of all 576 scenarios. This study incorporates both allocated costs for all screening-related procedures and attributable costs for various stages of lung cancer care, which were derived from government data in a province with universal health care, which allows for more comprehensive and accurate cost estimates compared with other studies. Furthermore, this study incorporates detailed information on smoking behavior and smoking-related mortality in contrast to previous studies. Finally, although the majority of previous studies only reported the number of life-years gained, this study reports a variety of benefits (such as lung cancer mortality reduction and the number of life-years gained) and harms (such as the number of false-positive results and the occurrence of overdiagnosis).
This study has some limitations; for example, quality of life was not incorporated in the analyses. There may be some differences in quality of life between annual and biennial screening, as more frequent screening will increase the impact of screening and follow-up–related effects on quality of life. However, results from the NELSON trial indicate that although CT lung cancer screening has a minor impact on quality of life in the short term, the long-term effects are negligible [46,47]. In addition, utility estimates for lung cancer care are highly variable [48].
Another limitation is that assumptions for follow-up procedures were based on data from the NLST, which may not be generalizable to a population-based setting, as screening algorithms with reduced false-positive rates are being investigated [1,49–51]. By reducing the false-positive rates, the number of unnecessary follow-up CTs and invasive diagnostic procedures may be reduced as well, further improving the cost-effectiveness of lung cancer screening.
Finally, although fully allocated costs for lung cancer care and observed costs for the administration of a cancer screening program were incorporated in the analyses, the government of Ontario only reimburses the physician costs of a CT examination. However, capital investments would be required to acquire the CT scanners necessary to implement a lung cancer screening program, which could influence the costs per CT examination. Conversely, the increased CT capacity could potentially lead to discounts on the costs per CT examination.
This study used a cost-effectiveness threshold of $50,000 Canadian dollars ($41,114 May 1, 2015, US dollars) per life-year gained, similar to previous Canadian cost-effectiveness studies [39]. However, the acceptable ratio between costs and effects differs between countries. For example, although a cost-effectiveness threshold of US$100,000 per quality-adjusted life-year has been proposed for the US, the United Kingdom’s National Institute for Health and Care Excellence uses a £20,000–£30,000 ($30,274–$45,411 May 1, 2015, US Dollars) threshold to determine cost-effectiveness [3,52,53]. Thus, the optimal screening scenario depends in part on the chosen cost-effectiveness threshold: if a cost-effectiveness threshold of $60,000 Canadian dollars ($49,336 May 1, 2015, US dollars) per life-year gained was chosen, Scenario #3 (annual screening for persons aged 55–75 who smoked at least 30 pack-years and currently smoke or quit smoking less than 10 y ago) would have been considered the optimal scenario. However, the ICER of Scenario #2 (annual screening for persons aged 55–75 who smoked at least 40 pack-years and currently smoke or quit smoking less than 10 y ago) remained below the proposed cost-effectiveness threshold of $50,000 Canadian dollars per life-year gained in 5 out of 7 sensitivity analyses (71.4%) with a range of $28,630–$73,507 Canadian dollars per life-year gained. This suggests that both the dominance and cost-effectiveness of Scenario #2 are robust across various sensitivity analyses.
Although our results suggest that a uniform biennial screening interval is dominated by a uniform annual screening interval, recent studies suggest it may be possible to identify individuals for whom biennial screening intervals could be recommended. NLST participants with a negative prevalence screen had a substantially lower risk of developing lung cancer compared to individuals with a positive prevalence screen [54]. Results from the NELSON trial suggest that the 2-y probability of developing lung cancer after a CT screen varied substantially by nodule size and volume doubling time [51]. Future research should evaluate whether the interval between screens can be varied based on previous screening results and what impact this has on cost-effectiveness. In addition, precision medicine could improve the treatment of selected individuals, and biomarkers might help to distinguish between indolent nodules and aggressive nodules requiring rapid diagnosis and treatment. The impacts of these future developments need to be assessed.
In conclusion, this study indicates that lung cancer screening can be cost-effective in a population-based setting if stringent smoking eligibility criteria are applied. Annual screening scenarios are more cost-effective than biennial screening scenarios.
|
10.1371/journal.pbio.1000543 | Phenotypic Consequences of Copy Number Variation: Insights from Smith-Magenis and Potocki-Lupski Syndrome Mouse Models | A large fraction of genome variation between individuals is comprised of submicroscopic copy number variation of genomic DNA segments. We assessed the relative contribution of structural changes and gene dosage alterations on phenotypic outcomes with mouse models of Smith-Magenis and Potocki-Lupski syndromes. We phenotyped mice with 1n (Deletion/+), 2n (+/+), 3n (Duplication/+), and balanced 2n compound heterozygous (Deletion/Duplication) copies of the same region. Parallel to the observations made in humans, such variation in gene copy number was sufficient to generate phenotypic consequences: in a number of cases diametrically opposing phenotypes were associated with gain versus loss of gene content. Surprisingly, some neurobehavioral traits were not rescued by restoration of the normal gene copy number. Transcriptome profiling showed that a highly significant propensity of transcriptional changes map to the engineered interval in the five assessed tissues. A statistically significant overrepresentation of the genes mapping to the entire length of the engineered chromosome was also found in the top-ranked differentially expressed genes in the mice containing rearranged chromosomes, regardless of the nature of the rearrangement, an observation robust across different cell lineages of the central nervous system. Our data indicate that a structural change at a given position of the human genome may affect not only locus and adjacent gene expression but also “genome regulation.” Furthermore, structural change can cause the same perturbation in particular pathways regardless of gene dosage. Thus, the presence of a genomic structural change, as well as gene dosage imbalance, contributes to the ultimate phenotype.
| Mammalian genomes contain many forms of genetic variation. For example, some genome segments were shown to vary in their number of copies between individuals of the same species, i.e. there is a range of number of copies in the normal population instead of the usual two copies (one per chromosome). These genetic differences play an important role in determining the phenotype (the observable characteristics) of each individual. We do not know, however, if such influences are brought about solely through changes in the number of copies of the genomic segments (and of the genes that map within) or if the structural modification of the genome per se also plays a role in the outcome. We use mouse models with different number of copies of the same genomic region to show that rearrangements of the genetic materials can affect the phenotype independently of the dosage of the rearranged region.
| Copy number variation (CNV) of genomic segments among phenotypically normal human individuals was recently shown to be surprisingly frequent [1],[2]. It covers a large proportion of the human genome and encompasses thousands of genes [3],[4]. About 58,000 human CNVs from approximately 14,500 regions (CNVRs) have been identified to date (http://projects.tcag.ca/variation/). They contribute to genetic variation and genome evolution [5]–[8] by modifying the expression of genes mapping within the CNV and in its flanks [9]–[13]. Consistently, initial cases of adaptive CNV alleles under positive selection were recently uncovered [14] and several structural variants were shown to be associated with “genomic disorders” [15]–[17] and susceptibility to disease (reviewed in [7],[18]–[21]). For example, a microdeletion and its reciprocal microduplication at chromosomal band 17p11.2 were shown to be associated with Smith-Magenis (SMS; OMIM#182290) and Potocki-Lupski syndromes (PTLS; OMIM#610883), respectively [22]–[24]. The Retinoic Acid Induced gene 1 (RAI1; GeneID: 10743) is thought to be the main dosage-sensitive gene within this genomic interval. Consistently, SMS patients with only RAI1 mutation have been identified [25]–[28]. However, accumulating evidence indicates that other factors also contribute to the spectrum of clinical findings in patients. For example, SMS patients with RAI1 mutations are less likely than SMS patients with the deletion to be short and suffer from hearing loss, cardiovascular, and renal tract abnormalities. On the other hand, they are at higher risk for obesity [29]–[33]. Mouse models of these syndromes were generated. These engineered animals recapitulate several of the multiple phenotypes present in the human patients. The SMS mice show craniofacial abnormalities, obesity, overt seizures, hypoactivity levels, and circadian rhythm anomalies, while the PTLS model is underweight and presents hyperactivity, learning and memory deficiencies, and social impairment [11],[27],[34],[35].
We took advantage of these models and of a third strain that is a compound heterozygote balanced for copy number—it harbors the SMS deletion on one allele and the PTLS duplication on the other—to tease apart the phenotypic consequences of gene dosage alterations versus genomic structural changes.
The functional impact of CNV of a given genomic interval remains unstudied at a genome-wide scale. Such a global assessment is achievable nowadays using the mouse as a model organism. Mouse models of the Smith-Magenis and Potocki-Lupski syndromes carry a deletion (strain Df(11)17/+) and its reciprocal duplication (Dp(11)17/+), engineered rearrangements involving the syntenic genomic regions at band MMU11B2, respectively [11],[22]–[24],[27],[28],[30],[34],[35]. These heterozygous mice and their wild type littermates (+/+) allow the study of the influence of one, two, and three copies of the same CNV in an otherwise identical genomic background (see below). A fourth strain (Df(11)17/Dp(11)17) obtained by mating the Dp(11)17/+ and Df(11)17/+ animals enables the generation of genomically balanced mice with two copies of that same CNV in cis, while they are in trans in +/+ animals (see Figure 1 for a schematic representation of the four genotypes).
To investigate the phenotypic outcome of modifying gene dosage or of maintaining gene dosage but with a structural change, we assessed 14 different phenotypes in the four different mouse genotypes (i.e., 1n, 2n, 3n, and 2n compound heterozygote) (Table 1). The decreased embryonic survival, craniofacial abnormalities, overt seizures, and altered neuromotor function observed in Df(11)17/+ and the learning and memory impairments shown by Dp(11)17/+ animals were absent in the genetically balanced Df(11)17/Dp(11)17 mice (summarized in Table 1; for details see Text S1, Figures S1–S2 and Table S1). Likewise, the significant differences in body weight and abdominal fat found in the SMS and the PTLS mouse models when compared to +/+ animals were absent in Df(11)17/Dp(11)17 (Text S1 and Figure S3). Furthermore, we found that “backing out of the test tube,” when confronted by a wild type mouse, was only correlated with copy numbers but not with structural changes per se (Text S1 and Figure S4). A summary of phenotypic differences between Rai1 +/− and Df(11)17/+ mice can be found in Text S1.
Anxiety was found increased in Dp(11)17/+ mice in the elevated plus maze test [11]. We found an overall significant difference in the percentage of observations in the open arms (F(3, 87) = 5.9; p = 0.001) and closed arms (F(3, 87 = 8; p<0.0001). Post-hoc analysis showed that Dp(11)17/+ mice spend significantly more time in the closed arms (62.1%±3%) than their wild type littermates (51%±1.9%) (p = 0.002). In contrast, the percentage of observations in the open arms was significantly increased for Df(11)17/+ mice (37%±2.5%), when compared with +/+ animals (29%±1.9%) (p = 0.023). The percentage of observations in the open arm was also significantly increased for Df(11)17/Dp(11)17 mice (36%±2.2%), when compared with +/+ (p = 0.045), however the p value is in the borderline range. The number of observations of Df(11)17/Dp(11)17 mice in the center and the close arm was always smaller than that of wild type. This is concordant with what we observed for Df(11)17/+ mice. While none of these differences are significant, both Df(11)17/+ and Df(11)17/Dp(11)17 mice behave similarly. No significant differences were observed when Df(11)17/Dp(11)17 were compared to the Df(11)17/+ mice (p>0.05). These results indicate that dosage of genes mapping within the engineered genomic interval is associated with the levels of anxiety in mice, since the gain or loss of genetic material are giving opposite phenotypes. However, structural changes play a role, as restoration of the number of copies (2n in cis) does not rescue the phenotype (Figure 2 and Table 1). This observation was similar to what was found for activity levels in the open field (Table1) [27].
Dp(11)17/+ mice showed a subtle impairment in the preference of a social target versus an inanimate target and a clear impaired preference for social novelty when compared to +/+ mice [11] in the three-chamber test [36] that is based on the tendency of a subject mouse to approach and engage in social interaction with an unfamiliar mouse. We performed this test in the four different groups of purebred mice with distinct CNV genotypes. The analysis of the sociability part of the test showed a significant effect of chamber side (F(1, 90) = 38.99, p<0.0001). Post-hoc analysis demonstrated that mice from all analyzed genotypes spend more time in the chamber side that contains the stranger 1 versus the side with the empty container (p<0.01 in all cases) (Figure 3A). In the preference for social novelty data, we observed a significant difference for chamber side (F(1, 90) = 9.6, p = 0.0025) and genotype (F(3, 90) = 5.74, p = 0.0012). Post-hoc analysis revealed that wild type (p = 0.04) and Df (11)17/+ mice (p = 0.0002) tend to spend significantly more time with stranger 2 than with stranger 1, but Dp(11)17/+ and Df(11)17/Dp(11)17 mice spent the same amount of time with stranger 1 and stranger 2 (p = 0.37 and 0.87, respectively). Moreover, when +/+ mice were compared with the other three genotypes we found that they spend significantly less time in the side of the stranger 1 than the Dp(11)17/+ mice (p = 0.0002) and Df(11)17/Dp(11)17 mice (p = 0.0003), but no significant differences were found when compared to Df(11)17/+ mice (p>0.05). In aggregate, these results suggest that gene copy number variation is playing a role in the preference to social novelty and that the duplication or deletion of this genomic interval is giving an opposite phenotype. Surprisingly, the response to social novelty is also modified in Df(11)17/Dp(11)17 mice, notwithstanding that gene dosage is normalized (Figure 3B and Table 1), suggesting that genomic structural changes are playing a role in this phenotypic outcome.
The phenotypic findings in mice prompted us to assess the effect of changing the number of copies of the SMS/PTLS CNV on tissue transcriptomes. We analyzed genome-wide expression levels in five organs affected in human patients (cerebellum, heart, kidney, testis, and hippocampus) from adult male individuals (at least three animals of each of the strains carrying one, two in trans, two in cis, and three copies of the MMU11B2 region; see Materials and Methods).
We ranked and chromosomally mapped the most differentially expressed transcripts. As anticipated, we observed in each of the analyzed tissues a significant overrepresentation of transcripts mapping to the rearranged interval (which we named SMS/PTLS genes; see legend of Figure 1 or Materials and Methods for a complete list of loci mapping to the engineered interval) amongst the top 100 (31 to 40 transcripts depending on the tissue) and top 1,000 (33 to 50 transcripts) most differentially expressed transcripts (all p<1×10−4, tested with permutations; Figure 4A–B). The expression levels of the transcripts, which vary in number of copies amongst the different strains, are compared in Figure 4C. We found a positive correlation between gene dosage and expression consistent with partial results already published [11]. These transcripts are expressed on average at 66%±15% of the level measured in wild type in Df(11)17/+ (one copy) and 138%±29% in Dp(11)17/+ animals (three copies). In particular, the expression levels of the murine orthologs of the two genes RAI1 (GeneID: 10743) and SREBF1 (6720), which were associated with schizophrenia [37]–[39], a phenotype absent from SMS and PTLS patients [33],[40],[41], show a strong relationship with gene dosage. The SMS/PTLS genes are, however, unchanged in Df(11)17/Dp(11)17 mice (1.02-fold (SD = 0.16) more, two copies in cis) compared to normal controls (two copies in trans), analogous to results recently obtained from cell lines of a man who carried a 22q11 deletion on one allele and a reciprocal duplication on the other allele [42]. Note that the loxP site inclusions necessary for the mouse engineering induced the loss-of-function of one Cops3 copy (GeneID: 26572) (Figure 1) [34], thus Df(11)17/Dp(11)17 and Dp(11)17/+ animals have only a single and two active copies of this gene, respectively. Consistently, we found Cops3 relative expression level to be downregulated in the compound heterozygous animals and unchanged in the PTLS mouse model (Figure 4C). The Df(11)17/+ and Df(11)17/Dp(11)17 strains carry two and three copies of Zfp179 (a.k.a. Rnf112, GeneID: 22671), respectively (Figure 1 and 4C), thus this gene could be considered in the “flanking” genes category in some strains (see below).
To confirm the transcriptome profiling results, we independently measured by Taqman quantitative PCR the relative expression levels of 43 genes in the hippocampus and cerebellum of males (N = 3) and females (N = 3) and the cortex, liver, and lung of female mice (N = 3) of the +/+, Dp(11)17/+, and Df(11)17/+ genotypes. The list of genes and assays used are presented in Table S2. They map either centromeric, within, or telomeric to the rearranged region. We found good reproducibility of the data for the three genes that were quantified with two different Taqman assays (Figure S5). Likewise, we noted a robust correlation between the Taqman and expression microarray results (correlation coefficient, R2 = 0.87; Figure S5A). Furthermore, the assays performed on female tissues demonstrated that the above described influences on the expression levels of genes situated within the rearrangement are not restricted to one sex and to the five tissues monitored by microarray (Figure S5B and S5C). Thus, the altered expression of SMS/PTLS genes are most probably relevant to the development of the phenotypic manifestations of PTLS and SMS mouse models that are absent in the Df(11)17/Dp(11)17 animals.
A second category of transcripts, those that map to the rest of mouse chromosome 11 (MMU11 genes), was significantly enriched within the top 1,000 most differentially expressed transcripts in all five tissues (all p<1×10−4, 97 to 138 transcripts, Figure 4B; “Most-diff” set of data, see below and Materials and Methods). This “flanking effect” might not be an effect of structural changes but could potentially be caused by linkage disequilibrium between the engineered interval and flanking polymorphisms. Consistently, retention of large blocks from the parental strain through genetic selection even after repeated backcrossing has been reported [43]–[45].
The SMS and PTLS mouse models were generated from a different genetic background (i.e., the AB2.2 ES cell line derived from a 129S5 mouse, see [34] for details) and were backcrossed for 12 generations to C57BL/6J-Tyrc-Brd. Genotyping of the entire length of MMU11 revealed that, whereas the region proximal to the engineered interval had recombined, the distal section had either only partially or not recombined at all to the C57BL/6J background in Dp(11)17/+ and Df(11)17)/+, respectively (Figure S6A). These sequence variants may have a significant impact on microarray-based transcriptome profiling [46]–[48]. For example, almost half of the reported 100 most significant cis-acting expression QTLs could be attributed to sequence diversity in probe regions in [46].
We thus devised a strategy to identify and discard the transcripts that could possibly be influenced by their 129S5 genetic makeup rather than by the modification of the number of copies of the CNV. As we found that 129S5 and 129S2 mice were genetically identical at all tested loci from the SMS/PTLS engineered interval to the telomere, we thought to use expression data previously established in our laboratory with the same microarray platform (GEO Series accession number: GSE10744) [12] to identify the transcripts that show a different level of expression between 129S2 and C57BL/6J animals in at least one of six major tissues (brain, liver, testis, kidney, lung, and heart) (false discovery rate<0.1; corrected for multiple testing) and that thus should be removed from our analysis (see Materials and Methods, Figure S7). This allowed establishment of a restricted set of data, named Most-diff-restricted, in which these transcripts were excluded (the unrestricted set was named “Most-diff”; see Materials and Methods). Within this constrained set, we found again that the SMS/PTLS transcripts were significantly enriched within the top 1,000 most differentially expressed transcripts in all five tissues analyzed (Most-diff-restricted set: all p<1×10−4, 26 to 40 transcripts). Similarly, the transcripts that map to the rest of the MMU11 chromosome were significantly enriched within the top 1,000 most differentially expressed transcripts in the cerebellum and hippocampus (Most-diff-restricted set: p<1×10−3, 94 and 103 transcripts, respectively) but not in the other three monitored tissues.
One could argue that this class of transcripts is still overrepresented in the two neuronal tissues because we were unsuccessful in identifying and discarding all transcripts that are influenced by linkage disequilibrium. Hence, to further assess a potential bias caused by the linkage disequilibrium between the engineered interval and flanking polymorphisms, we compared in three different tissues (cerebellum, kidney and testis) the relative expression of genes before and after their recombination to a C57BL6/J homozygous genetic background. We measured by quantitative PCR the relative expression levels of genes showing significant differences in expression between Dp(11)17/+ and +/+ in the microarray profiling experiments (see above) and mapping to the 11:76843886–92963733 interval in Dp(11)17/+ mice after 12 and 17 backcrosses (129S5/C57BL6/J heterozygous and C57BL6/J/C57BL6/J homozygous background, respectively) and compared it to that of wild type littermates (Figure S6A). The different assays are presented in Table S3. We found that 7 out of 14 (50%), 8 out of 16 (50%) and 12 out of 16 (75%) of the genes we studied in testis, kidney, and cerebellum, respectively, showed a change in expression level between the PTLS model and controls even after recombination, suggesting that the observed differences in expression of these genes are independent of the genetic background and not caused by linkage disequilibrium of the engineered region (Figure S6B–S6D). These results justify the strategy used above to discard 129 out of 248 probesets that could possibly be influenced by their 129S5 background.
Contrary to what we observed for the genes that map to the rearranged intervals, the “flanking” transcripts presented no correlation between gene dosage of the SMS/PTLS CNV and their expression levels (Figure 4C–D and Figure S8). In fact a majority (>55% within Most-diff-restricted and >80% within Most-diff) of the MMU11 transcripts showed a similar change in expression level in the Df(11)17/+, Dp(11)17/+, and Df(11)17/Dp(11)17 animals compared to normal controls in all analyzed tissues. As an important proportion of the MMU11 genes that do not vary in number of copies appeared to be affected in a consistent manner in the engineered animals, it is unlikely that their expression is solely directly or indirectly controlled by one or a combination of the 34 genes mapping to the rearranged interval (see Figure 1 or the Materials and Methods section for the complete list of these genes). If this would have been the case we might anticipate observing opposite changes in expression in the SMS and in the PTLS mice (see above and below). Consistently, we observe similar expression levels not only in the mice with one or three copies but also in the balanced heterozygote animals with two copies in cis of the SMS/PTLS CNV. In this latter strain, these changes in expression levels of the MMU11 transcripts are identified, although we register no modifications of the expression levels of the SMS/PTLS transcripts (Figure 4C and Figure S8B). Similarly, the analogous changes in expression reported in the different engineered genotypes could not be explained by the retention of promoters driving the introduced selection markers, as a previously shown possible explanation we needed to control for (e.g., [49]–[51], reviewed in [52]), because different cassettes are maintained in the three different models, i.e. puromycin and neomycin resistance genes in Df(11)17/+ and Df(11)17/Dp(11)17 and Hprt, tyrosinase and K14Agouti genes in Dp(11)17/+ and Df(11)17/Dp(11)17 [34].
One mechanism explaining the observed deregulation of MMU11 transcripts might be the dissociation of these transcripts from their long-range regulatory elements, a phenomenon known as position effect [53]. If the changes in gene expression were caused by the physical separation of cis-acting regulatory elements mapping to the rearranged interval and MMU11 genes, we should expect an enrichment of affected genes close to the breakpoints (i.e., the loxP sites necessary for the mouse engineering [34]). This is only partially the case (Figure S9). In fact, we find genes with modified expression mapping on the entirety of mouse chromosome 11, for example, tens of megabases from the breakpoints, suggesting that other mechanisms of regulation might also be at play (Figure S9). We find, however, no correlation between the distance from the breakpoints and the extent of expression change (Figure S10). Many of the transcripts that show changes in relative expression appear to cluster in discrete groups along the chromosome. We tested this assumption using a modified version of the method described by Tang and Lewontin to infer significance (see Materials and Methods) [54],[55] but found no significant clustering of the modified transcripts. We thus infer that the observed clustering is simply due to the non-homogenous distribution of genes along mouse chromosome 11 (Figure S9B). Similarly, we found no significant enrichment of genes that neighbor CpG islands within the set of MMU11 CNV-affected transcripts (Most-diff-restricted: p<0.25; Most-diff: p<0.15 tested with permutations; see Materials and Methods), which could have suggested that these genes are expressed in many tissues [56]. We found, however, that the MMU11 transcripts modified in expression were expressed in a significantly greater fraction of the tissues we assessed (average 2.6, median 3) relative to other transcripts (1.8, 2; two-tailed p<2.2×10−16, Mann-Whitney U test). They are, however, not expressed at higher levels than their unchanged counterparts (Figure S11).
Interestingly, the two tissues that show a significant number of differentially expressed genes mapping to MMU11, i.e. hippocampus and cerebellum, are part of the central nervous system (CNS). This observation suggests that copy number changes may have more of an effect on normal copy neighboring genes expressed in the brain. Other reports have shown that genes expressed in the brain have changed less than have genes expressed in other tissues during evolution [57] and that CNV genes expressed in the brain are more tightly regulated than other CNV genes [12]. The stricter expression regulation of genes with a function in the CNS is possibly brought about by their increased interdependency through multiple feedback loops, common long-range cis-acting regulatory units, and/or changes in the chromatin conformation. Thus, suggesting that perturbation to such “higher order” genome organization would be more identifiable and consequential in the CNS. Consistently, the phenotypes that persist upon restoration of gene dosage, modification of activity, anxiety, and sociability levels, are most probably from a neurological origin. We identified gene(s) that are modified in their relative expression levels in the Df(11)17/Dp(11)17 mouse (see above). The comparison of the hippocampal and cerebellar transcriptomes of these mice with that of +/+ littermates showed that expression levels of genes involved in detection of stimuli, visual perception, as well as neuronal differentiation were modified and, thus ultimately, might be at the origin of the change in phenotypic outcome (Text S1, Table S4–S6).
Taken together our results indicate that structural changes per se, i.e. without changes in gene dosage, have genomic consequences on gene expression far beyond the locus whose structure is varied and that structural variation can profoundly modify the phenotypic outcome.
Copy Number Variants (CNVs), because of their prevalence, e.g. 10% of the mouse autosomal genome and 60% of its duplicated regions [12],[58], constitute important contributors to intraspecific genetic variation. Multiple human CNVs have been associated with diseases, susceptibility to diseases, and adaptation (reviewed in [7],[8],[18]–[20]).
We show that mouse models of Smith-Magenis and Potocki-Lupski syndromes, engineered to have one and three copies, respectively, of the mouse chromosome 11 (MMU11) band B2 region (Figure 1) present altered expression of the genes mapping within the rearranged interval and diametrically opposing phenotypes in body weight, percent fat, anxiety, preference for social novelty, dominant behavior, and activity levels (Table 1). Similarly, the deletion and reciprocal duplication of the 1q21.1 region are associated with micro- and macrocephaly, respectively [59], while the reciprocal diametric changes in head size were reported for 16p11.2 rearrangements [60],[61]. These observations and the associations of these genomic disorders with autism spectrum disorder (ASD) (1q21.1 duplication and 16p11.2 deletion) and schizophrenia (1q21.1 deletion and 16p11.2 duplication) [59],[61]–[72] lend support to the hypothesis that these conditions are at different ends of a spectrum related to evolution of the social brain [73],[74]. SMS and PTLS, like 1q21.1 and 16p11.2 rearrangements, are so-called genomic sister-disorders—disease mediated by duplications versus deletions of the same regions—with overlapping phenotypic traits (for a complete list, see [75]) in which conditions/phenotypes appeared to be linked to gene dosage. However, patients presenting ASD and 1q21.1 deletions or 16p11.2 duplications, as well as individuals with schizophrenia associated with 1q21.1 duplications or 16p11.2 deletions, were also reported ([61]–[65],[72]; reviewed in [74]), suggesting that some conditions might be due to altered gene(s) function(s) through both under- and overexpression. Alternatively, we can hypothesize that some phenotypes are not associated with a specific number of copies of a particular CNV but rather that the simple presence of a structural change at a given position of the human genome may cause perturbation in particular pathways regardless of gene dosage.
Murine genes mapping centromeric or telomeric to the SMS/PTLS rearrangement show analogous changes in expression. Specifically, a MMU11 gene over- or underexpressed in the SMS mouse model has more than 50% chance to be also over- or underexpressed in the PTLS mouse model, respectively. Remarkably, affected genes are mapping on the entirety of the chromosome and not only in proximity to the breakpoints. The uncoupling between the number of copies of the CNV genes and the phenotype, here the effect on expression of genes outside of the rearrangement, is further illustrated by the fact that we detect the same changes in expression in the compound heterozygote, i.e. a mouse model with a normal number of copies in a cis configuration (Figure 1). Concomitantly, this restoration of gene copy number within a structural change was shown not to rescue all phenotypic manifestations observed in the SMS and PTLS mice. Indeed some complex phenotypes such as activity, anxiety, and preference for social novelty were still present in these animals. These observations suggest a contribution of genomic structural changes to the final phenotypic outcome and experimentally document that simple gene dosage alone cannot account for these phenotypes. The non-concordant absence of compensation in Df(11)17/Dp(11)17 mice (i.e. Df(11)17/Dp(11)17 mice anxiety mimics the phenotype observed in the SMS model, while their preference for social novelty is similar to that of PTLS animals; Table 1) further uncovers the complexity resulting of CNV-related genomic alteration.
The activity levels measured in the open field test exemplify the interaction between gene dosage and final phenotypic outcome of a specific CNV. Df(11)17/+ mice are hypoactive while Dp(11)17/+ are hyperactive, hence the opposing phenotypes implicate gene dosage in the final outcome. Consistently, Rai1 +/− heterozygote and Rai1 transgenic mice were found to be hypo- and hyperactive in the open field, respectively [27],[76]. However, the compound heterozygote Df(11)17/Dp(11)17 and Dp(11)17/Rai- mice [27] are also hyperactive in the open field, establishing that we are confronted with a complex phenotypic outcome. In conclusion, the presence of a CNV generates a phenotype through gene dosage imbalance and/or the presence of genomic structural changes. Further studies are warranted to resolve the underlying causes and assess the relevance of our findings beyond genetically engineered model and/or rare and highly penetrant CNVs.
Although we performed a broad battery of behavioral experiments and studied the gene expression profile in five tissues to address different aspects of SMS/PTLS phenotypes, there are still other facets that are yet to be studied. One of the most significant and consistent phenotypes displayed by almost all SMS patients is sleep disturbance, including early sleep onset and offset, repeated and prolonged nocturnal awakening, as well as excessive daytime sleepiness (“sleep attacks”). Sleep disturbance in SMS is accompanied by intrinsically inverted melatonin rhythms and is often claimed by patients and their families as one of the most challenging aspects of the SMS spectrum [33],[77],[78]. We suggest that with approaches similar to this study, by combining expression analyses in the suprachiasmatic nucleus (SCN) and performing circadian experiments of the SMS mouse models, valuable insights can be gained also for this important SMS phenotype.
Importantly, our results suggest that the pathways through which CNVs (including both deletions and duplications) result in complex traits, particularly those involving the CNS, might include not only alteration of the expression of genes included in the rearranged interval but also the subtle modification of the regulation of gene(s) mapping to the rest of the rearranged chromosome. These changes in expression levels might be triggered by a position effect, modification of the chromatin structure, perturbation of chromatin loops, disruption of long transcript structure, reflection of a regulatory interaction between chromosome homologues (e.g. transvection), and/or repositioning within the nucleus of a genomic region (e.g., in [79]–[83]; reviewed in [5]). Consistently, a balanced translocation was shown to significantly modify transcriptome profiles [84]. The results presented here also suggest that the chromosome and its gene collection are not randomly devised. The location and order are maintained possibly in relation to a higher level genomic organization required for proper regulation.
The potential unidirectionality of the long-range effects of CNVs on gene expression and phenotypic outcome independent of copy number change that has been uncovered in this report poses an important challenge in appreciating the contribution of this class of variation to phenotypic features. To include this variable in genome-wide [85] as well as in eQTL association studies [10], it might be necessary to combine all rearrangements that differ from normality regardless of their directionality.
The materials and methods used for this report can be accessed online (Text S1).
|
10.1371/journal.pntd.0004346 | The Ecological Dynamics of Fecal Contamination and Salmonella Typhi and Salmonella Paratyphi A in Municipal Kathmandu Drinking Water | One of the UN sustainable development goals is to achieve universal access to safe and affordable drinking water by 2030. It is locations like Kathmandu, Nepal, a densely populated city in South Asia with endemic typhoid fever, where this goal is most pertinent. Aiming to understand the public health implications of water quality in Kathmandu we subjected weekly water samples from 10 sources for one year to a range of chemical and bacteriological analyses. We additionally aimed to detect the etiological agents of typhoid fever and longitudinally assess microbial diversity by 16S rRNA gene surveying. We found that the majority of water sources exhibited chemical and bacterial contamination exceeding WHO guidelines. Further analysis of the chemical and bacterial data indicated site-specific pollution, symptomatic of highly localized fecal contamination. Rainfall was found to be a key driver of this fecal contamination, correlating with nitrates and evidence of S. Typhi and S. Paratyphi A, for which DNA was detectable in 333 (77%) and 303 (70%) of 432 water samples, respectively. 16S rRNA gene surveying outlined a spectrum of fecal bacteria in the contaminated water, forming complex communities again displaying location-specific temporal signatures. Our data signify that the municipal water in Kathmandu is a predominant vehicle for the transmission of S. Typhi and S. Paratyphi A. This study represents the first extensive spatiotemporal investigation of water pollution in an endemic typhoid fever setting and implicates highly localized human waste as the major contributor to poor water quality in the Kathmandu Valley.
| Aiming to understand the ecology of municipal drinking water and measure the potential exposure to pathogens that cause typhoid fever (Salmonella Typhi and Salmonella Paratyphi A) in Kathmandu, Nepal, we collected water samples from 10 water sources weekly for one year and subjected them to comprehensive chemical, bacteriological and molecular analyses. We found that Kathmandu drinking water exhibits longitudinal fecal contamination in excess of WHO guidelines. The chemical composition of water indicated site-specific pollution profiles, which were likely driven by localized contamination with human fecal material. We additionally found that Salmonella Typhi and Salmonella Paratyphi A could be detected throughout the year in every water sampling location, but specifically peaked after the monsoons. A microbiota analysis (a method for studying bacterial diversity in biological samples) revealed the water to be contaminated by complex populations of fecal bacteria, which again exhibited a unique profile by both location and time. This study shows that Salmonella Typhi and Salmonella Paratyphi A can be longitudinally detected in drinking water in Kathmandu and represents the first major investigation of the spatiotemporal dynamics of drinking water pollution in an endemic typhoid setting.
| Enteric (typhoid) fever is a severe systemic infection and a common cause of community acquired febrile disease in many low-income countries in Asia and Africa [1]. The infection is triggered by the ingestion of the bacteria Salmonella Typhi (S. Typhi) and Salmonella Paratyphi A (S. Paratyphi A). Both S. Typhi and S. Paratyphi A are human restricted pathogens (they have no known animal reservoir) and is it acknowledged that they are transmitted through contaminated food and water or via contact with fecal matter from acute or chronically infected individuals [1]. However, the predominant route of infection has never been rigorously investigated in an endemic setting outside a conventional case/control study design [2,3].
Typhoid fever is a common infection in Kathmandu (the capital city of Nepal) and our previously generated serological data implies that the local population has longitudinal exposure to both of these endemic pathogens [3,4]. Further studies, generated through investigating the spatiotemporal dynamics of typhoid fever in Kathmandu predicted that both S. Typhi and S. Paratyphi A are more likely to be transmitted through contaminated water than via human-to-human transmission in this setting [5,6]. Significantly, we found that typhoid fever cases cluster in areas with a high density of urban water sources, which are gravity driven and, therefore, rationally located at lower elevations. The various urban water sources in Kathmandu are most commonly in the form of sunken wells (as found in many urban and rural settings in lower-income countries), piped supplies into large communal holding tanks or the more traditional stone waterspouts (hitis/dhunge dharas) [7]. The iconic stone spouts are common across the Nepalese capital and the water flow into these sacred locations is gravity-dependent (Fig 1a), replenished by rainfall and snowmelt from the surrounding Himalayan Mountains. Natural soft-rock aquifers act as reservoirs for ground water and ultimately the untreated water enters the stone spouts from the aquifers through a series of ancient porous underground channels. Our previous analysis specifically identified locations around these stone spouts are hotspots for S. Typhi and S. Paratyphi A infections [5,6].
Hypothesizing that the local water, particularly the water accessed via the stone spouts, is a substantial public health risk for typhoid fever and other enteric infections in Kathmandu, we aimed to longitudinally assess bacterial contamination, the chemical composition and the ecological dynamics of enteric bacteria in the water supply in this location. To address this hypothesis, and focusing on water sources accessed by the local population, weekly water samples were collected over a one-year sampling period from ten locations and subjected to various physical, chemical, microbiological and molecular analyses.
The water sources for this study were the ten most commonly used water sources (identified by questionnaire) lying within a previously identified typhoid fever hotspot in Lalitpur, Kathmandu [6]. The selected locations were GPS located using an eTrex legend (Garmin) and consisted of five stone spouts, three sunken wells and two piped supplies. The location of these water sources are shown in Fig 1b and described in detail in Table 1. Daily rainfall data from Kathmandu Airport was provided by the Nepalese Department of Hydrology and Meteorology (http://www.dhm.gov.np/) and aggregated into weeks for the purposes of the analysis presented here.
Water was collected (when permitted by water flow), from all of the 10 locations once per week over one year from May 2009 to April 2010. From each of the sources mid-flow water samples were collected in two sterile bottles in volumes of 1 L and 500 ml. From the stone spouts and the piped supply, the stopper was aseptically removed and free flowing water was allowed to flow directly into the sterile bottle. For wells, a sterilized steel bucket (bleached and washed with autoclaved water prior to use) was lowered into the well until it was partially submerged, the bucket was then removed and the water was poured into the bottles. All the bottles were labeled with the source code, date and time of collection of the samples. After recording the water temperature the bottles were transported to the laboratory at ambient temperature and were processed within one hour for physical, chemical and microbiological analysis.
The Kathmandu Water Engineering Laboratory (http://www.sodhpuch.com/water-engineering-training-centre-p.) performed all chemical and physical analyses following their standard operating procedures for international water quality. The measured variables were pH (Hanna pH meter, calibrated with pH4, pH7 and pH9 buffers), temperature (Hanna digital thermometer with probe), conductivity (Hanna conductivity meter), color (Perkin Elmer’s LAMDA 650 UV spectrophotometer at 270 nm) turbidity (NEPHELOstar Plus nephlometer), hardness (EDTA titration), total alkalinity as CaCO3 (methyl orange), chloride (argentometric titration), ammonia (nesslerisation), total nitrate (Perkin Elmer’s LAMBDA 650 UV spectrophotometer at 275 nm), total nitrite (Perkin Elmer’s LAMDA 650 UV spectrophotometer at 275 nm), and trace elements and heavy metals (Atomic Absorption Spectrophotometric (AAS) method). All variables were recorded on the day of sampling and compared to WHO guidelines for water quality [8].
A modified most probable number (MPN) method was used to assess the microbiological quality of the water, specifically coliform contamination [9,10]. Briefly, five ten-fold serial dilutions were made from each water sample by inoculating 1 ml of undiluted water sample into 9 ml of MacConkey broth (Oxoid, UK). This was continued until a dilution of 1 x 10−5. A total of 30 tubes (five tubes for each dilution) were prepared for each sample. The inoculated broths were incubated at 44°C for 48 hours for the culture of thermotolerant coliforms. After incubation, each tube was examined and those that were positive (production of acid and gas) were counted. The number of positive and negative tubes in each of these three sets was noted in order and these data were used to estimate the coliform content using a five-tube MPN table.
To detect the presence of enteric bacteria with pathogenic potential (e.g. Salmonellae, Shigellae, Vibrionaceae, and E. coli) 10, 20, 50, 100 and 500 μl of undiluted water was directly plated onto Xylose lysine deoxycholate (XLD) and MacConkey agar plates. The plates were incubated at 37°C overnight and then observed for growth. To increase the likelihood of culturing Salmonellae and Shigellae, 100 ml of undiluted water was filtered through a membrane filter with a pore size 0.45 μm (Whatman, GE Life Sciences, PA, USA) using a sterile syringe. The filter paper was removed using sterile forceps and placed in 90 ml of typtic soya broth (Oxoid, UK). The soya broth bottles were agitated using a vortex to displace the organisms on the membrane and incubated for 18 hours at 37°C. After overnight incubation 1 ml of the pre-enrichment culture was transferred to 10 ml of selenite broth. Further, 1 ml of the pre-enrichment culture was transferred to 10 ml of Rappaport-Vassiliadis Broth (RVB). The incubated overnight broth was then plated onto XLD and MacConkey agar plates. The plates were incubated overnight at 37°C and then observed for growth. For the detection of Vibrionaceae, 1 ml of the undiluted water sample was diluted in 9 ml of alkaline peptone water. The suspension was then incubated overnight at 37°C and then plated on to MacConkey, XLD and thiosulphate-citrate-bile salts sucrose (TCBS) agar.
The colony morphologies including the form, size, surface appearance, texture, color, elevation and margin of all individual colonies were recorded from MacConkey and XLD plates. Of special interest were colonies that were circular, with an entire margin and slightly raised elevation that were non-lactose fermenting on both plates, with or without the production of hydrogen sulphide on the XLD plate. Individual colonies with the aforementioned characteristics were isolated and plated on nutrient agar and incubated at 37°C for 24 hours. Isolated colonies obtained on the nutrient agar plates were then subject to API20E testing to identify Enterobacteriaceae and other non-fastidious Gram-negative rods.
Total DNA from all water samples was extracted using the Metagenomic DNA Isolation Kit for Water (Epicentre Biotechnologies, WI, USA). Water samples were centrifuged at 1,000 rpm (Hettich Zentrifugen, EBA 21, Germany) for 5 minutes to remove large debris and then decanted into sterile containers. After centrifugation, 100 ml of the centrifuged water was filtered through a pre-sterilized filter with a pore size of 0.45 μm (Whatman, GE Life Sciences, PA, USA). Using sterile forceps and scissors the membrane was removed from the filter apparatus and cut into four pieces. The cut filters were then placed in a 50 ml sterile conical tube with the upper surface of the filter facing inwards. One milliliter of filter wash buffer containing 0.2% Tween-20 was added to the filter pieces in the tubes to remove organisms on the filter surface. The tube was agitated at high speed for approximately 2 minutes with intermittent breaks. The cell suspension was transferred to a clean micro-centrifuge tube and centrifuged at 14,000 X g (Thermo Fischer Scientific, IEC Micro CL17, Germany) for 2 minutes to pellet the cells. The supernatant was discarded. The cell pellet was re-suspended in 300 μl of TE buffer, and 2 μl of ready-lyse lysozyme solution and 1μl of RNAse A were added and mixed thoroughly. The tube was incubated at 37°C for 30 minutes and then 300 μl of 2 X meta-lysis solutions and 1 μl of Proteinase K were added to the tube and thoroughly mixed by vortexing. To ensure that all the solution was at the bottom of the tube, the tube was pulse centrifuged. The tubes were then incubated at 65°C for 15 minutes. The solution was cooled to ambient temperature and placed on ice for 5 minutes. 350 μl of MPC protein precipitation reagent was added to the tube and mixed thoroughly by vortexing vigorously for 10 seconds. The debris was pelleted by centrifugation for 10 minutes at 14,000 X g (Thermo Fischer Scientific, IEC Micro CL17, Germany) at 4°C. The supernatant was transferred to a clean micro-centrifuge tube and the pellet was discarded. To the supernatant, 570 μl of isopropanol was added and mixed by inverting the tube multiple times. The DNA was pelleted by centrifugation for 10 minutes at 14,000 X g (Thermo Fischer Scientific, IEC Micro CL17, Germany) at 4°C. The isopropanol was removed and the sample was briefly pulse centrifuged and any residual liquid was removed without disturbing the pellet. To the pellet 500 μl of 70% ethanol was added without disturbing the pellet. The tube was then centrifuged for 10 minutes at 14000 X g (Thermo Fischer Scientific, IEC Micro CL17, Germany) at 4°C. Ethanol was removed without dislodging the DNA pellet and the sample was briefly pulse centrifuged and any residual fluid was removed without disturbing the pellet. The pellet was then air dried for 8 minutes at ambient temperature before being resuspended in 100 μl of nucleic acid free sterile water (Epicentre Biotechnologies, WI, USA).
Quantitative Real-time PCR was performed on all extracted DNA to detect DNA sequences specific for S. Typhi and S. Paratyphi A as previously described [11]. Primer and probe sequences were as follows; S. Typhi; ST-Frt 5' CGCGAAGTCAGAGTCGACATAG 3', ST-Rrt 5' AAGACCTCAACGCCGATCAC 3', ST- Probe 5' FAM-CATTTGTTCTGGAGCAGGCTGACGG-TAMRA 3'; S. Paratyphi A; Pa-Frt 5'ACGATGATGACTGATTTATCGAAC 3', Pa-Rrt 5' TGAAAAGATATCTCTCAGAGCTGG 3', Pa-Probe 5' Cy5-CCCATACAATTTCATTCTTATTGAGAATGCGC-BHQ5 3'. Briefly, 5 μl of environmental DNA extractions (as above from 100 ml of water and resuspended in 100 μl of nucleic acid free sterile water) was used as the template for each experiment, i.e. 5 μl equated to 5 ml of water sample. Quantification was performed using standard curves where plasmid DNA carrying the target sequences were diluted in 10-fold serial dilutions ranging from 100 to 105 plasmid copies per μl; standard curves for assessing S. Typhi and S. Paratyphi A copy number were constructed by plotting the Ct value against the plasmid DNA copy number.
For the 16S rRNA gene surveying, variable regions 3 to 5 (V3–V5) of the 16S rRNA gene were PCR amplified from the water DNA extractions. The primers used were as described previously [12], see S1 Table for the full barcode and primer sequences (30 nucleotides for 454 adaptor, 12 nucleotides for unique recognition (tag) and 18 nucleotides to amplify the specific V3–V5 region) used for each sample in the present study. The conditions of PCR were as follows: 1U of AccuPrime Taq DNA Polymerase High Fidelity (Invitrogen, Carlsbad, CA USA), 200 mM of forward and reverse primer, 2 μl of template environmental DNA in a 20 μl reaction. The reaction was cycled for 1 x 94°C for 2 minutes, and then 20x (94°C for 30 seconds, 53°C for 30 seconds and 68°C for 2 minutes). Each sample was PCR amplified on four occasions, the resulting amplicons were pooled and then ethanol precipitated before resuspension in 20 μl of TE. The 16S rRNA gene amplicons were shipped to The Wellcome Trust Sanger Institute and pooled together into an equimolar mastermix, as measured by a Qubit fluorometer (Invitrogen, Carlsbad, CA, USA), prior to sequencing on a GS FLX Titanium 454 machine (Roche Diagnostics, Oakland, CA USA) using the Lib-L kit. The resulting sequence data is available at the European Nucleotide Archive under Study Accession Number ERP004371/Sample Accession number ERS373486. Sequence data was processed using the mothur software package (http://www.mothur.org/), following a previously described protocol [12]. This removed poor quality reads, and generated taxonomic classifications for each Operational Taxonomic Unit (OTU). Following these filtering steps 326,155 sequences remained (range of 1 to 7,396 sequences per sample).
We first tested for geographic differences in bacterial assays, using the non-parametric MANOVA implemented in the package ade4 [13] for the R software suite [14]. 9,999 random permutations of the data were used to assess statistical significance of Pillai’s statistic [15] and compute the associated p-value. After ruling out the presence of geographic differences between samples, data were aggregated across locations by computing average weekly profiles, from which temporal trends were more straightforward to investigate. As bacterial assays may each capture different aspects of water contamination, these data were subjected to a centered Principal Component Analysis (PCA) [13], which we used to derive a latent variable (the first principal component, PC1) as correlated as possible to all the different assays [16] and therefore reflecting the extent of bacterial contamination. PCA is ideally suited to derive synthetic variables, which capture the essential trends of variation in quantitative or binary data, and is thus readily applicable to water quality data including physical chemical properties, bacterial assays and meta-genomic variation.
The temporal trends in PC1 were visualized using ggplot2 [17] and modeled using a cubic spline of sample collection dates. Five breakpoints were used in the model as they gave the best visual fit, and no other number of breakpoints led to significantly better models. This model was compared to a model where PC1 was constant in time using a classical ANOVA comparing the residual variances of the two models.
As for bacterial assays, the various chemical properties measured captured potentially different aspects of water pollution. PCA was used to identify the main trends of variation amongst water samples [18,19]. Because measurements were made using different units and had inherently different scales of variation, a centered and scaled PCA was used [20]. Missing data were replaced by the average of the corresponding variables, as is customary in PCA [20]. Results of this first PCA were driven by an outlier, which turned out to be a sample of exceptionally poor water quality. In such cases, because PCA finds linear combinations of variables with maximum variance, the presence of an outlier may conceal other interesting structures [20]. As a consequence, this sample was removed in a second PCA, the results of which are shown in Fig 2. Differences in water chemistry of the water sources were tested using the same MANOVA procedure used in bacterial assays analysis.
16S rRNA gene survey data consisting of 93 samples and 11,212 OTUs were first transformed into compositional data [21], so that each sample was transformed into a composition of OTUs frequencies summing to 1. This transformation ensures that further analysis will only reflect differences in taxa composition, and not in absolute quantities of sequenced DNA. The Gini-Simpson index was computed for every sample as 1-∑ipi2 where pi is the relative frequency of OTU i in the sample. A centered PCA was used to analyze the metagenomic profiles, retaining the first five principal axes as they expressed most of the structured variation in the data (Fig 3a). The proportion of the total variation represented by the jth principal axis was computed as λj/∑j λj (Fig 3b). The contribution of an OTU i to a principal axis U defined by a vector of loadings [u1,u2,…,u11,212] was computed as ui2, which is justified by the fact that U has a norm of 1 (thus ∨U∨2 = ∑iui2 = 1). The OTUs contributing most to the retained principal axes were defined as OTUs with contributions greater than a given threshold (Fig 3c). Different sets of OTUs were defined using thresholds of 1%, 2%, 5%, 10%, 20%, 30%, 40%, and 50%, which resulted in retention of between 4 and 14 OTUs. A threshold of 10% was retained as it allowed for conserving essentially all of the variation of the 7 principal axes with only 10 OTUs, which still represented 80% of the total variation in the data.
Discriminant Analysis of Principal Components (DAPC, [22,23]) was applied to the 16S rRNA gene data to identify combinations of OTUs that differed most between stone spouts and well. While originally developed for genetic markers data, this method has since been applied to various other types of data, including 16S rRNA (e.g. [24,25]). Cross-validation was used to assess the optimal number of PCA axes to retain in the preliminary dimension-reduction step, using 100 independent replicates for each number of retained PCA axes and 70% of the samples as training set. The same analyses were repeated to investigate possible differences across the three locations.
A summary of the chemical and bacterial data generated for each of the 10 sampling locations (432 water samples) is presented in Table 2 (total data available in S1 Dataset). We firstly assessed the physical qualities and chemical composition of the water samples and compared these data to WHO guidelines (Table 2) [8]. The chemical analyses of the water samples signified that several sources had concentrations of iron and ammonia in excess of WHO guidelines, but notably, only nitrate levels and turbidity consistently exceeded WHO recommendations in all locations. These finding, with respect to nitrate were broadly consistent with previous single time point observations in this location [26,27]. Further investigation identified significant differences in the chemical profiles of the water from the various locations (non-parametric MANOVA: Λpillai = 0.250,p = 1×10−4 with 9,999 permutations). These disparities in chemical compositions could be summarized using a PCA, with water from two of the sunken wells (locations 4 and 6 (Table 1 and Fig 1b)) forming distinct clusters with independent chemical signatures (Fig 1c). These profiles were characterized by consistently high concentrations of iron and ammonia and greater turbidity at location 4 and greater conductivity, hardness, chlorides and nitrates at location 6 (Fig 1c). Distinctively, one of the piped water supplies (location 8) had consistently lower chemical contamination indices than all other sources (Fig 1c and Table 2). The differences in chemical composition between locations, confirmed by pairwise comparisons between the sampling locations (Wilcoxon rank test, all p-values <0.05 with Bonferroni correction), suggest water mineralization and implicate contaminants such as vehicle exhaust gases [28,29] and poor waste handling systems as the key drivers of poor water quality in these locations [30].
Of the different chemical pollutants observed in Kathmandu drinking water, the sustained contamination of water by nitrites and nitrates was the most alarming. Nitrites and nitrates can be introduced into the water through a range of processes including surface water infiltration, industrial pollution, agricultural fertilizer run-off and the leakage of sewerage systems [26,31]. Sustained exposure to nitrates can lead to a range of non-communicable diseases including methemoglobinemia, gastrointestinal cancer, bladder and ovarian cancers, and may expedite type II diabetes, thyroid hypertrophy and respiratory tract infections [32]. Nitrate excess is also a notable marker of fecal contamination, and we found nitrites and nitrates correlated positively with coliform concentration (see below) and weekly rainfall (p<0.001; Spearman’s rho). Consistently, the concentration of chloride, again a marker of sewage and manure contamination [33], also increased with the onset of seasonal rainfall (p<0.001; Spearman’s rho). More generally, several other chemical properties were also associated with weekly rainfall; positive correlations were observed for rainfall against turbidity, ammonia and hardness (Spearman’s rho, p-values <0.05 in six locations), and negative correlations were observed with rainfall against pH and alkalinity (Spearman’s rho, p-values <0.05 in four locations). Taken together, these results suggest that chemical pollution of drinking water in this setting is likely driven by a combination of rainfall runoff and localized contamination with human fecal waste.
To determine the extent of potential fecal contamination in the water samples, we estimated the concentration of fecal indicator thermotolerant coliforms using the minimal probable number (MPN) method. The WHO guidelines state that no water directly intended for drinking, or in the distribution system should contain thermotolerant coliforms (in a 100ml sample) [8]. We found that majority of the cultured water samples were contaminated with thermotolerant coliforms, suggesting that all the sampled water sources are prone to fecal contamination (Table 2). The concentrations of cultured coliforms across the samples ranged from 0.1 to 2.5 x 108 CFU/100 ml, these figures are again largely consistent with prior investigations of drinking water quality in this setting [8,26,31]. The water from the stone spouts (locations 1, 2, 3, 9 and 10 (Table 1) had higher coliform concentrations (median 94 CFU/100 ml; IQR: 4 to 1.6 x 104) than that from the sunken wells (locations 4, 5, 6) (median 8 CFU/100 ml; IQR: 1 to 7.9 x 104) and from the piped supplies (locations 7, 8) (Median 2 CFU/100 ml; IQR: 1 to 170) (p<0.001; Kruskal Wallis) (Fig 2a). The highest coliform concentrations across the sampled water sources were in the months of June, July and August and, similar to the chemical contamination, positively correlated with the period of increased weekly rainfall (Spearman’s rho = 0.27, p<0.001). The water sampled from source 2 (stone spout) consistently had the highest level of coliform contamination (median coliform concentration; 1.3 x 104 CFU/100 ml; IQR: 200 to 1.68 x 105) and had the highest single coliform count of 2.5 x 108 CFU/100 ml in August.
The alarming levels of coliform and chemical contamination highlight that water quality is a major public health issue in Kathmandu. The detection of thermotolerant coliforms is not suitable for the identification of specific waterborne pathogens, but is used as a general measure of bacterial contamination [34]. To address this methodological limitation, we concurrently performed membrane filtration and microbiological enrichment for a range of pathogenic bacterial species classified as high risk by the WHO, including Salmonella, Shigella, Vibrionaceae and E. coli [8]. Overgrowth on plates diminished the ability to accurately quantify these organisms but we repeatedly cultured and identified a wide range of organisms with pathogenic potential to humans including Vibrio cholerae 01, Shigella dysenteriae type-1, Pseudomonas spp., and Plesiomonas shigelloides (S2 Table). We additionally isolated multiple Salmonella spp. in the water through supplementary enrichment, yet, after serotyping, none were identified as S. Typhi or S. Paratyphi A.
The lack of confirmative cultures for S. Typhi or S. Paratyphi A is not uncommon given previous efforts to culture these pathogens from water. S. Typhi has been cultured from environmental samples previously [35], but is notoriously difficult to isolate from water in endemic locations. It has been suggested that S. Typhi is often present and viable, but in a non-culturable state [36]. To address this limitation we extracted and purified total nucleic acid from each of the water samples after filtration and performed appropriately controlled quantitative real-time PCR on all samples for chromosomal targets specific for S. Typhi and S. Paratyphi A. We found that 333 (77%) and 303 (70%) of 432 DNA extractions from the eater samples were PCR amplification positive for S. Typhi and S. Paratyphi A, respectively (Fig 2b), with 266/432 (62%) samples being PCR amplification positive for both. To confirm that the PCR amplicons were S. Typhi and S. Paratyphi A, a random cross section of 96 (48 for each serovar) PCR amplicons were successfully cloned, sequenced and confirmed to originate from either S. Typhi or S. Paratyphi A. Further, through inference from standard curve, we identified a significant difference between the number of copies/reaction between S. Typhi (median 208 copies/reaction; IQR: 72 to 603) and S. Paratyphi A (median 11 copies/reaction; IQR: 4.6 to 28), corresponding with inferred significantly different medians of 4,200 and 2,200 copies/100 ml, respectively (p<0.001; Kruskal Wallis) (Fig 2b).
The presence of S. Typhi and S. Paratyphi A DNA in the water samples did not differ significantly across locations (non-parametric MANOVA: Λpillai = 0.020,p = 0.66 with 9,999 permutations). Therefore, we propose that processes of fecal contamination of water with S. Typhi and S. Paratyphi A operate non-locally and are likely to equally contaminate water sources in this area. This observation confirms previous work predicting that infection with a specific S. Typhi genotype is a random process [6]. This lack of geographical structure allowed us to pool data from all locations, gaining additional power to investigate the temporal trends of the average weekly PCR amplification profiles through a PCA. The first principal component (PC1) captured a gradient of S. Typhi and S. Paratyphi A DNA that exhibited a marked increase during weekly periods of rainfall in all water sources, and a subsequent decrease during the dry season (May–October) (Fig 2c). A simple seasonal model of PC1 (using a spline of the collection dates with five breakpoints) showed that this temporal trend was a significantly better fit than a model where PC1 was constant over time (ANOVA: F = 4.2184,p = 0.0034). Notably, the increased presence of S. Typhi and S. Paratyphi A in the water samples displayed a substantial temporal lag, with PC1 reaching a peak between one to two months after the end of the monsoon rains (Fig 2c). This delay may reflect the time taken for the organisms to reach the water outlet from the source of contamination or a concentration effect reflecting lower ground water levels in drier periods. We speculate that a leaking sewerage system and fluctuations in the pressure in the water supply pipes (negative pressure in the pipe results in an influx of sewage into the water pipe) are a likely source of this contamination.
While the trend of contamination by S. Typhi and S. Paratyphi A was similar across all sampled locations, these organisms likely represent a microscopic fraction of the diverse communities of microorganisms transiting through these water sources [37]. Assessing these complex microbial communities is relevant for investigating the extent of contaminating organisms present in these water sources and also aids tracing the likely sources of bacterial contamination. (i.e. soil and/or fecal waste). 16S rRNA gene surveying provides a suitably broad approach in conducting this type of investigation [38]. Therefore, the longitudinal structures of the bacterial communities in the two water sources with the greatest coliform concentrations (locations 2 and 5) were compared by 16S rRNA gene amplification and pyrosequencing (data available in S2 Dataset) [12]. The resulting 16S rRNA gene data showed that the bacterial communities in 93 tested water samples were composed of 11,212 OTUs, more than half of which were observed only once (S1 Fig). The bacterial diversity, as measured by Gini-Simpson’s index [39], was high in all analyzed samples (S2 Fig). We observed an increase in the average number of taxa found during the dry season (S3 Fig), suggesting that low rainfall induces a concentration effect and that exposure to a wider range of taxa likely increases during this period.
For a more comprehensive analysis, we investigated the bacterial diversity using a PCA of the 16S rRNA gene survey data transformed into OTU frequency profiles. This analysis showed that >80% of the variation amongst the analyzed samples could be summarized in seven dimensions, each representing a different assemblage of the detected OTUs (S4 Fig). Closer examination of the relative contributions of OTUs to each axis revealed that nearly identical principal components could be obtained through a combination of the 10 most common OTUs, which belonged to the genera Acinetobacter (OTUs 00001 and 0019), Acidovorax (OTU00011), Comamonas (OTU00025), Flavobacterium (OTU00042), Bacillus (OTUs 00503 and 00155), Chryseobacterium (OTU00208), Staphylococcus (OTU00436) and Brevundimonas (OTU00707) (S4 Fig).
The ten most commonly detected OTUs were comprised of bacterial genera that are typically found in the environment, and most are not recognized as enteric organisms. While many of these genera are known aquatic organisms, there is some overlap with genera that have recently been demonstrated to originate from sample preparation procedures [40]. To address this limitation, we specifically focused on the relative contributions of nine bacterial families that are common constituents of the gastrointestinal tract of mammals (Bacteroidaceae, Clostridiaceae, Enterobacteriaceae, Erysipelotrichaceae, Lachnospiraceae, Lactobacillaceae, Prevotellaceae Ruminococcaceae and Veillonellaceae) [41]. This sub-analysis (519 OTUs) showed that 80% of the variation could be summarized in only five dimensions (Fig 3a and 3b). Furthermore, almost identical principal components could be obtained through a combination of just six OTUs, which belonged to the Enterobacteriaceae (OTUs 00024, 00185 and 01479), Bacteroidaceae (OTU00979), Clostridiaceae (OTU00263), and Prevotellaceae (OTU 2149) (Fig 3c). We additionally found that the OTU composition was not random, and varied substantially between the sunken well and the stone spout. A DAPC analysis of the 16S rRNA gene data was used to identify combinations of the 519 gastrointestinal OTUs showing the greatest difference between the well (location 5) and the stone spout (location 2). The corresponding DAPC was able to recover the type of water source of the samples (based on their OTU composition) in 81% of cases (Fig 3c). The OTUs exhibiting the greatest variation (OTU00263; Clostridium, OTU00185; Enterobacteriaceae, OTU00979; Bacteroides and OTU00024; Enterobacteriaceae) were additionally four of the six principal fecal OTUs driving the temporal trends (Fig 3d). These results suggest the existence of local ecological factors, such as proximal sewage pipes, strongly affect the composition of bacterial communities in different types of water sources.
The current daily demand for water in Kathmandu is estimated to be 200,000 m3/d, but is unable to be met by municipal supplies, resulting in deficits of 70,000 to 115,000 m3/d in the wet and dry seasons, respectively [26]. This water deficit means that drinking water distribution lines operate at low pressure as compared to the overused sewage pipes that are often co-located. This pressure differential between sewage and drinking water lines, their relative proximity and poor state of repair accounts for the ingress of sewage into drinking water lines and an overall deterioration in their quality and safety. Furthermore, the cost associated with municipal water has led to an upsurge in the use of stone spouts in the middle and low-income residents of Kathmandu. The users of stone spouts now represent >20% of the water usage in Kathmandu [26]. Bacterial contamination has previously been shown to be higher in stone spouts than sunken wells than piped supplies in Kathmandu, with similar trends seen with nitrate contamination [26]. Our data confirm these associations and for the first time we show seasonal variations in fecal contamination, chemical pollutants and, vitally, exposure to both S. Typhi and S. Paratyphi A. We suggest that stone spouts are the most contaminated type of water source in this location as they are maximally exposed to potentially contaminating sources of human sewage. Further, the shallow aquifers supplying the stone spouts are likely more contaminated than deep aquifers, as they are, again, closer to the sources of potential contamination.
Our work shows that Kathmandu drinking water exhibits year round fecal contamination and is far from compliant with World Health Organization (WHO) standards [8] and it will be challenge to meet the United Nations sustainable development goal six (http://www.un.org/sustainabledevelopment/water-and-sanitation/#). It has been known for >100 years that the provision of filtered water has a dramatic effect on communicable diseases. In a classic paper by Sedgwick and McNutt they describe the observed decrease in mortality from typhoid fever and other infections when clean water is supplied to an urban population [42]. Taking these historic findings in account, we surmise that stone spouts and sunken wells represent a major public health risk to those in Kathmandu, not only for typhoid but also for other communicable and possibly non-communicable diseases. The chemical composition of drinking water indicates localized, site-specific pollution profiles, consistent with complex populations of enteric bacteria, which show both temporal and location specific profiles. For the first time we provide a molecular proxy for S. Typhi and S. Paratyphi A persistently transiting through an urban water source, where they appear to reach a peak concentration one to two months after the end of the monsoon. Whilst these results are very specific to this setting, many of the mechanisms facilitating such a large degree of contamination (water shortages, negative-pressure pipes, urban development, poor sanitation) are likely to be common in other typhoid endemic settings in Asia and beyond.
In conclusion, our work shows that municipal water in the capital city of Nepal exhibits evidence of substantial bacterial and chemical contamination. Further, we additionally show evidence of longitudinal contamination of both S. Typhi and S. Paratyphi A, demonstrating the impact of human fecal contamination and outlining that both these organisms are being continually transmitted through these water supply systems. Future research should focus on investigating the main routes of these bacterial and chemical pollutants into this water supply system. Further, the ability to culture and genotype organisms from the environment and human infection would close the circle on the role of the water delivery systems in urban settings for the transmission of typhoid fever. Nepal is amongst the poorest countries in Asia (GDP per capita of USD 703 in 2013 [43]) and substantial investment is required to improve the capacity and quality of the water supply in addition to the sewage handling systems in Kathmandu. As a rapid improvement of the water systems is unlikely to occur given the recent earthquake and ongoing political difficulties, we advocate the use of home water filters and sterilization systems alongside vaccination campaigns as the major public health interventions for typhoid fever prevention in this setting.
|
10.1371/journal.ppat.1000005 | Methylated DNA Recognition during the Reversal of Epigenetic Silencing Is Regulated by Cysteine and Serine Residues in the Epstein-Barr Virus Lytic Switch Protein | Epstein-Barr virus (EBV) causes infectious mononucleosis and is associated with various malignancies, including Burkitt's lymphoma and nasopharyngeal carcinoma. Like all herpesviruses, the EBV life cycle alternates between latency and lytic replication. During latency, the viral genome is largely silenced by host-driven methylation of CpG motifs and, in the switch to the lytic cycle, this epigenetic silencing is overturned. A key event is the activation of the viral BRLF1 gene by the immediate-early protein Zta. Zta is a bZIP transcription factor that preferentially binds to specific response elements (ZREs) in the BRLF1 promoter (Rp) when these elements are methylated. Zta's ability to trigger lytic cycle activation is severely compromised when a cysteine residue in its bZIP domain is mutated to serine (C189S), but the molecular basis for this effect is unknown. Here we show that the C189S mutant is defective for activating Rp in a Burkitt's lymphoma cell line. The mutant is compromised both in vitro and in vivo for binding two methylated ZREs in Rp (ZRE2 and ZRE3), although the effect is striking only for ZRE3. Molecular modeling of Zta bound to methylated ZRE3, together with biochemical data, indicate that C189 directly contacts one of the two methyl cytosines within a specific CpG motif. The motif's second methyl cytosine (on the complementary DNA strand) is predicted to contact S186, a residue known to regulate methyl-ZRE recognition. Our results suggest that C189 regulates the enhanced interaction of Zta with methylated DNA in overturning the epigenetic control of viral latency. As C189 is conserved in many bZIP proteins, the selectivity of Zta for methylated DNA may be a paradigm for a more general phenomenon.
| γ herpesviruses are characterized by their life-long persistence in the infected host. This is due in part to their ability to establish latency in infected cells. Epstein-Barr virus (EBV) is almost ubiquitous within the human population. The virus establishes latency in B-lymphocytes and is thought to reactivate and undergo replication following physiological stimuli that lead to B-cell activation. During latency the DNA genome of EBV is methylated, resulting in epigenetic control of viral gene expression. A viral protein that is key to viral reactivation and replication, Zta, is unique amongst transcription factors in displaying enhanced binding to methylated DNA sites. Here, we define the molecular interactions that predispose Zta to interact with methylated binding sites and we identify a mutant of Zta that has lost the preference for methylated sites. This allows us to probe the relevance of the recognition of methylated DNA for the reactivation of EBV from latency and to suggest that EBV has evolved a mechanism to allow it to specifically recognize methylated DNA and overturn epigenetic silencing of its genome imposed by the host.
| Methylation of DNA is generally associated with inhibition of gene expression. This is mediated in part by the association of specific methyl-CpG binding proteins with methylated DNA, leading to transcriptional silencing and chromatin remodeling [1], and in part by the inability of some transcription factors to bind to methylated DNA [2]. A notable exception is the bZIP transcription factor Zta (also known as BZLF1, ZEBRA, Z), which displays enhanced recognition for methylated DNA [3],[4].
Zta is encoded by Epstein-Barr virus (EBV), a human herpes virus that infects the majority of the world's population. EBV causes infectious mononucleosis and is linked to malignancies such as endemic Burkitt's lymphoma, nasopharyngeal carcinoma, and Hodgkin's disease [5]. EBV infects then establishes long-term latency in B-lymphocytes [6],[5]. During latency, the EBV genome is heavily methylated and few viral genes are expressed. Disruption of EBV latency is sporadic, characterized by expression of the majority of the EBV gene complement, replication of the genome and release of virus [7],[8].
Zta is the first viral gene expressed during the switch to lytic replication. As well as enhancing its own expression in a positive feedback loop [9], Zta activates a second viral transcription factor, Rta, encoded by BRLF1 [10]. Both Zta and Rta are essential for viral replication and together promote expression of the remaining viral lytic genes [11]. The BRLF1 promoter, referred to as Rp, contains three Zta-response elements (ZREs), two of which (ZRE2 and ZRE3) include CpG motifs that are subject to methylation. A pivotal study demonstrated that the interaction of Zta with ZRE2 and ZRE3 is enhanced by methylation [3], a phenomenon believed critical for lytic cycle activation.
The basic region of Zta's bZIP domain contains a cysteine residue, C189, which regulates the redox-sensitivity of DNA-binding activity [12]. Substituting this cysteine with serine (ZtaC189S) is sufficient to prevent reactivation of EBV from latency and EBV genome replication [13],[12]. Here we explore the molecular basis for the dramatic effects of this point mutation. We demonstrate that C189 is critical for the activation of BRLF1 expression and for recognition of the methylated ZREs in Rp, both in vivo and in vitro. A previous study had shown that the point mutation S186A compromises Zta's ability to bind both the methylated and non-methylated forms of the ZREs within Rp [4]. We propose a model in which S186 and C189 contact the two cytosine methyl groups of a specific CpG motif, which is conserved between ZRE2 and ZRE3. The relevance of a regulatory role for C189 in methylated DNA recognition by Zta is discussed.
We and others have previously shown that altering a single cysteine residue within the DNA contact region of Zta (ZtaC189S) impairs its ability to disrupt EBV latency [13],[12]. Here we asked whether ZtaC189S is competent to initiate one of the earliest events in latency disruption, the transcriptional activation of BRLF1 (Figure 1A). Expression vectors encoding Zta and ZtaC189S were introduced into a Burkitt's lymphoma derived cell line, Raji, and subsequent transactivation of the endogenous viral Rp was assessed (Figure 1B). Basal expression of BRLF1 is low in the absence of Zta, but is enhanced 33-fold following Zta expression. In contrast, expression of an equivalent amount of ZtaC189S resulted in only 4-fold enhancement of BRLF1 expression. Therefore, ZtaC189S is severely compromised for its ability to transactivate the BRLF1 gene in Raji cells.
To investigate the molecular basis for this defect, we compared the ability of Zta and ZtaC189S to bind to the three ZREs present in Rp (Figure 2). All three of the ZREs contribute to Rp activity and are important for activation of the endogenous viral gene by chemical stimuli [14], although ZRE1 appears to be dispensable on a fully-methylated template [3]. Zta and ZtaC189S were produced by coupled transcription and translation in vitro and their ability to interact with each ZRE was assessed by EMSA (Figure 2). As summarized in Figure 2F, the proteins were equally capable of binding to non-methylated ZRE1 and ZRE2 but, in agreement with previous reports for Zta [3],[12], neither Zta not ZtaC189S showed a detectable interaction with non-methylated ZRE3. Thus, the reduced ability of ZtaC189S to transactivate BRLF1 is not due to an inherent defect in binding to the ZREs.
We next asked what effect methylation of ZRE2 and ZRE3 had on binding by Zta and ZtaC189S (ZRE1 lacks a CpG motif and hence is not subject to methylation). Note that methylation yields four methyl-cytosines in ZRE3 (two on each strand) but only two in ZRE2 (Figure 2). In line with previous reports, methylation of ZRE3 converted it from a marginal to a strong binding site for Zta and methylation of ZRE2 also enhanced binding (Figure 3 and [3],[12],[4]). Importantly, the ability of ZtaC189S to bind meZRE3 was markedly reduced compared to that of wild-type Zta (Figure 3). Binding of ZtaC189S to meZRE2 was only modestly reduced, to approximately 75% compared to wild type. An independent mutation of C189 to alanine (ZtaC189A) also resulted in decreased binding to meZRE3, suggesting that the decreased binding of the C189S mutant was due to loss of the cysteine and not, for example, merely due to phosphorylation of the newly introduced serine.
To substantiate the above findings, we used chromatin precipitation to evaluate the binding of Zta and ZtaC189S to the BRLF1 promoter in vivo. Cells (293-BZLF1-KO) that contain an episomal EBV genome [11] were transfected with vectors encoding either His-tagged Zta or ZtaC189S and equivalence of expression was confirmed by immunoblotting (Figure 4). Zta- and ZtaC189S-associated chromatin complexes were isolated and the co-precipitated DNA was amplified using quantitative PCR with primers spanning the ZREs within Rp (Figure 4). As expected [12], ZRE sequences in the precipitated chromatin were clearly enriched, whereas regions lying 5′ and 3′ of Rp were not. Fine mapping of the chromatin complexes allowed us to differentiate between binding to ZRE1, ZRE2 and ZRE3. Whereas Zta and ZtaC189S bound equally well to ZRE1, the interaction between ZtaC189S and ZRE2 was partly compromised (relative to wild-type Zta) and its interaction with ZRE3 was completely eliminated. As Rp is fully methylated in cells harboring EBV [3], the in vivo chromatin association and the in vitro DNA-binding analyses correlate well. Taken together, these data demonstrate a critical role for C189 in the interaction of Zta with meZRE3 and, to a lesser extent, meZRE2.
As the interaction of Zta with ZRE3 is methylation-dependent and the site contains four methylated cytosine residues, we explored the relevance of each methylation site (see Figure 5A for numbering convention). A series of double-strand versions of ZRE3 was generated, each containing a single methyl-cytosine, and binding to Zta was assessed in a competition assay. Initial validation of the assay revealed that whereas a 20-fold excess of fully methylated ZRE3 competes efficiently for Zta binding, non-methylated ZRE3 fails to compete, even at a 100-fold excess (Figure 5B,C). Each singly methylated version of ZRE3 competed for binding to some degree (Figure 5D–G) implying that all four methyl groups contribute to the interaction with Zta. The order of competition efficiency was methyl cytosine1′>methyl cytosine−2>methyl cytosine0≥methyl cytosine−1′. Thus, methylation within the CpG motif that is common to both ZRE2 and ZRE3 gave stronger competition than within the motif uniquely present in ZRE3.
To better understand the above observations, we modeled the structure of Zta's DNA-binding domain bound to methylated ZRE3 and compared it to the previously reported model of Zta-bound meZRE2 [15]. In the following description, we designate residues within the two Zta monomers and the corresponding DNA half-sites as “Left” or “Right”, and the two CpG motifs as motifs 1 and 2, as summarized in Figures 6A and D. The modeled ZRE2 and ZRE3 structures, both based on the crystal structure of Zta bound to an AP-1 site, differ exclusively in the right half-site, as the two ZREs are identical except at base pairs 1 and 2 (Figure 6A).
In the crystal structure, Zta makes base-specific contacts with the AP-1 site via residues N182 and R190. N182 interacts symmetrically with base pairs ±2 while R190 makes asymmetric contacts: R190Left interacts with the central guanine base while R190Right contacts the DNA phosphate backbone. In the modeled ZRE structures, the contacts mediated by R190 are conserved whereas those by N182 are not – a direct consequence of the DNA sequence difference. More specifically, the N182 contacts are disrupted in the left half-site of ZRE2 (because ZRE2 diverges from AP-1 at base pair –2) and in both half-sites of ZRE3 (divergent at base pairs ±2). The additional disruption in the right half-site likely explains why Zta binds more weakly to ZRE3 than to ZRE2 ([3]; Figure 2D and E).
How does methylation of ZRE3 enhance Zta binding? The mechanism previously postulated for ZRE2 is that methylation of CpG motif 1 results in a direct contact between the cytosine1′ methyl group and S186Left, enhancing affinity by stabilizing a hydrogen bond network that involves S186Left, N182Left and the CpG motif's Guanosine2′ base ([15]; Figures 6C and 7A). The same mechanism can be proposed for ZRE3, as ZRE2 and ZRE3 have identical left half-sites. Also predicted to contribute to enhanced binding is a hydrophobic contact between C189Left and the cytosine−2 methyl group (Figure 7A,C). This interaction, not previously described, is common to both the meZRE2- and meZRE3-bound Zta models. Thus, S186Left and C189Left are postulated to interact with CpG motif 1 so as to simultaneously engage the two methyl groups, which are located on complementary DNA strands (Figure 7A). Such interactions are consistent with the ability of ZRE3 singly methylated on cytosine+1′ or cytosine−2 to compete with fully-methylated ZRE3 for Zta binding (Figure 5).
The ZRE3 model also suggests how methylation of CpG motif 2 influences Zta binding. The cytosine–1′ aromatic ring is in a cation-π interaction with the guanidino group of R190Left (Figure 6B and 7B) [16]. Methylation would enhance this interaction by increasing the amount of negative charge in the π-electron system (CH3- is an electron-donating substitutent) and by introducing an additional van der Waals contact with the arginine side chain. On the complementary strand, the Cyt0 methyl group contacts both S186Right and R190Right. Although R190Right does not interact with a DNA base, S186Right is within hydrogen bond distance of the base-contacting residue N182Right; hence, methylation may indirectly influence interactions involving base pair +2 or +3. Although the atomic details cannot reliably be predicted, these probably differ from those previously postulated for S186Left and CpG motif 1, as the stereochemical environments of S186 and N182 differ between the two half-sites. Nevertheless, the prediction that Zta residues contact both methyl groups of CpG motif 2 agrees with the ability of ZRE3 singly-methylated at cytosine−1′ or cytosine0 to compete with fully methylated ZRE3, albeit more weakly than when methylated within CpG motif 1 (Figure 5). Indeed, the observation that methylation on CpG motifs 1 and 2 has a differential effect on Zta binding is consistent with these motifs mediating non-equivalent interactions with Zta residues, due to their asymmetry relative to the ZRE3 pseudodyad (Figures 6D, 7A and 7B).
Our modelling predicts that C189Left (marked by an asterisk in Figures 6 and 7) is sensitive to the methylation state of CpG motif 1 via a direct contact involving the residue's thiol group and the cytosine−2 methyl group (Figure 7C). This interaction is favored by the hydrophobic nature of the cysteine side chain [17]. Mutation to the more polar serine residue would destabilize this contact by unfavorably juxtaposing the serine hydroxyl and Cyt−2 methyl groups. This agrees with the decreased binding affinity observed for ZtaC189S toward both meZRE2 and meZRE3 (Figure 3). Moreover, the absence of this thiol-methyl contact in non-methylated ZRE sites explains why the C189S mutation has little effect on the binding to non-methylated ZRE2 (Figure 2D). A possible alternative explanation, that an altered S186 or R190 conformation causes the decreased affinity, appears unlikely given that C189 contacts neither residue (see, for example, Fig. 6C).
In contrast to C189Left, C189Right is too far from any cytosine methyl group to form a direct contact (Figure 6B,C). However, in the ZRE3 model, C189Right is in van der Waals contact with the thymidine−2′ methyl group (Figure 7D). This interaction is symmetrically equivalent to that of C189Left with the cytosine−2 methyl group and consequently should also be destabilized by the C189S mutation (Figure 6D). The interaction is unique to ZRE3, as the corresponding base in ZRE2 is guanine−2′, which lacks a methyl group. Thus, the prediction is that the C189S mutation destabilizes contacts in both half-sites of meZRE3, but in only one of ZRE2. This agrees neatly with the finding that the mutation more severely impairs Zta binding to meZRE3 than to meZRE2 (Figure 3).
Our modelling predicts that C189 directly contacts the cytosine−2 methyl group within CpG motif 1, which is common to ZRE2 and ZRE3. This complements a previous prediction that S186 contacts the cytosine1′ methyl group in the same CpG motif [15] (Figure 8A). We tested our hypothesis by assessing the ability of Zta to bind versions of meZRE2 that omit methylation of either the 1′ or −2 cytosine residues. As predicted, omitting either methyl group significantly reduces binding to each (the binding is not entirely abolished, as Zta binds non-methylated ZRE2 with appreciable affinity) (Figure 8B). We next investigated the contribution of the cytosine−2 methyl group to the binding affinity of the S186A and C189S point mutants. Our model predicts that omitting this group should more significantly compromise the binding affinity (relative to wild type) of the S186A mutant over that of the C189S mutant, because only the former conserves the cysteine thiol group. As shown in Figure 8C, ZtaS186A binds more weakly to meZRE2 than does wild type Zta, in agreement with previous studies [4]; omitting the cytosine−2 methyl group further decreases the relative binding affinity, consistent with the loss of an important contact. This is in stark contrast with the results observed for ZtaC189S: although the relative affinity of this mutant toward meZRE3 is weak, it is not significantly further reduced upon omitting the cytosine−2 methyl group (Figure 8D). On the contrary, binding is comparable to that of the wild type, which is dramatically reduced compared to binding to the fully methylated ZRE3; the methyl group of cytosine−2 enhances the DNA binding affinity of wild type Zta to a greater extent than that of the C189S mutant. The combined results strongly support the hypothesis that C189 directly contacts the cytosine−2 methyl group.
Zta is the only known example of a transcription factor whose binding to specific DNA sequence elements is enhanced by CpG methylation [3]. In other contexts, the distinction between methylated and unmethylated DNA is primarily orchestrated by proteins lacking specific DNA-binding activity, such as the methyl binding proteins MeCP2, MDB1, MDB2, MDB4 and Kaiso [1],[18]. A previous study demonstrated that mutations of S186 affect the ability of Zta to interact with methylated ZRE2 and ZRE3 [4]. However, these mutations also alter interactions with non-methylated ZRE2 and ZRE3 and compromise ZRE1 recognition. In contrast, the C189S mutant is compromised for binding methylated ZRE3, and to a lesser extent ZRE2, but retains wild type affinity towards many unmethylated sites, including ZRE1 (Figures 2 and 3). Thus, the C189S mutant provides a highly selective tool to address the relevance of methylated ZRE recognition for the disruption of latency in vivo. Consequently, our results strongly corroborate the hypothesis that Zta binding to methylated ZREs is essential for the reactivation of latent EBV in B-lymphocytes. More specifically, because ZtaC189S is severely compromised for meZRE3 (but not meZRE2) binding, the inability of this mutant to activate BRLF1 expression in Raji cells suggests that meZRE3 recognition is particularly critical for Rp activation in this cell line (Figure 1).
Our present findings are in apparent disagreement with a previous study showing that the ZtaC189S mutant is only marginally compromised in reactivating BRLF1 expression [12]. The previous results, which we have verified (data not shown), were obtained using EBV-positive 293 cells (BZLF1-KO-293), an epithelial cell line. In contrast, we observe defective Rp activation in Raji cells, which derive from a Burkitt's lymphoma cell line. EBV infects and replicates in both B-lymphocytes and epithelial cells. However, latency and reactivation from latency in vivo is predominantly associated with B-lymphocytes making our findings particularly relevant to understanding how Zta activates Rp in a physiological setting. We speculate that the difference in Zta behavior in B-lymphocytes and epithelial cells may partly reflect differences in the milieu of cellular transcription factors or in chromatin structure. Indeed, a recent study showed that, compared to Raji cells, the level of ZRE2 methylation is slightly lower in BZLF1-ZKO-293 cells, and substantially lower in another epithelial cell line (AGS cells) [4]. Further studies are required to establish whether and how EBV lytic activation is regulated in a cell-specific manner.
Our structural model of Zta bound to meZRE3 reasonably accounts for a number of experimental observations. In particular, the model rationalizes why the binding affinity of Zta for unmethylated ZRE3 is lower than for ZRE2, why methylation increases the binding affinity for both sites, why the C189S mutation compromises methyl-ZRE recognition, and why the latter effect is more pronounced for ZRE3 than for ZRE2. Furthermore, the model suggests that Zta residues contact all four methyl groups in ZRE3's two CpG motifs (Figure 7), consistent with interactions observed for all four singly-methylated ZRE3 variants in a DNA-binding competition experiment (Figure 5). Although unable to explain why methylation within CpG motif 1 yields stronger binding than within motif 2, the model is nevertheless consistent with a differential effect, as it predicts different stereochemical environments for the two motifs.
A key prediction resulting from the model is that a specific C189 residue in the Zta homodimer is sensitive to methylation of the CpG motif common to ZRE2 and ZRE3 by means of a thiol-methyl group contact (Figure 7C). This contact is lost or destabilized by the replacement of C189 by A or S, consistent with the decreased DNA-binding affinity observed for these mutants (Figure 3). In contrast, the homodimer's second C189 residue is predicted to be remote from a CpG motif. In the case of ZRE3, this residue can form a thiol-methyl group contact with a thymine base (Figure 7D). The stability of the latter contact is predicted to be susceptible to the C189S mutation regardless of CpG methylation status. However, the decreased affinity of this mutant can only be observed when the binding to methylated ZRE3 is assessed, as the affinity of wild type Zta for unmethylated ZRE3 is too low to appreciate any further decrease. C189 is implicated in the redox sensitivity of Zta's DNA binding activity, and nitrosylation of this residue has been evoked as one possible mechanism by which nitric oxide down-regulates EBV reactivation [12]. Our results concerning C189 suggest how such regulatory phenomena might potentially be linked to methyl-ZRE recognition and Rp activation.
Our current working model is that S186 and C189 interact with both methyl-cytosines of a specific CpG motif to enhance the binding of Zta to a methylated ZRE, thereby overturning epigenetic silencing of the viral genome. Such methylation-enhanced affinity is conceivably unique to Zta, as no other bZIP proteins are known to posses a serine residue equivalent to S186. On the other hand, C189 is relatively conserved among bZIP proteins: in a sequence alignment of 50 human bZIP proteins, over half conserve this residue, including c-Fos and c-Jun. It is therefore tempting to speculate that CpG methylation may enhance the affinity of certain cellular bZIP proteins for their cognate DNA target sites. This tantalizing potential for Zta's cellular homologs to overturn the epigenetic silencing of genes awaits further investigation.
293-BZLF1-KO cells [11] and Raji cells were maintained as described previously [13],[12].
Raji cells were transfected as described previously [13],[19]. The sequences of the primers used in determining BRLF1 gene expression are as follows:
5′-CAGAAAGTCTTCCAAGCCATCC and 5′-CAAACAGACGCAGCCATGA.
Western blot analysis was determined as described previously [19].
Expression vectors encoding histidine tagged Zta and ZtaC189S were generated by amplifying the coding sequence with the following oligonucleotides, which incorporate a hexa-histidine repeat at the amino terminus and sub-cloned into pcDNA3 (Invitrogen) using BamH1 and EcoR1.
5′CTGCACACCGGGGATCCATGCATCATCATCATCATCATATGATGGACCCAAACTCGACTTCT and 5′CTGCACACCGGGGAATTCTTAGAAATTTAAGAGATCCTCGTGTAA
Zta, ZtaC189A, ZtaS186A were generated by site directed mutagenesis.
Chromatin was prepared from 293-BZLF1-KO cells 48 hours post transfection [20]. His-tagged protein complexes were isolated on HIS-Select Nickel Affinity Gel slurry (Sigma-Aldrich). Primers adjacent to each ZRE were used to amplify the sites.
ZRE1F (cggctgacatggattactgg);
ZRE1R (tgatgcagagtcgcctaatg);
ZRE2F (cagcagagaggctcggtt);
ZRE2R (tgcaatatttcctccagaaa);
ZRE3f (ggacaagatgtcactcttt);
zre3r (gggaagaaagtatagctac);
Rta3′F (TCCCTGTATTCACTGAGCGTCG);
Rta3′R (GGTCCCTTTGCAGCCAATGC);
Rta 5′F (CTTCGGGATAGTGTTTCAGG);
Rta 5′R (CTCAGCCCGTCTTCTTACC).
Radio labeled probes were generated with [33P] or [32P] and analyzed as described previously [21],[22],[13],[19]. In vitro translated proteins were generated in a rabbit reticulocyte lysate system or a wheat germ translation system, radio labeled with [35S] methionine. Competition EMSA was undertaken using unlabeled Zta protein and a radio labeled Zta binding site (ZIIIB).
The 5′ oligonucleotide sequences of the ZRE1, 2 and 3 probes (Invitrogen) and their methylated versions, with methyl-cytosine marked as O, (Sigma) are as follows:
ZRE1:5′-GATCTCTTTTATGAGCCATTGGCA-3′
ZRE2:5′-GATCATAAAATCGCTCATAAGCTT-3′
ZRE3:5′-GATCTATAGCATCGCGAATTTTGA-3′
ZRE2-meth:5′-GATCATAAAATOGCTCATAAGCTT-3′
ZRE3-meth:5′-GATCTATAGCATOGOGAATTTTGA-3′
Δ1′ or Δ-2 versions of the methylated primers were also synthesized (Sigma).
The crystal structure of Zta bound to a 19-mer DNA duplex containing an AP-1 site (PDB accession id 2C9L) was used to model Zta bound to a methylated ZRE3 site. Base-pair replacements converting an AP-1 sequence to that of ZRE3 (see Figure 4A) were made in program O [23], with bases set to adopt ideal Watson-Crick base-pairing and the template DNA backbone kept fixed. Methylation at a given C:G base pair position was modeled by least-squares superposition of a m5C:G base pair taken from the crystal structure of the self-complementary DNA duplex CCAGGC(m5C)TGG [24]; PDB accession id 2D25. Figures 6 and 7 were prepared using Bobscript [25] and Raster3d [26].
EBV genome used type 1 NC_007605; Zta (BZLF1) swiss prot P03206; Burkitt's lymphoma OMIM # 113970. |
10.1371/journal.pcbi.1000702 | Mathematical Modelling of Cell-Fate Decision in Response to Death Receptor Engagement | Cytokines such as TNF and FASL can trigger death or survival depending on cell lines and cellular conditions. The mechanistic details of how a cell chooses among these cell fates are still unclear. The understanding of these processes is important since they are altered in many diseases, including cancer and AIDS. Using a discrete modelling formalism, we present a mathematical model of cell fate decision recapitulating and integrating the most consistent facts extracted from the literature. This model provides a generic high-level view of the interplays between NFκB pro-survival pathway, RIP1-dependent necrosis, and the apoptosis pathway in response to death receptor-mediated signals. Wild type simulations demonstrate robust segregation of cellular responses to receptor engagement. Model simulations recapitulate documented phenotypes of protein knockdowns and enable the prediction of the effects of novel knockdowns. In silico experiments simulate the outcomes following ligand removal at different stages, and suggest experimental approaches to further validate and specialise the model for particular cell types. We also propose a reduced conceptual model implementing the logic of the decision process. This analysis gives specific predictions regarding cross-talks between the three pathways, as well as the transient role of RIP1 protein in necrosis, and confirms the phenotypes of novel perturbations. Our wild type and mutant simulations provide novel insights to restore apoptosis in defective cells. The model analysis expands our understanding of how cell fate decision is made. Moreover, our current model can be used to assess contradictory or controversial data from the literature. Ultimately, it constitutes a valuable reasoning tool to delineate novel experiments.
| Activation of death receptors (TNFR and Fas) can trigger either survival or cell death according to the cell type and the cellular conditions. In other words, the same signal can have antagonist responses. On one hand, the cell can survive by activating the NFκB signalling pathway. On the other hand, it can die by apoptosis or necrosis. Apoptosis is a suicide mechanism, i.e., an orchestrated way to disrupt cellular components and pack them into specialized vesicles that can be easily removed from the environment, whereas necrosis is a type of death that involves release of intracellular components in the surrounding tissues, possibly causing inflammatory response and severe injury. We, biologists and theoreticians, have recapitulated and integrated known biological data from the literature into an influence diagram describing the molecular events leading to each possible outcome. The diagram has been translated into a dynamical Boolean model. Simulations of wild type, mutant cells and drug treatments qualitatively match current data, and predict several novel mutant phenotypes, along with general characteristics of the cell fate decision mechanism: transient activation of some key proteins in necrosis, mutual inhibitory cross-talks between the three pathways. Our model can further be used to assess contradictory data and address specific biological questions through in silico experiments.
| Engagement of TNF or FAS receptors can trigger cell death by apoptosis or necrosis, or yet lead to the activation of pro-survival signalling pathway(s), such as NFκB. Apoptosis represents a tightly controlled mechanism of cell death that is triggered by internal or external death signals or stresses. This mechanism involves a sequence of biochemical and morphological changes resulting in the vacuolisation of cellular content, followed by its phagocyte-mediated elimination. This physiological process regulates cell homeostasis, development, and clearance of damaged, virus-infected or cancer cells. In contrast, pathological necrosis results in plasma membrane disruption and release of intracellular content that can trigger inflammation in the neighbouring tissues. Long seen as an accidental cell death, necrosis also appears regulated and possibly involved in the clearance of virus-infected or cancer cells that escaped apoptosis [1].
Dynamical modelling of the regulatory network controlling apoptosis, non-apoptotic cell death and survival pathways could help identify how and under which conditions the cell chooses between different types of cellular deaths or survival. Moreover, modelling could suggest ways to re-establish the apoptotic death when it is altered, or yet to trigger necrosis in apoptosis-resistant cells. The decision process involves several signalling pathways, as well as multiple positive and negative regulatory circuits. Mathematical modelling provides a rigorous integrative approach to understand and analyse the dynamical behaviours of such complex systems.
Published models of cell death control usually focus on one death pathway only, such as the apoptotic extrinsic or intrinsic pathways [2],[3],[4]. A few studies integrate both pathways [5], some show that the concentration of specific components contribute to the decision between death and survival [6],[7] while other studies investigate the balance between proliferation, survival or apoptosis in specific cell types along with the role of key components in these pathways [8], but no mathematical models including necrosis are available yet. Moreover, we still lack models properly demonstrating how cellular conditions determine the choice between necrosis, apoptosis and survival, and how and to what extent conversions are allowed between these fates.
Our study aims at identifying determinants of this cell fate decision process. The three main phenotypes considered are apoptosis, non-apoptotic cell death (which mainly covers necrosis) and survival. Although the pathways leading to these three phenotypes are highly intertwined, we first describe them separately hereafter, concentrating on the players we chose to include in each pathway. Summarised in Figure 1A, this description does not intend to be exhaustive, but rather aims at covering the most established processes participating in cell fate decision.
Only the apoptotic caspase-dependent pathway downstream of FAS and TNF receptors is considered here. Upon engagement by their ligands and in the presence of FADD (FAS-Associated protein with Death Domain), a specific Death Inducible Signalling Complex (DISC-FAS or DISC-TNF in Figure 1) forms and recruits pro-caspase-8. This leads to the cleavage and activation of caspase-8 (CASP8). In the so-called type II cells, CASP8 triggers the intrinsic or mitochondria-dependent apoptotic pathway, which also responds to DNA damage directly through the p53-mediated chain of events (not detailed here). CASP8 cleaves the BH3-only protein BID (not explicitly included in the diagram), which can then translocate to the mitochondria outer membrane. There, BID competes with anti-apoptotic BH3 family members such as BCL2 for interaction with the proteins BAX or BAK (BAX will stand here for both BAX and BAK). Consequently, oligomerisation of BAX results in mitochondrial outer membrane permeabilisation (MOMP) and the release of pro-apoptotic factors. Once released to the cytosol, cytochrome c (Cyt_c) interacts with APAF1, recruiting pro-caspase-9. In presence of dATP, this enables the assembly of the apoptosome complex (referred to as ‘Apoptosome’ in Figure 1A, lumping APAF1 and pro-caspase-9), responsible for caspase-9 activation, followed by the proteolytic activation of pro-caspase-3 (CASP3) [9]. By cleavage of specific targets, the executioner caspases (CASP3 in the model) are responsible for major biochemical and morphological changes characteristic of apoptosis. SMAC/DIABLO (SMAC) is released during MOMP to the cytosol, where it is able to inactivate the caspase inhibitor XIAP [10]. CASP3 also participates in a positive circuit by inducing the activation of CASP8 [11],[12]. In type I cells, CASP8 directly cleaves and activates executioner caspases such as CASP3 (not described).
Here, we consider mainly a mode of cell death with morphological features of necrosis, which occurs when apoptosis is impeded in cells treated with cytokines [13] or in some specific cell lines such as L929 cells when exposed to TNF [14]. In primary T cells, if caspases are inhibited, activation of TNFR or FAS causes necrosis via a pathway that requires the protein RIP1 and its kinase activity (RIP1K) [13]. This RIP1-dependent cytokine-induced necrotic death defines necroptosis [15],[16]. A genetic screen recently identified other genes necessary for this type of cell death [17]. However, a precise description of this pathway is still lacking. Reactive oxygen species (ROS) were proposed to be involved downstream of RIP1 [18]. ROS are also thought to play a key role in the control of mitochondria permeability transition (MPT), since they are produced by damaged mitochondria and can oxidize mitochondrial components, thus favouring MPT [19],[20],[21]. Furthermore, the role of mitochondria in necrosis is highlighted through the involvement of MPT, which causes a fatal drop in ATP level and leads to necrotic death. Indeed, MPT results from the inhibition of ATP/ADP exchange at the level of mitochondrial membranes, or from the inhibition of oxidative phosphorylation decreasing cellular ATP level and causing energy failure [21],[22]. Although there is evidence that necrosis is also triggered by TNF- and FAS-independent pathways, these are not yet considered in this study. These pathways include, for example, calpain-mediated cleavage of AIF followed by its nuclear translocation [23],[24], or PARP-1-mediated NAD+ depletion [24],[25].
NFκB represents a family of transcription factors that play a central role in inflammation, immune response to infections and cancer development [26]. The ubiquitination of RIP1 at lysine 63 by cIAP leads to the activation of IKK and ultimately that of NFκB [27]. In different cell types, especially in tumour cell lines, activation of NFκB inhibits TNF-induced cell death [28]. This effect is mediated by NFκB target genes: cFLIP inhibits recruitment of CASP8 by FADD [29]; anti-apoptotic BCL2 family members inhibit MOMP and MPT [30],[31],[32]; XIAP acts as a caspase inhibitor [33]; and ferritin heavy chain [34] or mitochondrial SOD2 [35] decrease ROS levels (these mechanisms are represented in Figure 1A by a direct inhibitory arrow from NFκB to ROS). For the sake of simplicity, other NFκB target genes that are known to inhibit TNF-induced apoptosis are provisionally omitted in the model (e.g., A20; cf. [36],[37]).
Our goal here is to provide a simplified but yet rigorous model of the mechanisms underlying cell fate selection in response to the engagement of FAS and TNF receptors. We have proceeded in several steps. First, we have assembled a regulatory network covering the main experimental data. Species and interactions were selected on the basis of an extensive literature search and integrated in the form of a diagram or “regulatory graph”. This diagram is then translated into a dynamical model. Our analysis initially focuses on the determination of the asymptotic properties of the system for different conditions, which correspond to the possible phenotypes that the model can account for. Next, we analyse the different trajectories leading to each phenotype in the wild type and mutant situations. As quantitative data are still largely lacking for this system, we use a qualitative logical formalism and its implementation in the GINsim software [38]. As we shall see, proper model analysis can assess where and when cell fate decisions are made, provide novel insight concerning the general structure of the network, in particular concerning the occurrence of cross-talks between pathways, and predict novel mutant phenotypes and component activity patterns.
The information gathered in the literature has been integrated into a regulatory graph (Figure 1A). Our selection of molecular players (nodes of the graph) is based on our current understanding of the molecular mechanisms of cell fate decision. Documented positive or negative effects among pairs of components are represented by signed arcs (Figure 1). Each node or arc is annotated and associated with bibliographical references in the model file, as well as in the accompanying documentation (cf. supplementary Text S1).
Our current model encompasses three main pathways (Figure 1A): the activation of a caspase-dependent apoptosis pathway, the RIP1-kinase-dependent pathway leading to necrosis, and the activation of the transcription factor NFκB with pro-survival effects. Other pathways involved in cell death, such as growth factor receptors or other RTK (receptor tyrosine kinase), TLR (Toll-like receptor), and MAPK signalling pathways have been provisionally left out.
We defined specific “markers” or “read-outs” of the three cell fates. When caspase-3 is activated, the cells are considered to be apoptotic; when MPT occurs and the level of ATP drops dramatically, the cells enter non-apoptotic cell death; finally, when NFκB is activated, we consider that cells survive. Active survival is thus monitored here by the activation of NFκB pathway, in accordance with many studies, as opposed to passive survival, which occurs when no death signals are engaged. In reality, other pathways can interact downstream of NFκB activation, which can reinforce or shut off survival. For now, passive survival will be referred to as the ‘naïve’ state, i.e. the stable states with none of the three pathways activated.
As mentioned before, the pathways are highly intertwined (Figure 1A). For instance, the survival pathway interacts with the apoptotic pathway at different points: cFLIP inhibits CASP8; BCL2 blocks mitochondria pore opening through inhibition of BAX (and BAK, implicitly represented in our model); and XIAP blocks the activity of both CASP9 in the apoptosome and CASP3. Conversely, the apoptotic pathway negatively regulates NFκB activity through the CASP8-mediated cleavage and inactivation of RIP1 upstream of NFκB. Because RIP1 operates upstream of the necrotic pathway, this regulation also impacts necrosis. Moreover, for the apoptosome to form, dATP (or/and ATP) is (/are) needed. Consequently, in our model, when necrosis occurs, ATP production drops, terminating apoptosis. Regarding the influence of the survival pathway on the necrotic one, NFκB tentatively stimulates the production of anti-oxidants that shuts off ROS level. Both the necrotic and the apoptotic pathways are able to interact with the survival pathway through the action of cIAP1/2, referred to as cIAP in our model. More precisely, cIAP1 and 2 are E3-ubiquitin ligases that target RIP1 for K63-linked polyubiquitination. They are essential intermediates in the activation of NFκB downstream of TNF receptor [27]. Some synthetic molecules that mimic the N-terminal of SMAC IAP-interacting motif have been shown to induce cIAP1/2 auto-ubiquitination and subsequent proteasomal degradation, thus blocking TNF-dependent NFκB activation [39],[40]. Tentatively, mitochondrial permeabilization in the apoptosis or necrosis pathways could block TNF-induced NFκB activation through the release of SMAC into the cytosol thereby causing the inhibition of c-IAP1/2. Initially, cIAP was not included in the model, which led to discrepancies between model simulations and published data. Indeed, in FADD or CASP8 deletion mutants, our preliminary model predicted only survival (not shown), whereas both necrotic and survival phenotypes were observed in experiments in the presence of TNF or FAS [41],[42],[43]. The consideration of the path MOMP⇒SMAC = |cIAP⇒NFκB enabled us to eliminate the discrepancies, both necrotic and survival phenotypes were then obtained in the simulations, although it does not preclude other mechanisms.
To transform the static map shown in Figure 1A into a dynamical model accounting for the different scenarios or set of events leading to one of the three phenotypes, we have to define proper dynamical rules. Since there is little reliable quantitative information on reaction kinetics and cellular conditions leading to one or another phenotype, these rules must be sufficiently flexible to cover all possible scenarios following death receptor activation.
The nodes encompassed in the map represent different things: simple biochemical components (receptors, ligands, proteins or metabolites): TNF, FASL, TNFR, FADD, FLIP, CASP8, RIP1, IKK, NFκB, cIAP, BCL2, BAX, Cyt_c, SMAC, ROS, XIAP, CASP3, ATP); specific modified forms of proteins: RIP1K (active RIP1 kinase), RIP1ub (K63-ubiquitinated RIP1); complexes of proteins: DISC-TNF (corresponding to TRADD, TRAF2, FADD, proCASP8), DISC-FAS (corresponding to FAS, FADD, proCASP8), apoptosome; cellular processes: MPT (Mitochondrial Permeability Transition) and MOMP (Mitochondria Outer Membrane Permeabilisation).
A Boolean variable is associated with each of these nodes, which can take only two logical values: “0” (false), denoting the absence or inactivity of the corresponding component, and “1” (true), denoting its active state.
Furthermore, a logical rule (or function) is assigned to each node, defining how the different inputs (incoming arrows) combine to control its level of activation. For example, CASP8 can be activated (its value is set to “1”) by DISC-TNF or DISC-FAS, but only in the absence of cFLIP protein. This can be encoded into a logical rule as follows: (DISC-TNF OR DISC-FAS) AND NOT cFLIP. Several nodes correspond to simple inputs (TNF, FASL and FADD). Their initial values are kept fixed during most simulations.
On the basis of the regulatory graph and the associated logical rules, we then proceeded with the exploration of the dynamical properties of our model. We first focused on the identification of all stable states and on their biological interpretation. Then, we investigated the reachability of these stable states for different initial conditions, for both wild type and mutant cases. Details on the computational methods used are provided in the Methods section and in the supplementary Text S1. The logical model has been filed in the BioModels database with the reference MODEL0912180000.
Analysis of the cell fate decision model (Figure 1A) led to the identification of the 27 stable states showed in Figure 2. These stable states are the sole attractors of the system under the asynchronous assumption (see Methods). They thus represent all possible cellular asymptotical states. In other words, whatever the initial conditions, a wild type cell will end up in one of these states if we wait long enough. A closer look reveals that several stable states correspond to each cellular fate, with few differing (minor) component values.
This consideration led us to address the following questions: (i) Does a cluster structure exist in the distribution of internal stable states of the network? (ii) If so, in these clusters, could the corresponding states be interpreted as slightly different realisations of the same cellular phenotype? (iii) What would be the characteristic signature of each cluster (conserved values of variables inside each cluster)? (iv) What is the number of independent variables defining the internal stable states of the network?
Standard statistical methods and clustering algorithms are applied to group stable states. Figure 3 displays a projection of the internal (without inputs and outputs) stable values into the 2D space defined by the first two principal components of the corresponding distribution. The first two principal components explain 52% and 20% of the total variation, respectively (Table S1). The first principal component can be associated with the activity of NFκB pathway, while the second is determined mainly by ATP and MPT status. These factors do appear to determine the principal (independent) degrees of freedom for the internal state of the network. A typical trajectory starting from any set of initial conditions will thus quickly converge to the region under the influence of these three components.
The 2-D graph (Figure 3) reveals a striking separation of the stable states into 4 clusters: one cluster (blue circles) devoid of significant activity, which we call the ‘naïve’ state; one cluster (green rhombs) corresponding to survival, with NFκB pathway activated; and two clusters corresponding to the two different modalities of cell death, apoptosis (orange squares) and necrosis (purple triangles). K-means clustering using Euclidean and L1 distance perfectly reproduces these groupings, demonstrating that the compact groups easily distinguishable on the PCA plot indeed represent well-separated clusters in the original multidimensional space.
Some interesting conclusions and predictions can be drawn just by looking at the values of each component in each phenotypical group. For instance, in the necrotic (purple) stable states, when FADD is present (i.e. normal wild type conditions), RIP1 is always OFF and CASP8 ON, even though RIP1 is required and CASP8 is dispensable for necrosis to occur. This observation suggests a transient activation of RIP1 protein when switching on the necrotic pathway in response to death receptors. However, inactivation or cleavage of RIP1 is not per se a prerequisite for necrosis, nor is CASP8 activation. Indeed, for the mutant models in which CASP8 activation is impaired, such as CASP8 or FADD deletion, there exist necrotic attractors with RIP1 = 1 (not shown). Our model thus predicts that TNF-induced necrosis could occur despite CASP8-mediated cleavage of RIP1. An attractive experimental model in which such a transient activation of RIP1 could be tested is the mouse fibrosarcoma cell line L929. Upon TNF exposure, these cells die by necrosis [44] and they have a functional CASP8 [45], which is cleaved during TNF-induced cell death [46].
Since RIP1 controls both the activation of NFκB and the level of ROS, the same transient behaviour could be expected for the survival phenotype. However, this is not observed with our model, as RIP1 = 1 in all survival (green) stable states. This can be explained by the regulatory circuit involving RIP1 and NFκB, which is not functional in necrosis. Indeed, when NFκB is active, it can mediate the synthesis of cFLIP, an inhibitor of CASP8, itself an inhibitor of RIP1. Moreover, RIP1 is part of the positive circuit that keeps NFκB ON. The model thus suggests that a sustained RIP1 activity is needed for survival. How could this hypothesis be experimentally assessed? If an experiment would reveal that RIP1 is only transiently activated upon death receptor activation, while NFκB remains activated, the model would be contradicted. In that case, one would need to look for other components capable to maintain NFκB active.
The stable state analysis described above provides a first validation of the master model presented in Figure 1A. On this basis, we performed a more detailed analysis of the dynamics of the system. We investigated which cell fates (stable states) can be reached from specific initial conditions. Given a set of reachable stable states, can we say something about their relative “attractivity”?
To avoid the combinatorial explosion of the number of states to consider, we have reduced the number of components while preserving the relevant dynamical properties of the master model (Figure 1B). Details of how this reduction is performed are provided in the Methods section. The resulting network encompasses 11 components. The corresponding Boolean rules are listed in Table 1. The size of the transition graph (211 = 2048) is now amenable to a detailed dynamical analysis. First, the set of attractors of this reduced model is identified: 13 attractors are obtained, which are all stable states matching those found for the master model when the input variable FADD = 1. Recall that, in the master model, both values of FADD were considered leading to 13 stable states with FADD = 1 and 14 stable states with FADD = 0. Using the theoretical results presented in [47] (mainly Theorem 1), we can conclude that the 13 stable states of Figure 2 are the only attractors of the master model when FADD = 1.
Based on the reduced model defined in Figure 1B and Table 1, we derived 15 model variants representing biologically plausible perturbations. We will abusively use the term “mutant” to refer to these variants, even though they do not all technically correspond to mutations. For instance, the “z-VAD mutant” simulates the effect of caspase inhibitor z-VAD-fmk. Each mutant simulation consists in a local alteration of our reduced model, which can be qualitatively compared with results reported in the literature.
In the Boolean framework, such alterations amount to force the level of certain variables to zero in the case of a gene deletion, or to one in the case of a component over-expression. As we are using the reduced version of the master model, some perturbed components may be hidden by the reduction process. In such cases, we change the logical rules of their (possibly indirect) targets to take into account their effects. Table 2 lists the 15 variants of the model considered, along with the modified logical rules, the expected effects on the phenotypes according to the literature, and short descriptions of simulation results.
The references provided in Table 2 cover experiments performed on different cell types and with different experimental conditions. In contrast, our cell fate model represents mechanisms of cell fate decision in a generic cell, qualitatively recapitulating a wide variety of cellular contexts. Given a cellular system, its response to the activation of death receptors is determined by the logical rules. However, the generic model presented here considers equally all possible contexts and regulatory combinations. To evaluate the relative likelihood of having a particular response in a randomly chosen cellular system, we count the relative number of possible trajectories from the stimulated ‘naïve’ state to a given phenotype. This analysis gives an idea on what is possible or forbidden in a ‘generic’ cell.
Using dedicated methods and software [48], the set of reachable stable states is calculated, starting from selected physiological initial conditions, for the wild type and mutant models. The physiological state is defined by fixing the variables ATP and cIAP to “1” and all the other ones to “0”. Different combinations of TNF and FASL are considered. The probability to reach each phenotype is computed as a fraction of the paths in the graph that link physiological initial conditions to each cell fate (Figure 4 for reduced model and Figure S4 for master model).
As expected, the absence of TNF and FASL can only lead to the ‘naïve’ state (except of course when caspase-8 or NFκB are over-expressed, for obvious reasons). This means that the inputs (TNF and FASL) are needed for the system to effectively trigger the decision process. This was expected since intracellular death signals are not yet taken into account in the model. When TNF = 1 (Figure 4, right panel), for the wild type system, we observe that three outcomes or phenotypes are reachable from the initial condition, with different probabilities: ∼10% for necrosis, ∼30% for active survival and ∼60% for apoptosis. Although these probabilities cannot be directly compared with experimental results, they become useful when comparing different variants of the model. For instance, an increase (or decrease) of a phenotype probability between the wild type and a particular mutant can be interpreted as a gain (or a loss) of effectiveness of the corresponding pathway in that mutant. Such qualitative observations can then be confronted with published experimental results, which are summarized in the last column of Table 2.
In most cases, activation of FASL and TNF lead to similar effects (not shown), except in the case of the FADD deletion mutant (Figure 5). As expected, this mutant cannot lead to cell death when FASL is ON. In contrast, necrosis is still possible in the presence of TNF. Interestingly, TNF-induced apoptosis is expected to be blocked [49] whereas the qualitative analysis shows that apoptosis is actually reachable in the model. Nevertheless, the probability of this phenotype is very low (around 0.61%), which means that very few trajectories may lead to apoptosis and it would thus be difficult to obtain the corresponding cellular context.
In the reachability analysis presented above, the value of TNF and FASL are kept constant and therefore always ON (or always OFF) along all trajectories. These qualitative simulations are useful to characterize the asymptotic behaviour of the system when the death receptor is engaged for a sufficiently long time. The principle of ‘ligand removal’ experiments consists in characterizing the decision process when it is subject to a temporary pulse of TNF. Here, time is intrinsically discrete, meaning that the duration of TNF pulse denoted td is represented by an integer number. In order to simulate each experiment, N trajectories were generated, starting from the “physiological” condition with TNF = 1. At time td, the value of TNF is forced to zero. The probabilities to reach the different phenotypes are then calculated as explained in the Methods section. The average probabilities, over the N computed trajectories, are represented in Figure 6, for the wild type and the 15 mutants.
The purpose of this study is to investigate the dynamics of all the mutants and how they reach the various possible phenotypes for different lengths of TNF pulses. It provides a measurable way to assess the appearance or disappearance of certain phenotypes upon TNF induction. The curves of Figure 6 allow to link explicitly the graphs of Figure 4 when TNF is ON (right panels) and OFF (left panels) with the subjacent dynamics.
Let us compare the wild type case and the deletion of cFLIP as an example of how to read these graphs. For early events, the two cases behave similarly as expected (up to event 3). As TNF pulse is prolonged, the apoptotic phenotype becomes more and more pronounced and strongly favoured over the survival one in the cFLIP mutant as opposed to the wild type conditions. This leads to the complete disappearance of survival in the mutant. This observation reinforces the role of cFLIP in the control of the apoptotic pathway.
With the ‘ligand removal’ experiment, we can evaluate the number of steps, in the reduced model, that are needed for the cell to decide on its fate after TNF exposure. For almost all mutants and wild type case, the choice is made around step 4. This means that, after this point, even if TNF is removed, the cell has already committed to a specific fate.
One surprise arises from the non-monotonic behaviour of mutants for which apoptosis is suppressed (APAF1, BAX, caspase-8 and FADD deletions and z-VAD-fmk treatment), tentatively indicating a competition between components of the survival and necrotic pathways. Indeed several inhibitory cross-talks could explain this behaviour. These mutants also indicate the existence of an optimal TNF induction for which the maximum rate of necrosis is achieved (around step 2 in the corresponding mutants of Figure 6).
To complete our study of cell fate decision, we reasoned on the simplest model of cell fate that can be deduced from the master model described above.
The purpose here was to further simplify the network to obtain a formal representation of the logical core of the network. We have selected three components to represent the three cellular fates: NFκB for survival, MPT for necrosis and CASP3 for apoptosis. Based on reduction techniques and on the identification of all possible directed paths between these three components [50], a three-node diagram was deduced from the master model.
In this compact model, each original path (including regulatory circuit) is represented by an arc whose sign denotes the influence of the source node on its target. All original paths and the corresponding arcs are recorded in Table 3. In some ambiguous cases (e.g. influence of MPT on CASP3 or of NFκB on MPT), the decision on the sign of the influence is based on the Boolean rules and not on the paths only. Indeed, two negative and one positive paths link NFκB to MPT. Therefore, the sign of the arc depends not only on the states of BCL2 and of ROS, both feeding onto MPT, but also on the rule controlling MPT value. Since the absence of BCL2 and the presence of ROS (Boolean ‘AND’ gate) participate in the activation of MPT, if BCL2 is active, then MPT is set to 0, even when ROS is ON. By extension, if NFκB is ON, then MPT is 0, justifying the choice for a negative influence. In the case of mutations eliminating all the negative influences, however, a positive arrow must be considered.
The resulting molecular network is symmetrical: each node is self-activating and is negatively regulated by the other nodes (Figure 7, upper left panel). This is a conceptual picture representing the general architecture of the master model that can help address specific questions. Even for this relatively simple regulatory graph, there is a finite but quite high number of possible logical rules. For now, we use a simple generic rule involving the AND and NOT operators. For example, the logical rule for CASP3 is: NOT MPT AND NOT NFkB AND CASP3. This compact model has four stable states, each corresponding to one cell fate, along with the ‘naïve’ state (Figure 7, upper right panel). This is coherent with what was observed from the analysis of the complete model.
To validate our compact model, we verified that the simulations of known mutations correspond to the published observations. Here, when a hidden component is deleted, all the paths traversing this component in the original graph are broken. If all the paths corresponding to an arc of the compact model happen to be broken, then it is removed. In the case of auto-regulation, not only the link is broken but the node is also set to zero to avoid the node to become active in the absence of death receptor activation.
Let us consider the CASP8 deletion mutant to illustrate this approach (Figure 7, middle panels). For this mutant, several arrows in the compact model have to be deleted. For example, the arcs CASP3⇒CASP3 (paths 4+5 in Table 3) and CASP3⊣MPT (path 17) clearly depend on the activation of CASP8. Note, however, that CASP8 intervenes in other paths, which do not fully rely on its sole activity. In the case of the arc NFκB⇒NFκB, CASP8 depletion interrupts path 3, while path 2 can still enable the NFκB auto-regulation. Consistent with the results from the previous section, CASP8 depletion leads to the loss of the apoptotic fate while the ‘naïve’ stable state cannot be attained.
At this point, one could wonder how apoptosis could be re-established in a CASP8 mutant. The analysis of the broken paths suggests some experiments to bypass CASP8 and undergo CASP3 activation. On the basis of path 4, BAX, MOMP, SMAC and XIAP are identified as potential targets, while path 5 points to cytochrome c and apoptosome. One way to experimentally assess this possibility would be to inject exogenous cytochrome c as it was done in ‘wild type’ conditions [51], or yet provoke its release from the mitochondria by forcing the opening of the pores. This is possible only in the absence (or with low activity) of NFκB and in the presence of ATP. Again, since no quantitative information can be deduced from the path analysis proposed in this study, no prediction can be made on the concentrations of proteins needed to achieve a specific answer.
In a previous section, we postulated that an inhibition of the survival pathway by the necrotic pathway is necessary to reproduce some mutant phenotype. We suggested that cIAP could play this role. Let us now test this hypothesis with our conceptual model. We build the corresponding 3-node model without cIAP. In the current version, cIAP plays two important roles, first as a mandatory intermediate in the inhibitory effect of MPT (associated with necrosis) onto NFκB (survival) (path 13), next as an obligate intermediate in the self-activation of NFκB (path 2). The simulation (Figure 7, lower right panel) shows that in the absence of cIAP, it is impossible to obtain the necrosis cell fate in the CASP8 (and FADD) mutant(s), in agreement with our previous conclusions and in support of our suggestion. A complete list of all possible gene knockouts is provided in the Table S2.
This conceptual model analysis underlines the importance to simplify in order to better understand the general structure of the network and reason on it. Indeed, the simple 3-node network enables us to grasp global functional aspects and propose specific qualitative predictions.
Mathematical models provide a way to test biological hypotheses in silico. They recapitulate consistent heterogeneous published results and assemble disseminated information into a coherent picture using a coherent mathematical formalism (discrete, continuous, stochastic, hybrid, etc.), depending on the questions and the available data. Then, modelling consists of constantly challenging the obtained model with available published data or experimental results (mutants or drug treatments). After several refinement rounds, a model becomes particularly useful when it can provide counter-intuitive insights or suggest novel promising experiments.
Here, we have conceived a mathematical model of cell fate decision, based on a logical formalisation of well-characterised molecular interactions. Former mathematical models only considered two cellular fates, apoptosis and cell survival. In contrast, we include a non-apoptotic modality of cell death, mainly necrosis, involving RIP1, ROS and mitochondria functions.
Both the master and the reduced models were constructed on the basis of an extensive analysis of the literature. The master model (Figure 1A) summarises our current understanding of the mechanisms regulating cell fate decision and identifies the major switches in this decision. However, some important interactions, components (caspase-2, calpains, AIF, etc.) or pathways (JNK, Akt, etc.) have not yet been considered. This model was built to be as generic as possible. Most of the mutants considered were analysed in Jurkat cells, T-cells, or L929 murine fibrosarcoma cells, thus in very different cellular contexts (e.g. in response to TNF, Jurkat cells are resistant to cell death, whereas L929 cell lines undergo necrosis). We are trying to account for all those phenotypes in a unique model. The next step will be to provide a model variant for each cell type in order to better match cell-specific behaviours.
The reduced models can be used to simulate observed experiments and to reflect on the general mechanisms involved in apoptosis, survival or necrosis. This led us to identify the principal actors involved in the decision process. The presence of RIP1 or FADD, for example, proved to be decisive in our simulation. However, the role of cFLIP appears less obvious than previously suggested [7].
We can easily perturb the structure of the system in silico and assess the dynamical effects of such perturbations (e.g. novel knockouts). Our model can also be used to decide between antagonist results found in different publications. For instance, the inhibitory role of cIAP1/2 on the apoptotic pathway was initially attributed to a direct inhibition of caspases. However, detailed biochemical studies challenged this view [52],[53]. We have tested this hypothesis by adding an inhibitory arc from cIAP onto CASP8, but simulations do not support a functional inhibitory role of cIAP1/2, since survival is favoured over apoptosis in many mutants, thus making apoptosis a very improbable phenotype (Figure S1). Similarly, we tested the role of the feedback circuit involving CASP8 and CASP3. We found that the activation of CASP8 by CASP3 is not functional when TNF and FASL are constantly ON. However, when TNF or FAS signal is not sustained, CASP3⇒CASP8 activation becomes necessary to insure the persistence of the apoptotic phenotype. When TNF is sustained, this feedback is no longer needed (see Figures S2 and S3 for details).
The in-depth analysis of model properties led us to propose several predictions or novel insights. Some concern the structure of the network, as several interactions appear to be necessary to achieve specific phenotypes. For example, our simulations of FADD and CASP8 deletion mutants underline the need for a mechanism from the necrotic pathway that would inhibit the survival one. Here, we consider a mechanism involving MPT, SMAC and cIAP. Other simulations point to different roles of proteins: RIP1 activity is transient in necrosis whereas it is sustained in survival. Similarly, our model analysis shows the role played by the duration of the TNF pulses in the cell fate decision and enlights when this decision is made. Finally, some hints about possible scenarios for forcing or restoring a phenotype in mutants are provided.
Deregulations of the signalling pathways studied here can lead to drastic and serious consequences. Hanahan and Weinberg proposed that escape of apoptosis, together with other alterations of cellular physiology, represents a necessary event in cancer promotion and progression [54]. As a result, somatic mutations leading to impaired apoptosis are expected to be associated with cancer. In the cell fate model presented here, most nodes can be classified as pro-apoptotic or anti-apoptotic according to the results of “mutant” model simulations, which are correlated with experimental results found in the literature. Genes classified as pro-apoptotic in our model include caspases-8 and -3, APAF1 as part of the apoptosome complex, cytochrome c (Cyt_c), BAX, and SMAC. Anti-apoptotic genes encompass BCL2, cIAP1/2, XIAP, cFLIP, and different genes involved in the NFκB pathway, including NFKB1, RELA, IKBKG and IKBKB (not explicit in the model). Genetic alterations leading to loss of activity of pro-apoptotic genes or to increased activity of anti-apoptotic genes have been associated with various cancers. Thus, we can cross-list the alterations of these genes deduced from the model with what is reported in the literature and verify their role and implications in cancer.
For instance, concerning pro-apoptotic genes, frameshift mutations in the ORF of the BAX gene are reported in >50% of colorectal tumours of the micro-satellite mutator phenotype [55]. Expression of CASP8 is reduced in ∼24% of tumours from patients with Ewing's sarcoma [56]. Caspase-8 was suggested in several studies to function as a tumour suppressor in neuroblastomas [57] and in lung cancer [58].
On the other hand, constitutive activation of anti-apoptotic genes is often observed in cancer cells. The most striking example is the over-expression of the BCL2 oncogene in almost all follicular lymphomas, which can result from a t(14;18) translocation that positions BCL2 in close proximity to enhancer elements of the immunoglobulin heavy-chain locus [59]. As for the survival pathway, elevated NFκB activity, resulting from different genetic alterations or expression of the v-rel viral NFκB isoform, is detected in multiple cancers, including lymphomas and breast cancers [60]. An amplification of the genomic region 11q22 that spans over the cIAP1 and cIAP2 genes is associated with lung cancers [61], cervical cancer resistance to radiotherapy [62], and oesophageal squamous cell carcinomas [63].
A better understanding of the pro- or anti-apoptotic roles of these genes involved in various cancers and their interactions with other pathways would set a ground for re-establishing a lost death phenotype and identifying druggable targets. The cell fate model proposed here is a first step in this direction.
In the future, we will consider additional signalling cascades and their cross-talks, following the path open by other groups [64]. In parallel, we are contemplating the inclusion of other modalities of cell death such as autophagy [65], which inhibits apoptosis through BCL2 and is itself inhibited by apoptosis through Beclin1. The functioning of the intrinsic apoptotic pathway and the internal cellular mechanisms capable of triggering it could be investigated in more details, taking advantage of recent molecular analyses [66],[67]. Finally, when systematic quantitative data regarding the decision between multiple cell fates will become available, our qualitative model could be used to design more quantitative models adapted to specific cellular systems in order to predict the probability for a given cell to enter into a particular cell fate depending on stimuli.
The computation of trajectories in the state space consists in the calculation of sequences of states where each member of the sequence is a logical successor of the previous one. As we choose to use Boolean variables to encode the 25-dimensional master model, the state space is the set S = {0,1}25. Although finite, the size of this set is huge (more than 33 millions states). Furthermore, in the discrete framework, the mathematical definition of the trajectories assumes an updating rule for the variables. Two main strategies are usually considered to analyse discrete models of biological networks. The first one consists in updating all variables simultaneously, at each time step. This synchronous strategy [68] has the advantage to generate simple determinist dynamics, each state having one and only one successor. Drawing a directed arrow from each of the 225 states to its successor, one constructs the synchronous transition graph, comprising all synchronous trajectories of the system. The determinism of the synchronous transition graph is a very strong property that poorly portrays the complexity of the biochemical processes that are modelled (some processes are likely to occur faster than others). The second strategy, which is used in this paper, consists in considering that only one component is updated at each time, implying that a state may have several successors [69].
More precisely, to compute the set of asynchronous successors of a state x = (x1,…,xn)∈{0,1}n, one has to follow the three steps: (1) compute the state F(x) = (f1(x),…, fn(x)), where fi is the Boolean rule of the ith variable (F(x) is thus the synchronous successor of x); (2) select the indices i such that xi≠fi(x) (those are the indices of the variables that are liable to change when the system is in state x); and (3) for all such indices i, the state (x1,…,fi(x),…,xn) is an asynchronous successor of x.
According to this definition, in the asynchronous approach, no a priori hypothesis is made on the order of the events: all possible orders are considered, which is much more satisfying from a modelling point of view, as it is very difficult to know the relative speeds of the different processes involved in the master model. Note that the stable states of the model are independent on the choice of the strategy (synchronous or asynchronous). Therefore, the first analysis (based on the clustering of stable states) is valid regardless the updating strategy.
Drawing an arrow from each state to its asynchronous successors leads to the construction of an asynchronous transition graph, which comprises all possible asynchronous trajectories of the system. To each arrow starting from the same state is associated an equal probability (see [70] for details). This is a strong assumption, which is the main reason why the exact values of computed probabilities (of the different phenotypes) should not be compared to experimental data in a quantitative manner. Nevertheless, the same assumption has been made for all model variants (mutants and drug treatments), thereby allowing comparative studies. A systematic method to assess the impact of the probability distribution is a key point towards a finer quantitative analysis (work in progress).
As pointed earlier, the size of the transition graph is exponential with respect to the number of variables, which constitutes a first obstacle to the dynamical analysis. A second difficulty resides in the fact that the asynchronous graph is not deterministic, as each vertex may have more than one successor, which, given the size of the graph, makes the application of classical graph algorithms computationally heavy.
We have used a model reduction technique specifically adapted to discrete systems, which mainly consists in iteratively “hiding” some variables, while keeping track of underlying regulatory processes [47]. The main dynamical properties of the master model, including stable states and other attractors are conserved in the reduced model. Thanks to the computation of the reduced asynchronous transition graph, relevant qualitative dynamical properties of the model can be compared to experimental results for wild type and in different mutant cases.
To reduce the number of species in the master model, each logical rule is considered. For each removed component, the information contained in its rule is included in the rules of its targets such that no effective regulation is lost.
Many intermediate components could easily be replaced by a proper rewriting of the logical rules associated with their target nodes. For example, IKK has only one input (RIP1ub) and one output (NFκB). Since its role in our model merely consists in transmitting the signal from RIP1ub to NFκB, it can be easily replaced by a straightforward change in the logical rule associated with NFκB (implementing a direct activation from RIP1ub instead of IKK). We also relied on the results of the clustering of stable states and their associations with biologically plausible phenotypes to select the key components to keep in the reduced model: NFκB is the principal survival actor, while caspases-3 and -8, together with the mitochondrial membrane permeability variables (MOMP and MPT), determine apoptotic and non-apoptotic cell deaths.
Let us consider the example of the removal of BAX and BCL2 (Figure 1 A and B). The regulators (or inputs) of these variables are NFκB for BCL2 and CASP8 for BAX while their regulating targets (or outputs) are MPT for BCL2 and MOMP for BAX. BCL2 is directly activated by NFκB, and has two targets: MPT and BAX. Therefore, BCL2 removal is performed by replacing BCL2 by NFκB into the rules of the two targets, leading to the two new logical rules: MPT′ = ROS AND NOT NFkB and BAX′ = C8 AND NOT NFkB. Applying the same process to remove BAX, one obtains the following new rule for MOMP: MOMP′ = MPT OR (C8 AND NOT NFkB).
The variables MOMP and MPT have now as inputs the variables NFκB and CASP8. One can see that, in spite of the disappearance of variables BAX and BCL2, their regulating roles are still indirectly coded in the reduced system, ensuring that no “logical interaction” of the master model (i.e. activation or inhibition) is actually lost during the reduction process. Table S3 lists the variables of the master model that are removed to obtain the reduced model.
Some hypotheses were made when reducing the model. First, FADD is considered to be constantly ON in wild type simulations. Second, since the two complexes TNFR and DISC-TNF have been removed together with the input FADD, the two deaths ligands TNF and L have the exact same action in the reduced model. Indeed, we consider that, in response to FAS death receptor engagement as well as that of TNF; the activations of both the survival and necrotic pathways RIP1-dependent. In this case, one could then merge these variables and consider only one input that could be called “external death receptor”. However, we choose to keep the two variables TNF and FASL, in the FADD deletion mutant, the phenotype differs for TNF and FAS signal: actually, only for that mutant is the symmetry of TNF and FAS broken.
|
10.1371/journal.pntd.0005400 | Ultrasonographic evaluation of urinary tract morbidity in school-aged and preschool-aged children infected with Schistosoma haematobium and its evolution after praziquantel treatment: A randomized controlled trial | Schistosoma haematobium infections are responsible for significant urinary tract (UT) complications. Schistosomiasis control programs aim to reduce morbidity, yet the extent of morbidity in preschool-aged children and the impact of treatment on morbidity reduction are not well studied.
Our study was embedded in a randomized, placebo-controlled, single-blind trial in Côte d’Ivoire, which evaluated the efficacy and safety of three doses (20, 40 and 60 mg/kg) of praziquantel in school-aged (SAC) and preschool-aged (PSAC) children infected with S. haematobium. Enrolled children were invited to participate in an ultrasound examination prior and six months after treatment. At these time points 3 urine samples were collected for parasitological and clinical examinations.
162 PSAC and 141 SAC participated in the ultrasound examination at baseline, of which 128 PSAC and 122 SAC were present at follow-up. At baseline 43% (70/162) of PSAC had UT morbidity, mostly at bladder level and 7% had hydronephrosis. 67% (94/141) of SAC revealed mainly moderate UT pathology, 4% presented pseudopolyps on the bladder wall, and 6% had pyelectasis. At follow up, 45% of PSAC and 58% of SAC were S. haematobium positive, mostly harboring light infection intensities (41% and 51%, respectively). Microhematuria was present in 33% of PSAC and 42% of SAC and leukocyturia in 53% and 40% of PSAC and SAC, respectively. 50% (64/128) of PSAC and 58% (71/122) of SAC presented urinary tract morbidity, which was mainly mild. A significant correlation (p<0.05) was observed between praziquantel treatment and reversal of S. haematobium induced morbidity. Progression of UT pathology decreased with increasing praziquantel dosages. A worsening of morbidity was observed among children in the placebo group.
Bladder morbidity is widespread among PSAC. Praziquantel treatment is significantly associated with the reversal of S. haematobium induced morbidity, which underscores the importance of preventive chemotherapy programs. These programs should be expanded to PSAC to prevent or decrease the prevalence of morbidity in young children. This trial is registered as an International Standard Randomized Controlled Trial, number ISRCTN15280205.
| Schistosoma haematobium is a parasite that infects the human genito-urinary tract. People get infected with the parasite through contact with fresh water and children are at major risk. The complications linked to this infection are due to an inflammation caused by accumulation of the eggs in peri-bladder veins. If not treated, infections can last years and different degrees of severity are observed. These range from thickening of the bladder wall and blurriness of the mucosa to more serious lesions such as pseudo polyps and masses in the bladder that can, with time, evolve in cancer of the bladder. We analyzed preschool-aged children (PSAC) and school-aged children (SAC) with ultrasound before and after praziquantel treatment. Children were randomly assigned to different doses of praziquantel (20, 40 or 60 mg/kg) or to placebo at baseline. Six months after treatment all children underwent another ultrasound of the urinary tract. We included 162 PSAC and 141 SAC at baseline, of which 128 PSAC and 122 SAC had a second ultrasound evaluation six months afterwards. In addition, urine was sampled at both time points for presence of blood, proteins and signs of infection (leukocytes and nitrates). Six months post-treatment 45% of PSAC and 58% of SAC were S. haematobium positive. Already at the first screening 43% of PSAC and 67% of SAC had bladder lesions. After treatment 50% of PSAC and 58% of SAC still had pathology linked to the infection. We found a correlation between the treatment dose and healing of bladder lesions. On the other hand, we experienced an aggravation of lesions in the placebo group. Praziquantel is given to SAC as preventive chemotherapy every year at national level, where this parasite is endemic. This program should be expanded and include PSAC as well in order to reduce the consequences of infection.
| Schistosomiasis primarily caused by Schistosoma haematobium, S. japonicum and S. mansoni is a significant public health problem in low-income tropical and subtropical countries. It is an ancient disease with first reports on schistosomiasis dating back 4000 years ago [1]. Yet, still today an estimated 230 million people are infected [2]. Adult S. haematobium settle in the venous plexus of the genito-urinary tract of the infected host and produce fertilized eggs. Evidence suggests that morbidity is caused by the trapping of eggs within the urinary and genital tract, which induce a granulomatous host immune response. The granuloma formation induces a chronic inflammation resulting in disease manifestations. In more detail, morbidity includes a wide range of pathological presentations, from thickening of the bladder wall mucosa, ureteral dilatation and hydronephrosis, to presence of polyps and masses in the lumen, which could lead to bladder carcinoma in more severe cases [3,4].
S. haematobium infection is commonly detected by microscopic examination for eggs via urine filtration. Macro and microhematuria and proteinuria are indirect signs of infection, especially in school -aged (SAC) and preschool -aged (PSAC) children [5,6]. In addition, ultrasound examination of the urinary tract (UT) of infected subjects is an important tool to provide information on bladder and kidney lesions [7]. While intensity of infection as well as hematuria are important indirect indicators of morbidity [8,9], UT lesions could be quite different even at similar intensity of infection. Moreover this technique is useful not only at individual level, but also at community level, since it is well accepted, non-invasive and simple to perform [10,11]. Therefore, ultrasonography has been widely used to evaluate morbidity of UT due to S. haematobium infection [12–15] as well as its resolution after treatment [7,16–18]. It has been shown that UT lesions improve 12 months after treatment and, if not re-treated in case of reinfection, reappear 18 months after treatment [16]. It might be worth highlighting that studies in PSAC have been rare. To date, only few studies have included young children [4,15], which in general show a higher prevalence of morbidities in older children and adolescents. However, given that efforts are ongoing to include PSAC in preventive chemotherapy programs, it is crucial to have more data on the morbidity of PSAC and the impact of praziquantel in the prevention and reversal of morbidity at different follow up times, with the ultimate goal to define suitable control strategies. In addition, the optimal praziquantel dose in PSAC remains to be elucidated and findings on the reversal of morbidity might aid in the selection of optimal treatment dosages.
The aim of our study was therefore to evaluate morbidity in PSAC and SAC infected with S. haematobium and its resolution 6 months after treatment with different doses of praziquantel compared to placebo.
Ethical approval for the study was obtained by the National Ethics Committee of the Ministry of Health in Côte d’Ivoire (CNER, reference no. 037/MSLS/CNER-dkn) and the Ethical Committee of Northwestern and Central Switzerland (EKNZ; reference no. 162/2014). Parents/ guardians of enrolled children were informed about the trial, and written informed consent as well as signed assent was obtained before the first child was enrolled. This trial is registered as an International Standard Randomised Controlled Trial, number ISRCTN15280205. All children were treated with praziquantel at the end of the trial according to local guidelines (40 mg/kg).
Our study was embedded in a randomized, parallel-group, single-blind, placebo-controlled, dose ranging trial in PSAC (aged 2–5 years) and SAC (6–15 years) infected with S. haematobium. In both cohorts, 40 children per arm were randomized, using block randomization to 20, 40, 60 mg/kg praziquantel or placebo. The ultrasound evaluation was carried out in November 2015 and May 2016, in four different villages (Mopé, Diasson, Nyan, Massandji and Djiougbosso) in the Adzopè region of Côte d’Ivoire.
Details on the study procedures will be presented elsewhere. Briefly, all children provided three samples of urine on three different days at baseline, 21 days after treatment (follow up; not reported here) and six months after treatment. Urines were examined with the filtration method for detection of S. haematobium eggs according to standard procedures [19]. In addition, chemical examination of urines was performed using Multistix 10 SG Reagent Strips (Siemens Healthcare, Zurich Switzerland). From each child one stool sample was collected at baseline and 21 days post-treatment for the evaluation of co-infections with S. mansoni and soil-transmitted helminths. On the day of treatment all children provided one drop of blood for Plasmodium spp detection with rapid test (RDT) and hemoglobin measurement.
Before treatment all children underwent a physical examination performed by a physician and body temperature, blood pressure and pulse height and weight were recorded. Signs and symptoms of malaise were assessed with a questionnaire. S. haematobium egg-positive children fulfilling all inclusion criteria were assigned to one of the four following treatment arms: praziquantel 20 mg/kg (group 1), 40 mg/kg (group 2), 60 mg/kg (group 3) or placebo (group 4). Ultrasound was performed by a trained physician with Sonosite 180 Plus, probe Convex 3.5 mHz ultrasonography machine on the day of treatment. Children were asked to drink at least two full glasses of water before undergoing UT ultrasound. Ultrasound was performed if the bladder was at least 100 cc full and the ureter was considered dilated if its diameter measured >7 mm.
21 days and 6 months post-treatment all treated children were asked to provide three urine samples for detection of S. haematobium eggs and chemical examination. At the second follow up another sonography of urinary tract was performed from the same operator as at baseline.
Results were double entered in a database (Excel 2010), cross-checked and analyzed with Stata 12.0 (Lakeway Drive College station, TX, Unites States of America). The intensity of infection for S. haematobium was assessed by calculating the average of the egg counts from the triplicate urine filtration. Infection intensity was classified following WHO cutoffs [20].
Chi-squared analyses were performed to determine the associations between different markers of morbidity by sex, age, intensity of infection or markers of UT infections.
In November 2015 303 of the 348 children enrolled in the randomized controlled trial underwent an evaluation of the UT with ultrasound (Fig 1). Demographic, clinical and parasitological baseline data are presented in Table 1. Briefly, 162 of the 303 children were PSAC with a mean age of 3.8 (2–5) years. 46% of the preschoolers were male. 141 participants were SAC. Their average age was 8.9 (6–15) years and 44% were male. Six months after treatment (May 2016) 250 children (128 PSAC and 122 SAC) had an ultrasonography done for evaluation of UT lesions.
To our knowledge this is the first study that analyses urinary tract morbidity in school-aged and preschool-aged children affected by S. haematobium at baseline and six months after treatment with different praziquantel dosages and placebo. In settings where control of morbidity is the main goal of public health interventions, the most widely used criteria to determine it is the measurement of egg counts and urine analyses for hematuria and proteinuria, as indirect signs of UT impairment [3,12]. However, obviously a more accurate and specific evaluation of the organ pathology should be the way to follow [12,21–22]. Ultrasound examination allows to assess the damage of bladder wall and genito-urinary tract, which in combination with parasitological results and urine analyses are good indicators of consequences of chronic infection [4,12,13]. Ultrasonography has been applied since the ‘70s [21] for schistosomiasis to detect and describe the morphology of lesions. The need to implement diagnostic and monitoring with ultrasound is widely shared [4,10,21], but so far its use is still limited [21].
Since in schistosomiasis UT morbidity often occurs asymptomatic until an advanced grade of pathology [23,24], ultrasound offers the great advantage to spot early complications and progression of pathology in a non-invasive and easy to perform manner.
Our study confirms that early complications and bladder consequences of a S. haematobium infection are frequent also in preschool-aged children [5] (Fig 2). We recorded both direct and indirect signs of infection that give a full and detailed picture of UT status in infected children of different ages. In more detail, in our study most children (79%) had low intensity infection but nonetheless of these 54% of children (43% of PSAC and 67% of SAC) presented UT morbidities. As Hatz and colleagues pointed out [4], lesions of the bladder are observed also in absence of excretion of eggs, as these might be stuck and trapped in the wall resulting in an inflammatory reaction, that does not allow their release. Also for other helminthic infections, it has been demonstrated that morbidity (such as anemia, stunting) is mostly triggered by chronicity of infection rather than by its intensity [25,26].
According to our findings, children are not affected by severe morbidity, in fact, the greater part of hydronephrosis resolved immediately after urination. We also did not observe a frequent presence of pseudopolyps or masses in the bladder (6%). Our data are in line with findings by Koukounari and Njaanake [10,27], but in contrast to Elmadani, who described that more than 40% of children had masses in the bladder lumen and 30% had hydronephrosis after urination [13]. The prevalence of UT morbidity is indeed very different from one study to the other. For example, Heutier and colleagues registered a 70% prevalence of bladder lesions in an African village endemic for S. haematobium in children [28], whereas Ekwunife and Koukounari reported a lower rate of UT morbidity in infected children (38% and 6% respectively) similar to what we have found [10,12].
As already reported and underlined in several trials on Schistosoma morbidity [4,18,29–31], praziquantel treatment is crucial in decreasing morbidity with regard to healing lesions and pathology linked to the infection, especially at early stages of the disease. In the present study we went a step further and studied the effect of different praziquantel doses and placebo on UT morbidity. Strikingly, while in the placebo group almost 40% of children had progression of UT pathology over the 6 months course, this rate decreased with increasing dosages being only 5% in the children treated with 60 mg/kg praziquantel. In addition, all dosages of the drug were correlated with an improvement of the clinical picture. Overall, more than 90% of treated children experienced improvement of lesions, whereas in the placebo group this rate was only 10%.
In our study 74% of children had no residual urine after bladder void and 12% had a residual volume greater than 50%. We did not perform an uroflowrimetry to confirm pathological voiding, but children were asked about symptoms linked to urination discomfort and reported urge of voiding even if the bladder was almost empty and a feeling of incomplete voiding was present. Voiding impairment is difficult to assess and confirm, especially in children and in conditions of stress such as ultrasound performance and clinical examination. Nonetheless, we observed an improvement at follow up both in bladder filling and discomfort in urination, though this was not properly validated. Akpata stated that the above mentioned symptoms are better indicators of schistosomiasis than residual volume calculation [21]. Stiffness of detrusor muscle, polyps and hydronephrosis are signs of severe stage of the pathology, which fortunately was rare in our study cohort. This suggests that the annual drug administration that takes place in the area is a good strategy to fight morbidity and decrease UT impairment [4,27].
Urinary tract infections (UTI) are common in childhood, accounting for 6% of infection in this age range [32]. In our study leucocyturia was often documented (57%), especially among SAC (72% versus 48% in PSAC), whereas nitrituria was more frequent among PSAC (15 vs 3%) [33,34]. According to our data, lower urinary tract morbidity was correlated with a general worsening of the UT, revealing a higher rate of nitrates and proteins and blood cells in urines, which is an evident sign of mucosa damage. As in previous trials, we also found hematuria and proteinuria to be good indicators of UT pathology (66% and 56% respectively) [10,15]. The prevalence of UTI in our study was higher than earlier findings for same age group [32], but this is not surprising given the fact that almost half of the children had chronic infections with S. haematobium. Chronic infections with S. haematobium are the main cause of mucosa damage and hence more likely to develop subsequent bacterial infections [35]. After treatment we observed an improvement of microhematuria in treated children compared to placebo treated children in both age groups.
On the other hand, clinical symptoms documented, such as fever and cough, were not found to be related to an S. haematobium infection, but are likely triggered by other diseases. For instance, at physical examination splenomegaly and hepatomegaly was observed in 30% of children, which might be correlated to S. mansoni or to other co-infections (e.g. malaria, leishmaniasis).
In conclusion, we have demonstrated that extending treatment (40 or 60 mg/kg praziquantel) from school-aged to preschool-aged children is crucial, in order to prevent morbidity to a S. haematobium infection. We have shown that already a high percentage of PSAC present bladder inflammation and mucosa thickening due to S. haematobium infection, which could be decreased by including this age group in treatment programs. We observed a high re-infection rate with S. haematobium, therefore preventive chemotherapy must be conducted at least once a year in PSAC and SAC in order to decrease morbidity.
|
10.1371/journal.pntd.0005972 | Next-generation ELISA diagnostic assay for Chagas Disease based on the combination of short peptidic epitopes | Chagas Disease, caused by the protozoan Trypanosoma cruzi, is a major health and economic problem in Latin America for which no vaccine or appropriate drugs for large-scale public health interventions are yet available. Accurate diagnosis is essential for the early identification and follow up of vector-borne cases and to prevent transmission of the disease by way of blood transfusions and organ transplantation. Diagnosis is routinely performed using serological methods, some of which require the production of parasite lysates, parasite antigenic fractions or purified recombinant antigens. Although available serological tests give satisfactory results, the production of reliable reagents remains laborious and expensive. Short peptides spanning linear B-cell epitopes have proven ideal serodiagnostic reagents in a wide range of diseases. Recently, we have conducted a large-scale screening of T. cruzi linear B-cell epitopes using high-density peptide chips, leading to the identification of several hundred novel sequence signatures associated to chronic Chagas Disease. Here, we performed a serological assessment of 27 selected epitopes and of their use in a novel multipeptide-based diagnostic method. A combination of 7 of these peptides were finally evaluated in ELISA format against a panel of 199 sera samples (Chagas-positive and negative, including sera from Leishmaniasis-positive subjects). The multipeptide formulation displayed a high diagnostic performance, with a sensitivity of 96.3% and a specificity of 99.15%. Therefore, the use of synthetic peptides as diagnostic tools are an attractive alternative in Chagas’ disease diagnosis.
| Chagas disease, caused by the parasite Trypanosoma cruzi, is a life-long and debilitating illness of major significance throughout Latin America, and an emergent threat to global public health. Diagnostic tests are key tools to support disease surveillance, and to ultimately help stop transmission of the parasite. However currently available diagnostic methods have several limitations. Identification of novel biomarkers with improved diagnostic characteristics is a main priority. Recently, we conducted a large-scale screening looking for new T. cruzi antigens using short peptides displayed on a solid support at high-density. This led to the identification of several hundred novel antigenic epitopes. In this work we validated the serodiagnostic performance of 27 of these against an extended panel of human serum samples. Based on this analysis, we developed a proof-of-principle multiplex diagnostic kit by combining different validated reactive peptides. Overall, our data support the applicability of high-density peptide microarrays for the rapid identification and mapping epitopes that could be readily translated into novel and useful tools for diagnosis of Chagas disease.
| Chagas disease is a major health and economic problem in Latin America, for which no vaccine or appropriate drugs for large-scale public health interventions are yet available [1]. It is caused by the protozoan parasite Trypanosoma cruzi, found throughout the Americas in a variety of wild and domestic mammalian reservoirs, and it is usually transmitted by infected blood-sucking triatomine bugs. It is estimated that ~5.7 million people are currently infected with T. cruzi and that up to 120 million individuals living in endemic areas in Latin America are at risk of infection [2]. Chagas Disease remains the most important parasitic disease in the Western Hemisphere, with an estimated disease burden, as measured by disability-adjusted life-years, that is 7.5 times as great as that of malaria [2]. Increasing travel and immigration have also brought the risk of T. cruzi infection into non endemic countries [3]. Several efforts have successfully been undertaken to control transmission in Latin America, with a concomitant decrease in the number of acute vector-borne infections [4]. However, humans can also become infected with T. cruzi through the ingestion of tainted food and fluids, receipt of contaminated blood transfusion or organ transplantation, laboratory accidents, and from mother-to-child during pregnancy/delivery [1,4]. The diagnosis of Chagas disease is challenging because it is often asymptomatic in its acute phase and evolves into a chronic stage in which the disease presents diverse clinical forms [1]. In addition, and due to a major decline in parasitemia during the chronic phase, the detection of T. cruzi in blood samples by direct examination, hemoculture, or xenodiagnosis is difficult and time-consuming. Several PCR-based procedures have been reported that, although highly specific, present suboptimal sensitivity and require technological expertise and specialized expensive laboratory equipment [5]. In this framework, detection of anti-T. cruzi antibodies remains the most effective method for demonstrating direct exposure to the parasite [6]. At present, the most widely used serologic methods are indirect hemagglutination assay (IHA), indirect immuno-fluorescence assay (IIF), and enzyme-linked immunosorbent assay (ELISA) using total parasite homogenates or semipurified antigenic fractions [7]. Despite their satisfactory performance, these tests show variations in their reproducibility and reliability that can be attributed to poor standardization of the reagents or intrinsic variability of immune responses in patient populations [8–10]. In the absence of a single reference test showing 100% specificity and sensitivity, the current guidelines developed by the World Health Organization advise the use of two serologic tests for reaching a conclusive diagnosis. In the case of ambiguous or discordant results, diagnosis using a third technique should be conducted. In addition, there are other still unmet needs and gaps such as access to diagnostics in point-of-care sites for neglected populations [11,12], as well as development of much needed tests for early identification of congenital transmission; rapid assessment of drug treatment efficacy or prognostics tests for disease progression [10,13].
Recombinant DNA and peptide synthesis technologies historically allowed the production and one-step purification of large amounts of T. cruzi immunodominant antigens [14]. However, several studies showed that the use of single antigens in an assay did not confer the sensitivity required for a diagnostic test [14,15], which prompted the development of tests based on combinations of antigens[16,17], some of which were evaluated in multicenter trials and are commercially available [18–20]. Synthetic peptides are advantageous for diagnostic applications because they are: i) well defined (ease of quality control), ii) easily produced in large amounts, ii) highly pure and often cost-saving if compared to the production of natural or recombinant antigens in vitro [21]; and iv) also chemically stable (can be stored lyophilized or dessicated and tend to be stable for several years).
Short synthetic peptides spanning linear B-cell epitopes can also be used in serodiagnostic applications to increase specificity (that is, decrease the number of false positives) by replacing the use of whole protein antigens, therefore avoiding the display of unnecessary sequences that may lead to ‘false positive’ results. Specificity is a critical issue in serodiagnosis of Chagas Disease, where most reagents present cross-reactivity against other co-endemic parasites such as Leishmania spp. [18,21]. Peptide sensitivity, on the other hand can be increased using more densely presented immunoreactive epitopes (i.e. by creating a synthetic poly-epitopic molecule) or by combining multiple antigenic peptides in a single multiplex-assay [21–23]. A number of studies described the use of short peptides, containing either one or several epitopes for diagnosis of Chagas disease and other infectious diseases [23–34].
Recently, we have prioritized a number of candidate diagnostic targets from the genome of T. cruzi [35] and conducted a large-scale screening of parasite B-cell linear epitopes using high-density peptide microarrays [36]. This approach led to the identification of several hundred novel epitopes associated to chronic Chagas Disease, from which we selected 30 for further characterization. In this paper, we describe their diagnostic evaluation in ELISA format using a large panel of serum samples. In addition, and following an in silico-guided antigen combination strategy, we developed a proof-of-principle diagnostic kit based on these reactive peptides.
More than 2,000 candidate serodiagnostic peptides were previously identified by our group using a T. cruzi/Chagas HD peptide microarray [36]. To guide the selection of a subset of peptides for further serological characterization, a filtering strategy was conducted, as follows. First, peptides with serodiagnostic potential (high signal-to-noise ratio in the microarray experiments) were mapped to 187 distinct antigenic protein regions (stretches of adjacent peptides in a protein sequence). These antigenic regions may contain either a single B-cell linear epitope or, in some cases, a limited number of partially overlapping epitopes [37]. Next, antigenic regions were grouped into clusters of sequence-related peptides, in such a way that peptide sequences sharing stretches of 7 or more identical amino acids were put into the same cluster. We reasoned that peptides within a cluster may be both sequence and also likely antigenically related, whereas peptides from different clusters may likely represent the targets of different antibody specificities. From each cluster only a single antigenic region was kept (the one with highest microarray average seroreactivity). After this filter 95 unique antigenic regions were obtained (non-redundant, non-similar). From this set we selected 30 peptides from the top of the ranking for further characterization (the most reactive 15-mer from each antigenic region was selected). To minimize possible bias in our selection, the number of selected peptides from overrepresented sequences such as those from the mucin-associated surface protein (MASP) family [38] and from previously known antigens with mapped epitopes [24,39–43] was limited to 3 and 4, respectively. Sequence and features of our final set of synthetic peptides is summarized in Table 1.
Peptides in Table 1 were synthesized and used in ELISA assays as described below (see also Results) to screen for reactivity against Chagas positive and negative (control) samples. Once we obtained a first matrix of reactivity of peptides vs individual serum samples, we applied the EpiSelect algorithm to guide the selection of sets of peptides for the formulation of multiepitope assays. Implementation of the algorithm has been described [47], but briefly the algorithm aims to find the smallest selection of peptides (epitopes) that in concert maximizes the coverage (reactivity) against a given set of subjects. The input to the algorithm was the matrix of peptide reactivity values determined by ELISA, encoded as z-scores defined as the number of standard deviations above background. Positive peptides were defined using a z-score threshold of 3.
Synthetic peptides were purchased from Schafer-N (Copenhagen, Denmark). Peptides were synthesized using standard FMOC chemistry, purified by HPLC (> 90% purity) and characterized by mass spectroscopy. A C-terminal cysteine residue was included in all peptides for conjugation to maleimide-activated BSA. An additional amino acid residue (leucine) was added at the N-terminus of peptide p1, to avoid the partial deamination associated with an N-terminal glutamine [48]. Lyophilized peptides were resuspended in sterile-filtered water (Sigma Product w3500), and conjugated to maleimide-activated BSA (mBSA, Sigma-Aldrich Product B7542) according to the manufacturer’s protocol, using a molar ratio of 35:1 peptide to mBSA [49]. Peptide-mBSA conjugates were stored in 50% glycerol at -20°C until use. Peptides that failed to solubilize under these conditions were discarded for the analysis.
Human serum samples from T. cruzi-infected patients used in this study were obtained from the Laboratorio de Enfermedad de Chagas, Hospital de Niños "Dr. Ricardo Gutierrez" (HNRG, Buenos Aires, Argentina) (n = 80). Human serum samples from patients with American Tegumentary Leishmaniasis (ATL) used in this study were obtained from the Instituto de Patología Experimental, Universidad Nacional de Salta (IPE, Salta, Argentina) (n = 19). All procedures were approved by the research and teaching committee and the bioethics committee of both institutions, and followed the Declaration of Helsinki Principles. Written informed consent was obtained from all individuals (or from their legal representatives), and all samples were decoded and de-identified before they were provided for research purposes. Chagasic patients were in the asympomatic chronic stage of the disease without cardiac or gastrointestinal compromise (age range: 11 to 51 years old, median age: 20). Serum samples were collected from clotted blood obtained by venipuncture and analyzed for T. cruzi-specific antibodies with the following commercially available kits: ELISA using total parasite homogenate (Wiener lab, Argentina) and IHA (Polychaco, Buenos Aires, Argentina). ATL patients were diagnosed using a combination of techniques: direct observation of parasites (amastigotes) on smears of dermal scrapings; a test of delayed-type hypersensitivity (Montenegro skin test); and a clinical assessment (see [50]). The negative panel was composed of samples from healthy, non-infected individuals that gave negative results in the aforementioned tests, and were obtained either from the blood bank “Fundación Hemocentro Buenos Aires” (FHBA Buenos Aires, Argentina) (n = 82) or from IPE (n = 18). Samples from FHBA were also negative for HIV, Hepatitis B, Hepatitis C, HTLV I and II, Treponema pallidum (syphilis) and for Brucelosis (Huddlesson test).
To calculate the minimum sample size required to estimate sensitivity or specificity for a specified interval of confidence and precision under a normal approximation, we used the following formula:
n=Z2P^(1−P^)d2
Where Z is the z-score from a standard normal distribution (e.g. 1.96 for a 95% confidence interval), P^ is the pre-determined (guess) value of sensitivity (or specificity) based on previous experience/judgment, and d is the required precision [51]. Therefore, for Z = 1,96 (95% CI), P^ = 0.99, and d = 0.05 (5% error), the estimated sample size is 73. Therefore 73 is the minimum number of Chagas positive samples (to estimate sensitivity) and Chagas negative samples (to estimate specificity).
Microplates containing 96 or 384-wells (Thermo Scientific ImmunoPlates, MaxiSorp) were coated overnight at 4°C with 100 ng/well of peptide-mBSA or with different peptide mixtures (80 ng/well of each one) in PBS pH 7.4. Blank signal was determined using mBSA-coated wells. After 4 washings with TBS-T (50 mM Tris-HCl (pH 7.6), 150 mM NaCl, 0.05% (v/v) Tween20), the plates were blocked for 1 h at room temperature with 100 μl/well of assay buffer (3% (w/v) skimmed milk in TBS-T). The plates were washed and incubated for 1 h with human sera diluted as indicated (1:100 or 1:10) in assay buffer at room temperature. Optimization of the assay conditions was performed by a checkerboard titration analysis using 10 ng or 80 ng of peptide-mBSA, and different dilutions of secondary antibody (peroxidase-conjugated goat anti-human IgG antibodies (Sigma-Aldrich, St Louis, MO) (1:5,000; 1:10,000; 1:20,000 and 1:80,000). After washings, 100 μl of secondary antibody diluted as indicated (1:10,000 for assays using a single peptide per well, or 1:80,000 for multiepitope assays) in assay buffer were added to each well and incubated for 1 h at room temperature. Following additional washings with TBS-T, the reaction was developed with tetramethylbenzidine for 15 min (TMB, Sigma-Aldrich, St Louis, MO) and stopped by addition of 0.2 M sulphuric acid. Absorbance values were measured at 450 nm in a microplate absorbance reader (FilterMax F5 Multimode, Molecular Devices, Sunnyvale, CA, USA). All serum samples were tested in duplicate. Values were averaged and blank-corrected.
The same 16 serum samples from healthy blood donors were tested in each ELISA plate. The cut-off value was determined for each peptide and for each plate using the mean of the control blood donor samples plus 3 SD (the cut-off was set accounting for multiple-hypothesis testing). For each peptide or peptide mixture, standardized reactivity scores (z-scores) and the diagnostic analytical characteristics of sensitivity, specificity and AUC (Area under the ROC–Receiver Operating Characteristic–curve, as a performance metric) were calculated. Reagent sensitivity was calculated as the number of positive subjects (i.e. infected patients samples that were reactive against a particular peptide) over the total number of infected subjects tested; specificity was calculated as the number of negative subjects (non-infected control subjects that were seronegative against a particular peptide) over the total number of non-infected control subjects tested and AUC was calculated using the from the z-scores of infected subjects and non-infected subjects. For receiver operating characteristic (ROC) analyses [52], the results were expressed as the percentage of reactivity of the mean absorbance at 450 nm of the positive reference control serum included in each assay run. The Mann-Whitney test and ROC analysis were performed using the GraphPad Prism software (version 6 for OSX; San Diego, CA, USA) or ROCR R package [53].
Based on our previous screening of serodiagnostic peptides for Chagas Disease using HD peptide microarrays [36], 30 peptides were selected for further serological characterization and downstream validation. The strategy for selection of these peptides is outlined in Fig 1 (see also Methods), and essentially was guided to select a non-redundant set of peptides showing the highest antibody-binding signal in any array. After removing 3 peptides that showed solubility problems, the remaining 27 peptides were coupled to a carrier protein (mBSA) and assayed in ELISA format against a sera panel of 62 chronically infected Chagasic patients and 16 healthy controls. Initially, all human sera were tested at 1:100 dilutions. The panel of peptides included 16 peptides corresponding to previously uncharacterized T. cruzi proteins (novel antigens) that emerged during our screening [36], 7 peptides representing novel epitopes in previously characterized B-cell antigens and 4 peptides corresponding to previously known linear B-cell epitopes, which were used as positive controls (see Table 1 and S1 Fig). We also included in our panel an additional peptide (p17) as an internal negative control. Although belonging to a validated T. cruzi antigen [54], this peptide was derived from a protein region that showed consistently very low signal in all microarray replicates.
Diagnostic sensitivity, specificity and AUC values for each peptide are shown in Table 2 (complete data available in S1 Table). The diversity of reactivities in the collection of sera samples when assayed against individual peptides is also evident when visualizing the data in the form of a heatmap plot (available in S2 Fig). As shown, promising diagnostic performances were observed for most of the assayed peptides. Sensitivity values ranged from 30–92% (>50% in 22 out of 27), and specificity values were extremely high, which is consistent with our screening strategy [36]. In this context, it is worth noting that sensitivity values of all individual T. cruzi antigens described so far and proposed and/or included in serodiagnostic tests ranged from 80–99% [14].
Overall, and as previously reported for the TSSA antigen [37], a strong correlation between assays in the standard ELISA format and in microarray format was observed for each peptide (Table 1), thus providing additional validation and support for the use of HD-peptide arrays for discovery of new serology-based biomarkers.
We further evaluated the diagnostic specificity of the 16 best performing peptides (see Table 1) by using an extended panel of 61 control sera obtained from healthy subjects (Chagas-negative samples). As before, individual peptides coupled to mBSA were assayed in ELISA format. Diagnostic specificities and ROC-AUC were re-calculated for each peptide (top entries in Table 2). The average specificity was 97.23% and in all cases specificities > 95% were observed. Notably, most of the positive responses observed in this expanded set of Chagas-negative samples correspond to only 3 of the 61 sera samples tested. These samples (also negative for the highly-sensitive trans-sialidase inhibition assay [55]) were highly reactive against more than half of the peptides (12, 11 and 9 peptides each, see S1 Table in the ‘Additional negative sera’ section), suggesting a broad and yet-to-be explained cross-recognition towards T. cruzi-derived sequences. If these Chagas-negative serum samples were removed, specificity values of our peptides would increase up to an average 98.5%.
Based on the results described above, we undertook an in silico-guided approach to design a multiplex assay with improved diagnostic performance. Using ELISA data from individual peptides, we applied the EpiSelect algorithm [47] (see Methods) to identify several optimal (minimal) virtual peptide sets that in concert provided maximal coverage of the analyzed subjects. This analysis was performed after removing data from the 9 serum samples that were previously used in microarray experiments, to avoid optimistically biased results. The analysis performed on the tested peptides and 53 Chagas-positive subjects showed that 3 peptides were enough to reach a theoretical sensitivity of 100% (Fig 2). Data used for this analysis is available in S1 Table. The optimal set was composed by peptides {pc1, pc2, and p6}, resulting in an average of 2.51 reactive peptides per subject, closely followed by the peptide set {pc2, p11, and p6} with an average of 2.43 reactive peptides per subject. The reactivity patterns for these sets are shown in Fig 2 and S1 Table. Interestingly, at least 1 of the 3 novel peptides p6 (as in Fig 2), p2 or p8 (alternatives) would be required to achieve a sensitivity of 100% with a 100% specificity (see also S1 Table).
Other peptides such as p5, p7, p11, p12, p16, p19 and p24 also displayed excellent diagnostic characteristics, with individual high sensitivity (> 70%) and specificity (up to 95%). Hence, these peptides can be eventually incorporated into the multiplex design to increase its robustness (for example, to increase the number of reactive peptides per subject).
Based on these analyses, we prepared and tested a number of multi-epitope peptide combinations in ELISA format against an extended panel of sera from chagasic (positive) and healthy (negative) subjects. One such combination {pc1, pc2, pc3, p6, p13}, was tested against 22 positive and 24 negative serum samples and gave a diagnostic sensitivity of 72.7% and a specificity of 91.7%. Following the same methodology (S1 Table), we tested a slightly different formulation of peptides (pc1, pc2, p6, p7 and p24) against an increased number of sera samples (53 Chagas-positive and 31 Chagas-negative) obtaining an improved performance, with a sensitivity of 92.45% and a specificity of 93.55%.
Finally, with the aim of obtaining a peptide combination with enhanced robustness, we re-analyzed the reactivity profile of each individual serum sample (S1 Table) against our panel of peptides, and identified a few Chagas positive subjects that gave low or even negative reactivity to many peptides. From this analysis, we identified peptides that would theoretically maximize the sensitivity of the multiplex assay, despite not showing the best possible coverage of our subject (sera) collection. Thus, we arrived at a high performance multi-epitope formulation of seven peptides {pc1, pc2, pc3, p6, p7, p13, and p24}. To validate this final formulation, we increased the amount of coated peptide to 80 ng of each peptide per well and the serum concentration to 1:10. After these modifications, the performance of this formulation, when tested against 82 Chagas-positive and 80 Chagas-negative sera samples gave a sensitivity of 96.34% and a specificity of 100%, with an AUC value of 0.9974 (Fig 3).
We have also assessed the performance of this multiepitope formulation against a panel of 19 sera from subjects with positive diagnosis for American Tegumentary Leishmaniasis (see Methods), and another 18 negative (control sera) from the same endemic region. Only a single (negative) subject gave a positive response in our multiepitope assay (Fig 3C). Except for this case, the observed absorbance in the ELISA assays was nil. The specificity of the multiepitope formulation for this panel was 97.30%, with an overall specificity (considering all negative samples from all panels) of 99.15%. Table 3 summarizes the performance of this combination of peptides. This therefore represents a highly promising novel multiepitope formulation for the diagnosis of Chagas Disease.
Serological diagnostics methods for infectious diseases have usually evolved from first-generation lysate-based reagents. Through time, more defined formulations of diagnostic reagents have followed. Second-generation diagnostic kits based on purified antigenic fractions or third-generation kits based on recombinant proteins are now in widespread use. To develop new diagnostic tools that are simple and have few manipulation steps, one of the central aspects that currently limits the suitability of diagnostic kits is the need to produce, prepare and purify the antigens, along with the corresponding quality control. Short synthetic peptides can be produced cheaply in large quantities, and are chemically stable and amenable for long-term storage. Synthetic peptides have been already tested in a wide range of diagnostic applications and proved valuable for diagnosis of viral, bacterial, parasitic and autoimmune diseases [21,30–34]. Therefore, fourth-generation diagnostic kits based on well-defined peptidic antigens are now within reach.
Here we present a next-generation diagnostic formulation for Chagas Disease based on short peptides. Significant efforts have been invested by various groups over time to identify and test antigenic peptides for serodiagnosis of Chagas Disease, some of which displayed promising analytical characteristics. For example, peptides Ag2/B13/Pep2, TcD/Ag13, TcE and TcLo1.2, have been combined to create a multi-epitope recombinant neo-protein of excellent performance [24], and peptides from the cytoplasmic repetitive antigen (CRA)/Ag30 and flagellar repetitive antigen (FRA)/Ag1 [54] have been recently shown to present good specificity and sensitivity [56].
The advent of novel high-throughput approaches spawned by the post-genomic era is starting to impact on the discovery of new biomarkers and the development of diagnostic tools for a number of important pathogens [10]. We have recently showed the utility of a fast approach to screen for new T. cruzi antigens that is based on high-density peptide microarrays [36]. The advantage of this platform is that it allows to identify antigens and at the same time obtain a fine mapping of their linear epitopes. Using this strategy we have identified and mapped the epitopes of >90 novel T. Cruzi antigens [36].
As a followup of this first screening for peptidic antigens, we provide here an extensive serological characterization of 27 peptides, 18 of which represent novel epitopes that were mapped using our strategy, or represent recently discovered antigens but for which no fine epitope mapping was yet available (see Table 1). For example, even though the trans-sialidase/SAPA antigen (accession number X57235, TcCLB.509495.30 is the most similar genome locus tag) has been known for quite some time, peptide p13 (also annotated as ‘trans-sialidase’) is not derived from the originally described antigen, but from another member of the superfamily (TcCLB.506961.25) with only 29% identity to the original trans-sialidase/SAPA. Therefore, p13 is a new/novel antigen and epitope that bear no resemblance to any of the mapped epitopes already described [43,57]. Similarly, even though the proteins encoded by the genes TcCLB.511633.79 (microtubule-associated protein), or TcCLB.506391.30 (EF-hand protein 5) were already described and used as antigens [20,46], this is the first time that their fine mapped epitopes are tested for diagnostic purposes. Other peptides such as p16, p7, p11 and p19 are part of proteins that have been identified as potential antigens [35] but with no other serological evidence before our microarray experiments. Peptide p1, on the other hand, was derived from a member of the Mucin-Associated Surface Protein (MASP) family [38], which is a large family of genes which were shown recently to be the target of the adaptive immune response in an animal model of infection [58]. The MASP protein encoded by gene TcCLB.507071.20 was selected from the genome, as part of an effort to obtain a detailed characterization of the antigenicity and epitopes of this gene family in human infections [59]. Peptide p6 contains a slightly different version of the sequence TTRAPSRLREID, which has been identified as the major and conserved linear B-cell epitope included within the otherwise highly polymorphic TcMUCII family of T. cruzi proteins [44,60]. Whereas peptide p2 is a novel epitope from a putative 60S ribosomal protein L7a, that we have also previously identified as a potential antigen [35].
Using a panel of Chagas-positive and negative (control) samples, we performed a thorough serological characterization of the selected peptides. This allowed us to obtain a relatively large matrix of ELISA responses for all peptides against individual serum samples. This led us to identify a number of peptides with promising diagnostic potential, such as peptides p1, p7, p11, p16 and p19, which presented sensitivities above 80%, with no false positive responses in the first evaluation using a small panel of 16 sera, and only a few false positive responses (with specificities from 96.5% to 100%) in a second evaluation using a larger panel of sera. These sensitivities are similar to those originally reported in the first characterizations of validated serodiagnostic antigens such as TcD (95% for chronic subjects [61]) and SAPA (10% for chronic subjects, 90% for acute infection [62]), which were later improved when developed into a multiantigen diagnostic reagent (e.g. the Chagatest kit of Wiener Labs that includes these antigens claims a sensitivity of 98.8%[63]). Hence, even if some peptides displayed sensitivities that were not very high when assessed singly, they were high enough as to keep them under consideration for development of an assay based on combinations of peptides.
The matrix of ELISA responses was then used to guide the rational formulation of a multiepitope diagnostic reagent using a well-defined algorithm for the inclusion of peptides. The first combinations tested did not achieve a significantly high performance, even if the theoretical prediction (Fig 2) would suggest otherwise. One reason for this is that even though the input to the EpiSelect algorithm included the level of response of each subject against each peptide (represented as the number of standard deviations above negative controls), the effect of combining peptides produced a higher background signal that was not predicted by the algorithm. Another reason was the inclusion in our panel of Chagas-positive sera of several samples with moderately low antibody titers overall (see for example the 9 sera grouped in the bottom branch in S2 Fig). Despite these pitfalls, the detailed data present in this matrix was pivotal in identifying peptides for inclusion in the final multiepitope formulation. The rationale for inclusion of peptides was the ability of a given peptide (as observed in the matrix) to potentially overcome a negative response for a given serum sample. For example, peptides p6 and p2, followed by p11 represented an optimal complement of the two best performing peptides, pc1 (from the antigenic repeat of the CA-2/B13 antigen Ag2) and pc2 (the serodiagnostic epitope TcE) for diagnosis. Also, peptide p13 when combined with peptides pc1 and pc2 was one of the few peptides that provided relatively high signal in the ELISA assay against the group of sera with relatively low overall responses. The fact that we could consistently increase the performance of each combination upon following this rationale shows the usefulness of this approach.
Interestingly, all peptides in the final multiepitope formulation are highly conserved (see S1 Text). A sequence similarity search across available complete genomes (e.g. those from the CL-Brener [64] and Sylvio X10 [65] strains using BLASTP) or from draft assemblies (Tula cl2, Esmeraldo cl3, Dm28c or JRcl4 in the TriTrypDB resource [66], release 30 from February 2017, using TBLASTN) shows that all peptides are highly conserved across strains representing different evolutionary lineages of the parasite (TcI, TcII, TcV, TcVI).
The observed diagnostic performances for all peptides and peptide combinations tested were very promising, particularly considering that all assays were based on short synthetic peptides. Our final best performing multi-epitope combination was based on a combination of seven antigenic peptides. With an equimolar mixing of peptides, we attained a very high (>96%) level of sensitivity and specificity. These are highly promising values for a first optimization attempt; the final ELISA assay/formulation could be indeed further improved using different blocking reagents, coupled detection system and, most importantly, by adjusting the relative concentration of different peptides in the final mixture.
Analysis of potential cross-reactivity with other co-endemic diseases and pathogens is essential to validate any diagnostic reagent. In the case of Chagas Disease, cross-reactivity against infections with Leishmania species is a particular concern [67]. We have included a panel of serum samples from confirmed cases of tegumentary leishmaniasis from the northern province of Salta, Argentina to assess the performance of our formulation. This also gave us the opportunity to improve the assessment of specificity by analizing a paired set of negative (control) samples (chagas-negative and leishmaniasis-negative) from the same endemic region. From a set of 37 of these samples which were negative for Chagas Disease, only one gave a positive cross-reactive response (Fig 3). Although this highlights the need to perform a more extensive characterization of this cross-reactive sample (e.g. against our complete panel of peptides), and eventually revise the combination of peptides in our formulation, the current multiepitope assay has a sufficiently high specificity at this stage (99.15%), comparable to other commercially available kits [63] that can certainly be improved by optimization of the assay or by replacing of cross-reactive peptides.
Besides the obvious attention to the diagnostic performance of the identified peptides, these results serve to validate the use of high-density peptide microarrays as a fast screening platform. The fact that all selected peptides gave positive responses against several Chagas-positive subjects show that this technology can be trusted to rapidly identify and map epitopes of complex pathogens. It is also worth mentioning here that there are about a hundred additional antigenic regions within the signal range observed in the peptide microarray screening from which these peptides were identified [36] and that await further serological characterization. This observation, together with the fact that the microarray screening only covered ~3% of the parasite proteome, show that there is still a large repertoire of Chagas-specific antibody specificities that remain serologically unexplored.
The results presented herein hence provide a novel, robust multi-epitope formulation as a basis for the development of improved peptide-based serodiagnostics for Chagas Disease. In contrast with chimeric DNA constructs that encode multiepitope recombinant proteins, the fact that this diagnostic reagent is based on the combination of short peptides that can be synthesized separately and easily formulated in a mix-and-match approach, means that it can be improved successively over time with only a reasonable effort.
|
10.1371/journal.pbio.0050063 | Cdk5 Is Involved in BDNF-Stimulated Dendritic Growth in Hippocampal Neurons | Neurotrophins are key regulators of neuronal survival and differentiation during development. Activation of their cognate receptors, Trk receptors, a family of receptor tyrosine kinases (RTKs), is pivotal for mediating the downstream functions of neurotrophins. Recent studies reveal that cyclin-dependent kinase 5 (Cdk5), a serine/threonine kinase, may modulate RTK signaling through phosphorylation of the receptor. Given the abundant expression of both Cdk5 and Trk receptors in the nervous system, and their mutual involvement in the regulation of neuronal architecture and synaptic functions, it is of interest to investigate if Cdk5 may also modulate Trk signaling. In the current study, we report the identification of TrkB as a Cdk5 substrate. Cdk5 phosphorylates TrkB at Ser478 at the intracellular juxtamembrane region of TrkB. Interestingly, attenuation of Cdk5 activity or overexpression of a TrkB mutant lacking the Cdk5 phosphorylation site essentially abolishes brain-derived neurotrophic factor (BDNF)–triggered dendritic growth in primary hippocampal neurons. In addition, we found that Cdk5 is involved in BDNF-induced activation of Rho GTPase Cdc42, which is essential for BDNF-triggered dendritic growth. Our observations therefore reveal an unanticipated role of Cdk5 in TrkB-mediated regulation of dendritic growth through modulation of BDNF-induced Cdc42 activation.
| Accurate transmission of information in the nervous system requires the precise formation of contact points between neurons. Regulation of these contact sites involves fine tuning the number and branching of dendritic processes on neurons. Throughout development, several secreted factors act to regulate dendrite number and branching. One important family of these factors is neurotrophins, which are indispensable for the survival and development of neurons. For example, stimulation of hippocampal neurons with one neurotrophin, brain-derived neurotrophic factor (BDNF), increases the number of dendrites directly extending from the cell body. Here, we report that BDNF-stimulated dendritic growth requires phosphorylation of the BDNF receptor, TrkB, by a kinase known as cyclin-dependent kinase 5 (Cdk5). Inhibiting phosphorylation of TrkB by Cdk5 essentially abolishes the induction of dendrites by BDNF. Our observations reveal that Cdk5 serves as a regulator of neurotrophin function. Since Cdk5 and neurotrophins both play essential roles in neuronal development, our findings suggest that the interplay between Cdk5 and TrkB may also be implicated in the regulation of other biological processes during development.
| Neurotrophins are indispensable for multiple aspects of neuronal development, such as the maintenance of neuronal survival, regulation of neuronal architecture, and synaptic plasticity. Members of the neurotrophins include the prototypic member nerve growth factor (NGF), brain-derived neurotrophic factor (BDNF), neurotrophin (NT)–3, and NT-4/5. Downstream responses of neurotrophins are transduced by a family of receptor tyrosine kinases (RTKs) known as Trks, and also the low-affinity neurotrophin receptor p75. Although all neurotrophins bind p75, they associate with different Trk receptors with rather remarkable selectivity. NGF interacts selectively with TrkA, while BDNF and NT-4/5 bind preferentially to TrkB. NT-3, on the other hand, associates with TrkC with high affinity, although it also binds TrkA and TrkB with low affinity. Similar to other RTKs, activation of Trks leads to dimerization and autophosphorylation of the receptors, followed by the recruitment and initiation of a myriad of signaling pathways including the Ras/MAPK, PI3K, and PLCγ pathways [1,2].
Interestingly, recent studies have demonstrated that activity of cyclin-dependent kinase 5 (Cdk5), a serine/threonine kinase, is required for the downstream actions of a RTK, ErbB. Cdk5 was found to phosphorylate ErbB2/3, a phosphorylation that is essential for the activation of the receptors [3,4]. Cdk5 is a member of the cyclin-dependent kinase family, but it is unique in several aspects. First of all, it is activated by the neural-specific non-cyclin activators p35 and p39. Secondly, Cdk5 is not involved in the regulation of cell cycle control, but is implicated in neuronal migration, synapse functions/maintenance, and neuronal survival [5,6]. The importance of Cdk5 in neuronal development and migration is underscored by the aberrant phenotypes exhibited by mice lacking Cdk5 and its activators. Cdk5 knockout mice and p35/p39 double knockout mice both exhibit perinatal death with severe cortical lamination defects [7,8]. Furthermore, swollen soma and nuclear margination is evident in Cdk5-deficient neurons, implicating Cdk5 as an essential regulator of neuronal survival [7]. Interestingly, truncation of the Cdk5 activator p35 into p25 has also been associated with prolonged Cdk5 activation in a number of neurodegenerative diseases [9], thus revealing that precise regulation of Cdk5 activity is essential for maintenance of neuronal survival [10]. Furthermore, an increasing number of studies are pointing to an essential role of Cdk5 at the synapse, where it is not only involved in the formation and maintenance of synapses, but is also indispensable for the regulation of synaptic transmission and synaptic plasticity [5]. While the mechanisms by which Cdk5 regulates such diverse functions remain to be unraveled, the identification of ErbB receptors as Cdk5 substrates suggests that Cdk5 may exert its biological effects by modulating signaling pathways downstream of RTK activation. This piece of evidence, together with the abundant expression of Cdk5 and Trk receptors in the nervous system and their shared implication in a number of biological functions, prompted us to further examine if Cdk5 also regulates the signaling of Trk receptors.
In the current study, we report the identification of TrkB as a substrate of Cdk5. More importantly, we found that Cdk5-mediated phosphorylation of TrkB is essential for BDNF-induced dendritic growth through the modulation of Cdc42 activity. Our findings provide evidence for a crosstalk between the Cdk5 and neurotrophin signaling pathways, and lend further support to the idea that Cdk5 is a modulator of RTK signaling.
Given the increasing evidence implicating Cdk5 in the modulation of RTK signaling, we sought to examine if Cdk5 may also play a role in Trk signaling. Literature search revealed that TrkA, TrkB, and TrkC all contain serine- or threonine-directed proline residues at the intracellular juxtamembrane region of the receptors, but only TrkB and TrkC contain Cdk5 consensus sites S/TPXK/H/R (Figure 1A). To explore the potential interplay between Trk receptors and Cdk5, we first examined if Trk receptors associated with Cdk5 or p35. TrkA, TrkB, or TrkC was overexpressed together with Cdk5 or p35 in COS7 cells, and immunoprecipitation was performed with Cdk5, p35, or pan-Trk antibody. Interestingly, all three Trk receptors were observed to associate with Cdk5 (Figure 1B) and p35 (Figure 1C), while no association was observed when immunoprecipitation was performed with IgG control. Since both TrkB and its ligand BDNF are abundantly expressed in the brain throughout development, we next proceeded to verify the interaction between TrkB and Cdk5/p35 in postnatal brains. We found that TrkB associated with both p35 and Cdk5 in postnatal day 7 (P7) rat brain lysates (Figure 1D). Furthermore, Flag-tagged Cdk5 pulled down TrkB from the membrane fraction of adult brain lysates (Figure 1E). These observations collectively suggest that TrkB interacted with Cdk5/p35 in both postnatal and adult brains. Since both p35 and Cdk5 are present in brain lysates and likely exist as a complex, the observed interaction between TrkB and Cdk5/p35 did not provide specific information on whether TrkB associated specifically with Cdk5 or p35. To delineate between these two possibilities, the interaction between TrkB, p35, and Cdk5 was examined in p35+/+ and p35−/− brain lysates (Figure 1F). Interestingly, we found that in the absence of p35, the association between Cdk5 and TrkB was essentially abolished, indicating that p35 was required for the association between Cdk5 and TrkB in vivo.
We next proceeded to examine if Trk receptors, TrkB in particular, served as Cdk5 substrates using in vitro kinase assay. TrkA, TrkB, and TrkC were overexpressed in COS7 cells and immunoprecipitated by pan-Trk antibody. Incubation with Cdk5/p25 revealed that TrkB and TrkC, but not TrkA, were phosphorylated by Cdk5/p25 in vitro (Figure 2A). This is in agreement with the lack of Cdk5 consensus sites in TrkA, and points to the possibility that Cdk5 may phosphorylate TrkB and TrkC at the Cdk5 consensus sites at the juxtamembrane region (Figure 1A). To examine this possibility, a GST fusion protein containing only the juxtamembrane region of TrkB was prepared. In vitro kinase assay verified that Cdk5/p35 phosphorylated TrkB at the juxtamembrane region (Figure 2B). It has previously been proposed that p25 and p35 may confer different substrate specificities. Results from our in vitro kinase assay suggested that Cdk5 phosphorylated TrkB regardless of whether it was activated by p25 or p35, although further studies will be required to delineate the relative contributions of p25 and p35 to endogenous phosphorylation of TrkB by Cdk5.
We were next interested in identifying the Cdk5 phosphorylation site(s) on TrkB. Three TrkB-juxtamembrane region mutants were generated: TrkB M1, where Ser478 was mutated to alanine; TrkB M2, where Thr489 was mutated to alanine; and TrkB DM, where both Ser478 and Thr489 were mutated to alanine. Interestingly, phosphorylation of the TrkB-juxtamembrane region was almost completely abolished when Cdk5/p25 was incubated with TrkB M1 or TrkB DM (Figure 2C), thus revealing that Ser478 was required for Cdk5-mediated phosphorylation of the TrkB-juxtamembrane region. We further verified the importance of this site for Cdk5-mediated phosphorylation of TrkB by generating a phospho-specific TrkB antibody against Ser478. Preincubation of the antibody with blocking peptide prevented detection of Ser478-phosphorylated TrkB, indicating that the antibody was sufficiently specific (Figure 2D). Full-length TrkB mutants lacking the potential Cdk5 phosphorylation sites were overexpressed with or without Cdk5/p35 in HEK293T cells. Interestingly, Ser478-phosphorylated TrkB was not observed in the absence of Cdk5/p35, indicating that Cdk5 was essential for the phosphorylation of TrkB at Ser478 in HEK293T cells. More importantly, when TrkB mutants lacking Ser478 were expressed (TrkB M1 and TrkB DM), phosphorylation of TrkB at Ser478 was essentially abolished (Figure 2E). Taken together, our observations indicate that Cdk5 phosphorylated TrkB at Ser478 at the juxtamembrane region of TrkB.
To further examine if Cdk5 is essential for phosphorylation of TrkB at Ser478 in vivo, we examined the effect of inhibiting Cdk5 activity on phospho-Ser478 (p-Ser478) TrkB levels in cortical neurons. We found that at basal level, TrkB was weakly phosphorylated at Ser478. Interestingly, stimulation with BDNF led to a marked increase in p-Ser478 TrkB levels, indicating that phosphorylation of TrkB at Ser478 was at least in part ligand dependent. Remarkably, treatment with Cdk5 selective inhibitor roscovitine (Ros) almost abrogated the BDNF-triggered increase in p-Ser478 TrkB levels (Figure 3A), suggesting that Cdk5 was involved in the BDNF-stimulated component of TrkB Ser478 phosphorylation. To further establish the involvement of Cdk5 in Ser478 phosphorylation of TrkB in vivo, the levels of p-Ser478 TrkB in cdk5+/+ and cdk5−/− brain lysates were examined. Importantly, we found that Ser478-phosphorylated TrkB was basically undetectable in Cdk5−/− brain lysates (Figure 3B). Similarly, cortical neurons prepared from Cdk5−/− brains exhibited undetectable levels of p-Ser478 TrkB. In addition, BDNF stimulation failed to trigger an increase in p-Ser478 TrkB levels (Figure 3C). These observations strongly suggest that Cdk5 is essential for phosphorylation of TrkB at Ser478 in vivo, and that BDNF-stimulated increase in Ser478 phosphorylation of TrkB requires Cdk5 activity.
Since BDNF stimulation was observed to increase Ser478 phosphorylation of TrkB, and Cdk5 was required for phosphorylating TrkB at Ser478, we were interested to examine if BDNF stimulation affects Cdk5 activity. BDNF has previously been observed to increase Cdk5 activity after 3 d of BDNF stimulation in cortical neurons [11]. In agreement with this observation, we found that BDNF treatment led to an increase in Cdk5 activity within 15 min of BDNF stimulation (Figure 4A). More importantly, addition of Trk inhibitor K252a essentially abolished BDNF-triggered increase in Cdk5 activity, indicating that the increase in Cdk5 activity was dependent on TrkB activation (Figure 4B). It has previously been demonstrated that Cdk5 activity is enhanced by phosphorylation at Tyr15 [12]. Given the activation of tyrosine kinase activity of TrkB upon ligand stimulation, we were interested to investigate if BDNF treatment leads to phosphorylation of Cdk5 at Tyr15, thereby enhancing its activity. We found that BDNF stimulation enhanced association between Cdk5 and TrkB in cortical neurons (Figure 4C). More importantly, in vitro kinase assay using purified TrkB and Cdk5 revealed that TrkB phosphorylated Cdk5 at Tyr15 (Figure 4D and 4E). TrkB-mediated phosphorylation of Cdk5 was abolished with the addition of Trk inhibitor K252a, further verifying that Tyr15 phosphorylation of Cdk5 was TrkB dependent (Figure 4E). These observations collectively indicate that upon BDNF stimulation, Cdk5 was recruited to TrkB and phosphorylated by TrkB at Tyr15, thus leading to enhanced Cdk5 activity to promote phosphorylation of TrkB at Ser478.
Given the neural-specific nature of Cdk5, its abundant expression throughout development, and its essential role in the phosphorylation of TrkB at Ser478, we were interested in examining the biological significance of this phosphorylation on the downstream functions of BDNF/TrkB signaling. As a first step, we examined if Cdk5-mediated phosphorylation of TrkB affects TrkB activation and downstream signaling cascades. Interestingly, we found that inhibition of Cdk5 activity by Cdk5 selective inhibitor Ros only marginally affected tyrosine phosphorylation of TrkB and initiation of downstream signaling pathways including phosphorylation of Erk1/2, Akt, and CREB (data not shown). Indeed, BDNF-stimulated increase in TrkB tyrosine phosphorylation was weakly affected in cdk5−/− cortical neurons (Figure 3C). Furthermore, activation of Akt and Erk1/2 following BDNF stimulation was also comparable in cdk5+/+ and cdk5−/− cortical neurons (data not shown). Our observations thus revealed that Cdk5-mediated phosphorylation of TrkB did not significantly affect activation of the receptor, nor its initiation and recruitment of downstream signaling pathways.
Although Cdk5-mediated phosphorylation of TrkB had negligible effect on the downstream signaling of TrkB, it cannot be ruled out that Ser478 phosphorylation of TrkB is essential for the downstream functions of BDNF/TrkB signaling. We thus sought to examine if Cdk5-mediated phosphorylation of TrkB affects its downstream functions. BDNF has been observed to stimulate dendrite growth and development in hippocampal neurons [13,14]. In accordance with earlier observations, BDNF treatment led to a marked increase in the number of primary dendrites in hippocampal neurons (Figure 5A), although the length and branching of dendrites were not affected (data not shown). Interestingly, treatment with Cdk5 selective inhibitor Ros almost completely abolished the BDNF-stimulated dendritic growth, without affecting the basal number of dendrites (Figure 5A). Furthermore, overexpression of dominant negative (DN) Cdk5 (Figure 5B) and transfection with Cdk5 short interfering RNA (siRNA) (Figure 5C) both abrogated BDNF-induced increase in primary dendrites. More importantly, BDNF similarly failed to induce an increase in primary dendrites in cdk5−/− hippocampal neurons (Figure 5D). These observations collectively reveal that Cdk5 activity was required for BDNF-induced increase in primary dendrites in hippocampal neurons. To verify the importance of Ser478 phosphorylation of TrkB in BDNF-triggered dendritic growth, TrkB wild-type (WT) or TrkB M1 was overexpressed in hippocampal neurons. Remarkably, overexpression of TrkB M1 similarly abolished the BDNF-induced increase in primary dendrites (Figure 5E). Taken together, our data indicate that Cdk5-mediated phosphorylation of TrkB at Ser478 was required for BDNF-triggered dendritic growth in hippocampal neurons.
Rho GTPases, including RhoA, Rac1, and Cdc42, are key regulators of actin cytoskeleton dynamics. Since BDNF stimulation has been observed to activate Rac1 and Cdc42 in neurons [15], we were interested to delineate if Rho GTPases contribute to BDNF-stimulated dendritic growth. To investigate if Rho GTPases are involved, and to identify the Rho GTPase(s) implicated, hippocampal neurons were transfected with WT or DN Rac1, Cdc42, or RhoA. We found that while overexpression of WT and DN Rac1 increased the basal number of dendrites in the absence of BDNF treatment, overexpression of both forms of Rac1 abolished BDNF-stimulated dendritic growth. On the other hand, while overexpression of DN RhoA slightly enhanced primary dendrites irrespective of BDNF stimulation, overexpression of both WT and DN forms of RhoA inhibited BDNF-stimulated dendritic growth. Remarkably, in contrast to the inhibition of BDNF-stimulated dendritic growth in cells overexpressing WT Rac1 and RhoA, BDNF stimulation of hippocampal neurons overexpressing WT Cdc42 resulted in an increase in primary dendrites, which was nearly abolished by overexpression of DN Cdc42 (Figure 6A). Our observations therefore suggest that while Rac1 and RhoA may also modulate BDNF-stimulated dendritic growth, it is the activation of Cdc42 following BDNF stimulation that most likely mediates the increase in primary dendrites by BDNF.
To examine if phosphorylation of TrkB by Cdk5 affects dendritic growth through modulating BDNF-triggered activation of Cdc42, we first examined if the BDNF-induced increase in Cdc42 activity was affected by treatment with Cdk5 selective inhibitor Ros. In agreement with earlier findings, BDNF treatment resulted in an increase in Cdc42 activity. Interestingly, treatment with Ros significantly reduced BDNF-induced Cdc42 activity in cortical neurons (Figure 6B), suggesting that Cdk5 activity was involved in BDNF-triggered activation of Cdc42. To investigate if the reduction in Cdc42 activity contributes to the abrogation of BDNF-induced dendritic growth following attenuation of Cdk5 activity, the effect of overexpressing constitutively active (CA) Cdc42 with TrkB M1 on BDNF-induced dendritic growth was examined. Remarkably, overexpression of CA Cdc42 reversed the abrogation of BDNF-induced dendritic growth by TrkB M1 (Figure 6C). More importantly, while overexpression of CA Cdc42 had negligible effect on BDNF-stimulated increase in primary dendrites in Cdk5+/+ neurons, overexpression of CA Cdc42 similarly rescued the lack of dendritic growth in cdk5−/− neurons following BDNF stimulation (Figure 6D). These observations strongly suggest that Cdk5-mediated phosphorylation of TrkB at Ser478 was essential for the BDNF-triggered increase in primary dendrites through modulating BDNF-induced Cdc42 activity.
In the current study, we report the identification of TrkB as a novel Cdk5 substrate by providing evidence that Cdk5 phosphorylates TrkB at Ser478, located at the intracellular juxtamembrane region of the receptor. The near absence of Ser478-phosphorylated TrkB in cdk5−/− brain underscores the importance of Cdk5 in this phosphorylation in vivo. More importantly, we found that Cdk5-mediated phosphorylation of TrkB is required for BDNF-stimulated increase in primary dendrites. Furthermore, we demonstrated that Cdk5 activity is involved in BDNF-induced increase in Cdc42 activity, which underlies BDNF-induced dendritic growth in hippocampal neurons. Overexpression of CA Cdc42 restored BDNF-stimulated increase in primary dendrites in cdk5−/− neurons, lending further support that Cdk5-mediated phosphorylation of TrkB at Ser478 is essential for BDNF-induced Cdc42 activation and increase in primary dendrites. Our findings therefore reveal an unanticipated role of Cdk5 in mediating downstream functions of Trk signaling.
Activation of Rho GTPases has been implicated in a number of functions downstream of neurotrophin stimulation. For example, a recent study reported that synaptic maturation involves BDNF-stimulated increase in Cdc42 activity [16]. In addition, activation of Cdc42 is involved in the regulation of retinal growth cone filopodia by BDNF [17]. Activation of Rac1 following neurotrophin stimulation has also been observed to mediate neuronal migration triggered by neurotrophin treatment [18]. Our observation that BDNF-stimulated increase in Cdc42 activity contributes to the increase in primary dendrites corroborates these studies. It is interesting to note that overexpression of WT and DN Rac1 and RhoA also inhibited BDNF-induced increase in primary dendrites. While it is rather intriguing to observe similar actions by the WT and DN forms of these two Rho GTPases, our observation nonetheless suggests that Rac1 and RhoA may also play a role in BDNF-stimulated dendritic growth. Further studies will be required to delineate their involvements in BDNF-dependent regulation of dendritic development.
Although different Rho GTPases have been identified as essential downstream mediators of neurotrophin functions, much less is known about the mechanisms by which neurotrophin treatment results in Rho GTPase activation, and how this process is regulated. The activity of Rho GTPases is controlled by a number of factors. Conversion from the GDP-bound, inactive state to the GTP-bound, active state is facilitated by guanine nucleotide exchange factors (GEFs). The activated Rho GTPases then translocate to the plasma membrane, where they activate other downstream effectors such as PAK1 to modulate actin dynamics [19]. Indeed, neurotrophins have recently been observed to induce Rho GTPase activity through recruitment of a number of GEFs. TrkA was demonstrated to bind to Kalirin, an association that is essential for NGF-induced Rac1 activation and neurite outgrowth [20]. Furthermore, NGF treatment induces plasma membrane translocation of the GEFs Vav2 and Vav3, an event that is required for activation of Rac1 and Cdc42 and the induction of neurite outgrowth following NGF treatment in PC12 cells [21]. NGF also stimulates activation of the Rac-specific GEF p-Rex1 in PC12 cells [18]. Two recent studies reveal that neurotrophin stimulation in Schwann cells also leads to Rho GTPase activation through activation of GEFs. TrkC activation results in activation of the Cdc42-specific GEF Dbs [22] and Rac-specific GEF Tiam1 [23], both of which are required for NT-3-stimulated Schwann cell migration. Finally, TrkB was also recently demonstrated to bind and phosphorylate Tiam1 to mediate a BDNF-triggered change in cell shape [24].
On the other hand, recent studies accentuate the importance of membrane recruitment of Rho GTPase to lipid rafts for the function of these Rho GTPases. Lipid rafts are microdomains in plasma membrane rich in cholesterol and sphingolipids. Targeting of activated Rac1 to lipid rafts is required for activation of downstream effector Pak1 [25]. More importantly, neurotrophin-triggered Rac1 activation and morphological changes in hippocampal neurons have also been observed to require localization of Rac1 to lipid rafts [26]. Finally, BDNF has also been observed to increase Cdc42 activity in cerebellar granule neurons through enhancing calcium influx following the activation of PLCγ and PI3K pathways, a series of events that are essential for BDNF-mediated growth cone turning [27]. While a number of mechanisms have been postulated to underlie neurotrophin-mediated activation of Rho GTPases, it appears that the mechanisms implicated may vary with different downstream functions of Trk activation and the GEF involved.
In the current study, we demonstrated that Ser478 phosphorylation of TrkB by Cdk5 is essential for the Cdc42-dependent increase in primary dendrites triggered by BDNF, thus adding a new regulatory component to the mechanisms involved in Rho GTPase activation by neurotrophin. Although the precise downstream pathways by which this phosphorylation affects Cdc42 activation remains to be determined, our observations provide some interesting insights. First of all, while inhibition of Cdk5-mediated TrkB phosphorylation at Ser478 essentially abolished BDNF-induced increase in primary dendrites, it was surprising to observe that Cdk5 activity had a negligible effect on TrkB activation and initiation of downstream signaling pathways. This suggests that Cdk5 activity probably did not affect BDNF-dependent activation of Cdc42 and the induction of primary dendrites through modulating activation of downstream signaling. This is unexpected because BDNF-stimulated increase in primary dendrites was previously observed to depend on PI3K/Akt pathways in cortical neurons [28]. Nonetheless, accumulating evidence reveals that the location at which Trk receptors are activated may play a pivotal role in determining the precise downstream significance of Trk activation. For example, BDNF-induced increase in primary dendrites was recently demonstrated to involve TrkB activation in the lipid rafts [13]. In addition, retrograde transport of activated Trk receptors as signaling endosomes is emerging as a key regulator of neuronal survival [29]. Since we examined changes in TrkB downstream signaling cascades only in total lysates, it remains possible that Cdk5 activity may specifically affect TrkB signaling only at certain subcellular/plasma membrane compartments.
Secondly, overexpression of CA Cdc42 restored BDNF-induced dendritic growth in cdk5−/− neurons and in neurons overexpressing TrkB M1 (Figure 6), suggesting that maintenance of Cdc42 activation was sufficient to overcome the lack of BDNF-stimulated dendritic growth when Cdk5-mediated TrkB phosphorylation was absent. It thus appears that Cdk5 may impair BDNF-induced Cdc42 activation by affecting activation of the Rho GTPase. On the other hand, it should also be noted that overexpression of both the DN and CA forms of Cdc42 had a negligible effect on the basal number of primary dendrites in both cdk5+/+ and cdk5−/− neurons (Figure 6). Our observation is in agreement with an earlier study demonstrating that overexpression of DN or CA Cdc42 had no effect on the number of primary dendrites in chick spinal neurons [30]. In addition, it is consistent with the observation that modulation of Cdk5 activity or overexpression of TrkB M1 affected only BDNF-induced dendritic growth, without affecting the basal number of dendrites. Nonetheless, the inability of CA Cdc42 to mimic BDNF in the induction of primary dendrites suggests that activation of Cdc42 per se was insufficient to trigger dendritic growth in the absence of BDNF, and that additional, BDNF-dependent event(s) are required for the induction of dendritic growth by BDNF. Although the precise pathways implicated remain to be identified, it is tempting, in light of the emerging importance of lipid rafts in the activation of Rho GTPase, to speculate that BDNF may be required to stimulate translocation of activated Cdc42 to lipid rafts. In support of this hypothesis, it was observed that activation of Rac1 depends on the translocation of the activated Rho GTPase to lipid rafts [25,26]. In addition, in the absence of cholesterol, CA Rac1 failed to translocate to plasma membrane in fibroblasts [25]. More importantly, depletion of cholesterol similarly abolished BDNF-induced increase in primary dendrites in hippocampal neurons [13]. These observations collectively suggest that the inability of CA Cdc42 to increase dendritic growth in the absence of BDNF treatment may be related to the lack of CA Cdc42 translocation to lipid rafts, which may potentially be induced by BDNF treatment. A thorough investigation of the importance of lipid rafts in Cdc42 activation and primary dendrite induction by BDNF will shed light on the mechanisms by which BDNF-triggered dendritic growth is regulated.
Given the near absence of Ser478-phosphorylated TrkB in cdk5−/− brain, we believe that Cdk5 functions as the predominant kinase for this phosphorylation in vivo. Nonetheless, it was interesting to note that prior to BDNF stimulation, a basal level of Ser478-phosphorylated TrkB was detected in cortical neurons that was not inhibited by pretreatment with the Cdk5 inhibitor Ros. This may suggest that other serine kinases are present to phosphorylate TrkB at Ser478 in the absence of BDNF stimulation. Nonetheless, given the marked inhibition of BDNF-stimulated increase in TrkB phosphorylation by Ros, we believe that Cdk5 is essential for the BDNF-dependent component of TrkB phosphorylation at Ser478.
Given the abundant expression of Cdk5 and TrkB in neurons throughout development, and their respective concentration at the synapse, it would be interesting to examine if Cdk5 activity is also involved in other downstream functions of TrkB signaling, such as the regulation of neuronal survival and synaptic plasticity. Preliminary findings from our laboratory reveal that Cdk5 activity is also required for BDNF-stimulated neuronal survival in cortical neurons (unpublished data). In addition, the juxtamembrane region of Trk receptors has been associated with the regulation of Trk receptor internalization [31] and degradation [32]. Further investigation of whether this phosphorylation also affects the internalization and degradation of the receptor would provide further insights into the biological significance of this phosphorylation. In addition, since Cdk5 was observed to associate with TrkA without phosphorylating the receptor, further delineation of the consequences of this interaction would be essential for thoroughly understanding the crosstalk between Trk receptors and Cdk5. A preliminary study revealed that, similar to TrkB, TrkA phosphorylates Cdk5 at Tyr15 (unpublished data). The differential interaction of TrkA and TrkB with Cdk5, together with the differential localization of TrkA and TrkB in different neuronal populations, may provide a novel mechanism by which Cdk5 can regulate the signaling of different neuronal populations. In conclusion, our findings have provided evidence for a regulatory role of Cdk5 in Trk-induced dendritic growth, and lend support for an emerging role of Cdk5 as a regulator of RTK signaling. Given the importance of neurotrophin/Trk signaling in almost all aspects of neuronal development and function, our findings will likely have far-reaching implications for further elucidating the signaling mechanisms involved in the regulation of neuronal survival, synapse formation, and synaptic plasticity.
The antibodies against Trk (C-14), Cdk5 (DC-17), p35, and Shc were purchased from Santa Cruz Biotechnology (http://www.scbt.com). The antibodies against TrkB and SH2B were from BD Biosciences (http://www.bdbiosciences.com). The polyclonal antibodies recognizing phospho-TrkA (Tyr490), p44/42 mitogen-activated protein kinase (Erk1/2), phospho-p44/42 mitogen-activated protein kinase, AKT, phospho-AKT (Ser473), CREB, and phospho-Ser133 CREB were obtained from Cell Signaling Technology (http://www.cellsignal.com). Antibodies specific for actin and β-tubulin type III were from Sigma-Aldrich (http://www.sigmaaldrich.com). Antibody against the p-Ser478 of TrkB was raised by synthetic peptide (CISNDDDSApSPLHHIS; Bio-Synthesis, http://www.biosyn.com) and purified using AminoLink Kit (Pierce, http://www.piercenet.com).
Expression vectors of p35, Cdk5, and DN Cdk5 were prepared as previously described [3]. Flag-tagged and GST-tagged Cdk5 were generated by PCR, and subcloned into the mammalian expression vectors pcDNA3 (Invitrogen, http://www.invitrogen.com) and pGEX-6P-1 (Amersham Biosciences, http://www5.amershambiosciences.com), respectively. HA-tagged and GST-tagged Rac1, Cdc42, and RhoA constructs were gifts from Yung-Hou Wong (Hong Kong University of Science and Technology, Hong Kong). The expression vectors of TrkA, TrkB, and TrkC were constructed as described [33]. Three TrkB mutants lacking the potential Cdk5 phosphorylation sites were constructed by mutating Ser478 (TrkB M1), Thr489 (TrkB M2), or both Ser478 and Thr489 (TrkB DM) to alanine using the overlapping PCR technique, followed by subcloning into pcDNA3. GST-TrkB-Juxta construct was generated by PCR and subcloned into pGEX-6P-1. Protein purification was performed according to the manufacturer's protocol.
Stealth RNAi molecules for Cdk5 were prepared as previously described [34]. The sequences used were: Cdk5 siRNA, CCUCCGGGAGAUCUGUCUACUCAAA; and control siRNA (Cdk5), CCUAGGGCUAGCUGUUCAUCCCAAA.
Cdk5 and p35 knockout mice were kindly provided by A. B. Kulkarni (National Institutes of Health, Bethesda, Maryland) and T. Curran (St. Jude Children's Research Hospital, Memphis, Tennessee), and L. H. Tsai (Harvard Medical School, Boston, Massachusetts), respectively. Mice from different stages were collected and genotyped as described [7,35].
Rat cortical and hippocampal neuron cultures were prepared as previously described [33,34]. Subsequent to digestion with 0.25% trypsin in Hank's Balanced Salt Solution without Ca2+ and Mg2+ at 37 °C for 5 min, the reaction was stopped by 2.5% heat-inactivated horse serum. The dissociated neurons were seeded in culture dishes coated with 10 μg/ml poly-D-lysine. Two hours later the medium was replaced by neurobasal medium supplemented with 2 mM L-glutamine and 2% B27 supplement.
Selective Cdk5 inhibitor Ros (Calbiochem, http://www.merckbiosciences.com/html/CBC/home.html) was used to inhibit Cdk5 activity in primary neuron cultures. Primary cultures at 3 d in vitro (DIV3) were treated with or without BDNF (50 ng/ml) in the presence of Ros (10 or 25 μM) or DMSO for 3 d before harvesting or fixation.
For transfection of primary cultures, cortical and hippocampal neurons were seeded on coverslips in 12-well dishes at a cell density of 2 × 105 per coverslip. Neurons were transfected using calcium phosphate precipitation at DIV3. Twenty-four hours after transfection, the cultures were treated with BDNF for 3 d.
Primary hippocampal neuron cultures on coverslips in 12-well dishes were seeded at a cell density of 5 × 104 per coverslip for siRNA transfection. Cultures were transfected at DIV3 with Lipofectamine 2000 transfection reagent following the manufacturer's protocols (Invitrogen). The transfected cells were incubated at 37 °C with 5% CO2 for 24 h before treatment, and were then treated with BDNF for 3 d.
COS7 cells and HEK293T cells were obtained from American Type Culture Collection (http://www.atcc.org). Both cells were maintained in DMEM supplemented with 10% heat-inactivated fetal bovine serum, penicillin (50 units/ml), and streptomycin (100 μg/ml) at 37 °C with 5% CO2. COS7 cells and HEK293T cells were transfected using Lipofectamine Plus transfection reagents following the supplier's instructions (Invitrogen). The cells were treated and harvested 24 h after transfection.
Cells were lysed at 4 °C for 30 min in lysis buffer (RIPA: 1× PBS, 1% NP40, 0.1% SDS, and 0.5% sodium deoxycholate) with various protease inhibitors (1 mM phenylmethylsulfonyl fluoride [PMSF], 1 mM sodium orthovanadate [NaOV], 2 μg/ml antipain, 10 μg/ml leupeptin, 30 nM okadaic acid, 5 mM benzamidine, and 10 μg/ml aprotinin). Brain tissues were homogenized in lysis buffer (0.5% NP-40, 20 mM Tris-HCl, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, and 1 mM NaF [pH 7.5]) supplemented with various protease inhibitors (1 mM PMSF, 1 mM NaOV, 2 μg/ml antipain, 10 μg/ml leupeptin, 30 nM okadaic acid, 5 mM benzamidine, and 10 μg/ml aprotinin). Proteins were resolved by SDS-PAGE and subsequently electro-transferred onto a nitrocellulose membrane. Immmunoblots were probed with the desired primary antibodies at 4 °C overnight. After washing with TBS-T, the corresponding HRP-conjugated secondary antibody was added and incubated for 2 h at room temperature. Proteins were then visualized using enhanced chemiluminescence Western blotting detection reagents with reference to the supplier's instructions (Amersham Biosciences).
For immunoprecipitation, 1–2 mg of protein lysates was incubated with 1 μg of the corresponding antibody at 4 °C overnight with rotation. Forty microliters of protein G Sepharose (Amersham Biosciences) pre-washed with 1× PBS was added and rotated at 4 °C for 1 h. After intense washing with the lysis buffer, the immunoprecipitated protein and its associated proteins were analyzed by SDS-PAGE and Western blotting.
Flag-tagged protein was overexpressed in COS7 cells and the cell lysate was obtained as described above. The cell lysate obtained was incubated with anti-Flag M2 affinity gel (Sigma-Aldrich) at 4 °C overnight with rotation. The Flag-tagged protein was pulled down by the affinity gel, and the affinity gel was washed twice with lysis buffer (0.5% NP-40, 20 mM Tris-HCl, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, and 1 mM NaF [pH 7.5]) with various protease inhibitors (1 mM PMSF, 1 mM NaOV, 2 μg/ml antipain, 10 μg/ml leupeptin, 30 nM okadaic acid, 5 mM benzamidine, and 10 μg/ml aprotinin). One to two milligrams of proteins prepared from brain tissues was incubated with the affinity gel, and the Flag-tagged protein pulled down by the affinity gel for 1 h. The affinity gel was washed twice with lysis buffer supplemented with protease inhibitors. The proteins pulled down by the Flag-tagged protein were subjected to Western blot analysis.
Recombinant Cdk5/p35 and Cdk5/p25 were kindly provided by Shin-Ichi Hisanaga (Tokyo Metropolitan University, Tokyo). TrkA, TrkB, and TrkC were immunoprecipitated from transfected HEK293T cells, and used as substrates for reconstituted Cdk5/p35 or Cdk5/p25 in the in vitro kinase assay. The kinase assay was performed at 30 °C for 30 min in kinase buffer containing 100 μM [γ-32P] ATP as described [36]. To examine if TrkB phosphorylated Cdk5, recombinant TrkB kinase domain (Upstate Biotechnology, http://www.upstate.com) was incubated with GST-Cdk5 for 30 min at 30 °C, with or without Trk inhibitor K252a pretreatment (100 nM) for 10 min, in the presence of 100 μM [γ-32P] ATP or cold ATP. To examine if BDNF stimulated Cdk5 activity, primary cortical neurons were treated with BDNF with or without 30 min of K252a pretreatment (100 nM). The immunoprecipitated Cdk5/p35 complexes from the lysates were washed three times with lysis buffer and twice with kinase buffer. The in vitro kinase reaction was performed at 30 °C for 30 min with kinase buffer containing 100 μM histone H1 peptide and 100 μM [γ-32P] ATP as described [37]. The phosphorylated proteins were resolved by SDS-PAGE. After the gel was dried, the phosphorylated proteins were visualized by autoradiography. For TrkB-mediated Cdk5 phosphorylation, the phosphorylated protein was resolved by SDS-PAGE, and blotted with phospho-Cdk2 (Tyr15; Santa Cruz Biotechnology) or phosphotyrosine antibody (4G10; Upstate Biotechnology).
GTPase activity was measured as described [38]. Briefly, cultured cortical neurons at DIV7 were pretreated with DMSO or Ros for 30 min, followed by treatment with BDNF for another 5 min. Cells were lysed at 4 °C and incubated with Pak1-PBD agarose with constant rocking at 4 °C for 1 h. The proteins bound to the beads were washed three times with lysis buffer at 4 °C, eluted in SDS sample buffer, and analyzed for bound Cdc42 by Western blotting using monoclonal antibody against Cdc42 (Upstate Biotechnology). GTPase activity was quantified by densitometry analysis of the blots.
Following fixation in 4% paraformaldehyde and 5% sucrose in PBS with Ca2+ and Mg2+ for 30 min, the cells were washed three times with PBS, and were blocked with 1% bovine serum albumin and 10% goat serum for 20 min. The cells were then incubated with the corresponding primary antibody (1:150–500) at 4 °C overnight, and were subsequently washed with PBS three times. Following incubation with FITC or rhodamine conjugated secondary antibody (1:1,000) for 1 h at room temperature, the cells were washed again, stained with DAPI, and mounted with coverslips and MOWIOL (Calbiochem). Mounted cells were visualized under fluorescent microscope (Leica, http://www.leica.com).
All data were expressed as mean ± standard deviation. Statistical significance was determined by one-way analysis of variance followed by Bonferroni's post hoc test with 95% confidence. A p-value of smaller than 0.05 was considered as statistically significant.
|
10.1371/journal.ppat.1005804 | An Epithelial Integrin Regulates the Amplitude of Protective Lung Interferon Responses against Multiple Respiratory Pathogens | The healthy lung maintains a steady state of immune readiness to rapidly respond to injury from invaders. Integrins are important for setting the parameters of this resting state, particularly the epithelial-restricted αVβ6 integrin, which is upregulated during injury. Once expressed, αVβ6 moderates acute lung injury (ALI) through as yet undefined molecular mechanisms. We show that the upregulation of β6 during influenza infection is involved in disease pathogenesis. β6-deficient mice (β6 KO) have increased survival during influenza infection likely due to the limited viral spread into the alveolar spaces leading to reduced ALI. Although the β6 KO have morphologically normal lungs, they harbor constitutively activated lung CD11b+ alveolar macrophages (AM) and elevated type I IFN signaling activity, which we traced to the loss of β6-activated transforming growth factor-β (TGF-β). Administration of exogenous TGF-β to β6 KO mice leads to reduced numbers of CD11b+ AMs, decreased type I IFN signaling activity and loss of the protective phenotype during influenza infection. Protection extended to other respiratory pathogens such as Sendai virus and bacterial pneumonia. Our studies demonstrate that the loss of one epithelial protein, αVβ6 integrin, can alter the lung microenvironment during both homeostasis and respiratory infection leading to reduced lung injury and improved survival.
| The lung undergoes daily assault by microbes and other inhaled particulates and must maintain the balance between clearance of harmful microorganisms while protecting the delicate lung structure to avoid acute lung injury. Not surprisingly, this is a complex process requiring communication between the lung epithelial cells (first site of attack by invaders) and the cells of the intrinsic immune response. We demonstrate the αVβ6 integrin is an important player in this interface. Loss of αVβ6 during influenza infection, and other respiratory infections, leads to bolstered protection against severe lung disease and improved survival in mice. Even in the absence of infection, β6 KO animals have a distinct anti-microbial lung microenvironment as evidenced by increased type I IFN activity and activated alveolar macrophages that appear “poised to defend”. These studies explore how the epithelial-specific αVβ6 integrin regulates the lung microenvironment to alter alveolar macrophage activity through a TGF-β-dependent mechanism, leading to changes in both the homeostatic lung and responses to respiratory infections as a way of balancing microbial clearance with protection of the lung from excessive damage.
| At each breath, the lung is challenged by a large number and diversity of microbes and other foreign material such as pollen and dust. Many inhaled microbes cause lethal infections if not contained by the lung immune system, which has evolved to balance rapid and efficient microbial clearance with protection of the delicate lung structure from excessive damage. Lung damage caused by microbial pathogens is the cause of acute lung injury (ALI), which leads to increased edema, alveolar permeability, and impaired oxygen exchange. In severe cases, ALI can result in impaired gas exchange function and ultimately death (acute respiratory distress syndrome or ARDS) [1]. To mitigate lung damage after infection, inflammation resolves, returning to homeostasis that restores normal lung function [2,3]. The air-interface structure of the lung involves a complex immune cell population including resident interstitial macrophages and dendritic cells, and GM-CSF-dependent alveolar macrophages [4–7]. A key question in understanding pulmonary immunity concerns how the balance between effective immune surveillance and maintenance of lung anatomy and physiology is achieved through life.
Integrins are heterodimers composed of α and β subunits that regulate a plethora of cellular functions including cell-matrix and cell-cell adhesion, cell activation, and the recognition and post-translational processing of molecules [8]. In the lung microenvironment, β6 (encoded by Itgb6) exclusively pairs with the αV subunit and expression of the αVβ6 heterodimer is limited to epithelial cells. In this context, αVβ6 plays a key role in controlling several steps in lung homeostasis. Mice lacking the β6 subunit (β6 KO) have highlighted the central role αVβ6 integrin plays in balancing the pulmonary environment during injury [9,10]. For example, in both bleomycin-induced and PAR1-mediated ventilator-associated lung injury and airway hyper-responsiveness models, β6 KO mice were protected from ALI [11–13]. β6 KO mice also had reduced IL-1β-mediated lung injury in a model of bronchopulmonary dysplasia due to decreased inflammation and cellular infiltrate to the lung [14]. Alveolar macrophages isolated from β6 KO are characterized as large and foamy and express high amounts of MMP12 [15,16].
Severe respiratory infections are often associated with ALI/ARDS [17–19], and the pathophysiology is similar for a variety of infectious pathogens. For example, patients with SARS and MERS-CoV were found to have increased infiltration of inflammatory macrophages, accumulation of debris in the lung, and diffuse alveolar damage [20,21]. Similarly, elderly patients suffering from RSV sustained widespread alveolar damage [22]. Bacterial pneumonia can result in increased formation of hyaline membranes [23]. Deposition of collagen and fibrin, while necessary for lung repair, can also build up and block gas exchange [24]. In the case of influenza, the major complication leading to ALI/ARDS is pulmonary edema and impaired fluid clearance due to dysregulation of transporters that clear fluid from the alveolus such as ENaC and Na+K+ATPase [23,25,26]. The virus itself can also directly kill epithelial cells [27]. ALI/ARDS has a mortality rate of 30–50%, therefore it is critical from a public health perspective to understand the mechanisms at play [28,29].
Given the role of the β6 integrin in modulating ALI, we hypothesized that its upregulation during respiratory infections such as influenza would be important for viral pathogenesis. To test this, β6 KO mice were infected with the 2009 pandemic H1N1 influenza virus. β6-deficient mice were protected from influenza virus-induced disease as well as a broad range of respiratory pathogens, in most cases independent of effects on overall microbial numbers. Mechanistically, we found that epithelial β6 controls the homeostatic lung interferon response. In the absence of β6, type I interferon signaling is constitutive, causing the host to have an advantage over the spread of the virus. This protective phenotype was reversed by exogenous TGF-β1 or elimination of the type I interferon receptor, suggesting that αVβ6 controls a communication system between lung epithelia and immune cells through a TGF-β-dependent mechanism. These studies have important implications as transient inhibition of αVβ6 may represent a potential therapy for the management of acute lung injury.
β6 expression is induced upon mechanical or inflammatory injury and is an important mediator of ALI [10,12,14,30]. Thus, we hypothesized that upregulation of the β6 integrin during respiratory infections would be involved in viral pathogenesis. To test this, we measured β6 expression by quantitative RT-PCR at different times post-influenza infection. Consistent with lung injury models, β6 mRNA was significantly upregulated in whole lung homogenates by 3 days post-infection (dpi) and expression remained elevated through 5 dpi (p < 0.0001) correlating with the appearance of ALI (Fig 1A). We then intranasally infected WT and Itgb6-deficient mice (β6 KO mice) with A/California/04/2009 (CA/09) H1N1 virus and monitored morbidity for 12 dpi. Compared to WT controls, β6 KO mice lost significantly less weight (4, 6, and 8 dpi p < 0.0001) and began to recover by 10 dpi, while WT mice lost weight and either succumbed to infection or were euthanized due to morbidity between 6–10 dpi (Fig 1B and 1C). The dose of CA/09 virus used (104 TCID50) was lethal to all WT mice, while 70% β6 KO mice survived the infection (p = 0.0090). Mouse lethal dose 50 (MLD50) studies demonstrated the resistance of the β6 KO to influenza infection, with WT mice having an MLD50 of 102.5 versus 105.3 for the KO to the CA/09 virus. Protection was not limited to H1N1 infection; β6 KO mice were also protected from an emerging strain of avian influenza virus associated with severe and even fatal human respiratory disease [17], A/Anhui/1/2013 (H7N9) influenza virus (p = 0.0326, Fig 1D). β6 KO mice were also significantly protected from a lethal challenge of Sendai virus (p = 0.006, Fig 1E) and Streptococcus pneumonia (p = 0.0082, Fig 1F).
A major risk factor for influenza infection is secondary bacterial pneumonia [31]. As β6 KO mice were protected from individual influenza and S. pneumoniae challenges, we tested the β6 KO mice with a secondary bacterial challenge model. Mice were inoculated intranasally with a sublethal dose of A/Puerto Rico/8/34 H1N1 influenza virus, and 7 dpi administered a low dose of S. pneumoniae (D39X strain) [32]. This dual infection model is generally lethal in WT mice, although neither challenge alone causes death [33]. Survival of WT mice after secondary challenge was ~30%, while β6 KO mice were significantly protected (p = 0.0226) with 70% survival (Fig 1G) highlighting the importance of β6 integrin in the pathogenesis of respiratory infections.
To determine if the enhanced survival seen in the β6 KO mice was associated with decreased ALI, we examined histological sections for evidence of tissue injury, alterations of the alveolar capillary barrier, and inflammatory responses [1]. Although groups were histologically similar at 3 dpi, at 7 dpi inflammation and thickened septa involving extensive areas of alveolar parenchyma was seen in WT lungs. In marked contrast, evidence of inflammation and tissue damage in the β6 KO lungs was generally limited to terminal bronchioles and small numbers of adjacent alveoli (Fig 2A). Infected WT mice had increased TNF-α and IL-6 throughout the duration of infection, while the inflammatory cytokine response in β6 KO mice was less robust (Fig 2B). Evidence of alterations to the alveolar capillary barrier was found in infected WT mice beginning at 5 dpi as compared to KO mice (Fig 2C–2F) including heavier lungs as determined by wet/dry weight ratios (5 dpi p = 0.0077, 7 dpi p = 0.0063) indicating increased edema (Fig 2C), accompanied by increased total protein (p < 0.0001, Fig 2D) and albumin in the bronchoalveolar lavage fluid (BALF) (7 dpi p < 0.0001, Fig 2E) and accumulation of Evans blue dye in the lungs of WT mice (p = 0.0071, Fig 2F). Overall, β6 KO had decreased inflammation, acute lung injury, and improved survival during influenza infection.
The most likely mechanism of protection in influenza infected β6 KO mice was decreased viral titers. However, tissue culture infectious dose 50 (TCID50) and quantitative real-time PCR showed no significant differences in viral titers, although RT-PCR suggested a trend towards lower titers in β6 KO mice (Fig 3A and 3B). We therefore spatially monitored viral spread in the lungs by immunohistochemistry for viral nucleoprotein (NP) and use of a reporter virus [34,35]. Viral spread into the alveolar spaces was reduced in β6 KO mice in comparison to WT mice, which had extensive NP staining of type II pneumocytes and macrophages in the alveolar spaces by 5 dpi. In contrast, NP staining in β6 KO mice was largely restricted to the terminal bronchiolar epithelium and adjacent alveoli (Fig 3C). This is even more evident using our NLuc CA/09 reporter virus, which clearly demonstrates decreased virus in β6 KO ([34,35], S1 Fig). Finally, WT lungs had a higher percentage of sites of active infection at 5 and 7 dpi as well as more viral antigen detected in the alveolar spaces (Fig 3D and 3E). Collectively, viral spread in KO lungs was limited compared to WT mice, with less extensive involvement of alveolar epithelial cells possibly explaining the improved lung function relative to controls. These findings are consistent with studies suggesting that increased survival during influenza infection can be independent of changes in viral titers [36–40].
To understand the underlying mechanism(s) for the decreased viral spread within the lungs of β6 KO mice, we performed flow cytometry on BALF and lungs at different times post-infection. No significant differences were noted in numbers of TNF-α/iNOS-producing (tip)DCs, neutrophils, CD4+ T cells, CD8+ T cells, or influenza PB1-specific CD8+ T cells in either BALF (Fig 4A–4E) or whole lung (Fig 4F–4J). However, we noted substantial differences in the resident lung F480+ CD11c+ CD11b+ populations (macrophages and dendritic cells), namely that β6 KO mice lacked a conventional CD11c+ alveolar macrophage (AM) population (Fig 5A). AMs have characteristic flow cytometry properties as compared to dendritic cells (DCs) or incoming activated bone marrow-derived inflammatory monocytes including autofluoresence in the FITC channel, high constitutive expression of CD11c, F4/80 and SiglecF and low expression of CD11b [4,41,42]. By contrast, incoming activated bone marrow-derived inflammatory monocytes express CD11b, providing a means to distinguish between resident AM, inflammatory bone marrow-derived monocytes and DCs [4,43]. When we examined the lung macrophage/DC populations in uninfected β6 KO mice we found they had higher numbers of autofluorescent cells indicating AMs (Fig 5B) but instead of the normal CD11chi CD11b- phenotype, the vast majority of the β6 KO AM were MerTK+, CD64+, CD11c+, and uniformly CD11b+ (Fig 5C; See S2 and S3 Figs for gating strategies). These cells maintained expression of F4/80 and SiglecF, suggesting they are AMs with altered properties. They were also phenotypically distinct from those of WT mice, with a ‘foamy’ appearance typical of activated macrophages (Fig 5D). Importantly, the β6 KO CD11c+ CD11b+ AM were present at the earliest times tested (d3 after birth), continued to be present at all points of adulthood (S4A Fig), and were not affected by reducing neonatal exposure to airborne particulate matter by using HEPA-filtered cages or low dust cage bedding (S4B Fig). Even throughout the course of influenza infection very few CD11chi CD11b- cells were present in the lungs of β6 KO mice (Fig 5E). As the lungs from β6 KO were overtly normal, we concluded the CD11c+ CD11b+ cells were sufficient to perform the normal activities of lung homeostasis [12,14,16,30].
Most pulmonary AM originate from embryonic erythro-myeloid progenitors whose replacement by blood monocytes under steady state is extremely slow [42,44]. To test whether the β6 KO CD11c+ CD11b+ AM were predominantly derived from embryonic progenitors, we created mice where the blood monocyte pool was severely depleted by loss of the chemokine receptor CCR2, which is the predominant mediator for monocyte egress from the bone marrow. In the lungs of β6/CCR2 doubly-deficient mice, CD11c+ CD11b+ AM were present in identical amounts compared with those of β6 KO mice, arguing these cells are unlikely to be derived from peripheral blood monocytes (S4C Fig). Although other unknown pathways to recruit monocytes to the lung cannot be discounted, they are not as well characterized or understood as the CCR2 pathway. Taken together, we concluded loss of an epithelial integrin alters the phenotype of lung AMs.
Next we tested whether the β6 KO CD11c+ CD11b+ AM were a consequence of exposure to an altered lung microenvironment caused by loss of β6, or were caused by a cell-intrinsic mechanism. Adoptive transfer of congenically marked WT AM into β6 KO animals led to the CD11c+ CD11b+ AM phenotype after 7 days while transfer of β6 KO AMs into WT animals results in a WT CD11chi CD11b- phenotype (Fig 6A–6C). Importantly, the effect of the β6 KO lung microenvironment was not unique to transferred AM. Adoptive transfer of peritoneal resident macrophages into β6 KO animals resulted in conversion into the CD11c+ CD11b+ AM phenotype (Fig 6D and 6E) highlighting that it is the altered lung microenvironment of the β6 KO that controls AM phenotype. Donor cell recovery was consistent for all experiments (S5 Fig).
What could be different about the β6 KO lung microenvironment? β6 integrin is the primary way to endogenously activate transforming growth factor-β (TGF-β) in the lung, which is known to restrain AM activity [16,45]. Thus, we next asked if the altered macrophage phenotype in the β6 KO mice was TGF-β1 dependent. As expected, endogenous lung TGF-β1 activity was reduced in β6 KO mice as compared to WT (Fig 7A). To test the hypothesis that reduced TGF-β1 activity contributed to the altered lung microenvironment of β6 KO mice, we intranasally administered exogenous TGF-β1 to WT and β6 KO mice every other day for 3 weeks (Fig 7B) as described [46]. β6 KO mice treated with TGF-β1 had increased numbers of CD11c+CD11b- macrophages (Fig 7C, quantification Fig 7D), partially reversing the CD11c+CD11b+ phenotype, and were no longer protected from influenza infection (Fig 7E) suggesting that the loss of β6-mediated TGF-β1activation is important for both protection and the ‘activated’ macrophage phenotype.
β6 KO mice had an altered lung environment resulting in decreased viral spread. Type I IFNs are amongst the most potent cytokines limiting viral spread. Thus, we tested the hypothesis that β6 KO mice had increased type I IFN than WT. We used phosphorylated STAT1 levels (pSTAT1) as a marker for type I IFN signaling. Increased pSTAT1 along with increased type I IFN gene expression and unchanged expression of type II IFN-dependent targets such as IDO1 are indicative of a type I IFN dominated environment. Further, STAT1 is a robust marker of type I IFN signaling due to its weak activation by other cytokines such as IL-6 [47]. Immunoblot analysis of whole lung homogenates showed the β6 KO mice had robust type I IFN signaling at baseline that increased 3–5 dpi (Fig 8A). We sought to determine the source of type I IFNs and how epithelial-expressed β6 regulated the IFN response. We noted that type I IFN mRNAs were not substantially altered in uninfected β6 KO mice (S6A–S6D Figs). Furthermore, ELISA measurements for type I IFNs did not show substantial differences between WT and β6 KO mice within whole lung homogenates. Therefore, we generated β6 KO mice crossed to an IFN-β-YFP reporter strain. In these mice, YFP expression in the whole lung was predominantly found in CD45+ cells and increased in the absence of β6 (Fig 8B); although immunofluorescent microscopy of frozen lung sections from WT and β6 KO YFP reporter mice showed increased YFP+ cells scattered throughout the lung; including in the epithelia (Fig 8C and 8D). Further delineation of the exact cell types in this system expressing YFP is precluded by the strong autofluorescence in the FITC channel in lung macrophages [4]. Thus, we examined isolated AMs and primary epithelial cells grown at the air-liquid interface from β6 KO and WT mice for increased type I IFN signaling.
Sorting autofluorescent CD11c+CD11b- AM from uninfected WT mice and comparing them by microarray to autofluorescent CD11c+CD11b+ ‘AM’ from β6 KO demonstrated that β6 KO AM showed a striking enrichment of gene ontogeny and pathway terms linked to immune responses (S7 Fig), particularly genes associated with type I IFN responses (Fig 8E, Table 1). There was no difference in IDO1 expression, a cardinal IFN-γ regulated gene [48] (Table 1). These data suggest lung macrophages in the absence of β6 had increased IFN signaling. To test this further, we measured phosphorylation of IRF3 and STAT1, hallmarks of the type I IFN response, in the alveolar macrophage populations from WT and β6 KO mice. CD11c+ CD11b+ macrophages from uninfected β6 KO mice had increased phospho-IRF3 and phospho-STAT1 relative to WT CD11c+CD11b- AM (Fig 8F).
To determine if type I IFN signaling was also increased in epithelial cells, primary bronchotracheal epithelial cells (mTEC) from WT and β6 KO mice were isolated, differentiated and grown at the air-liquid interface prior to infection with CA/09 virus (MOI 0.1). Phosphorylated and total levels of STAT1 and IRF3 were quantitated at 24 hours post-infection by immunoblot (Fig 8H). β6 KO mTEC exhibited significant increases in total STAT1 and IRF3 levels as compared to WT mTEC after infection as well as at baseline, although differences in phosphorylated STAT1 and IRF3 were minimal (Fig 8H and S6E Fig). Based on these data, we concluded that regulation of type I IFNs was dysregulated in the lungs of β6 KO mice at baseline, and that hematopoietic cells as well as epithelial cells may be involved.
Finally, to determine if this is regulated by TGF-β1, lung homogenates from mice administered exogenous TGF-β1 were monitored for total and phosphorylated STAT1. Treatment decreased phospho-STAT1 in whole lung homogenate (Fig 8I). Macrophages isolated from TGF-β1-treated β6 KO mice also displayed significantly decreased mRNA encoding levels of the IFN associated transcripts Irf7, Ifit1, and Oas1g that were comparable to levels in WT macrophages (Fig 8G) highlighting that the loss of β6 activated TGF-β results in an altered lung macrophage population and increased type I IFN activation. We question whether this was the reason for the enhanced protection from influenza infection.
To determine if the increased type I IFN signaling in β6 KO mice was responsible for the enhanced protection, we crossed β6 KO with Ifnar-/- mice (β6/IFNAR double KO), which lack the type I IFN receptor and all type I IFN signaling, then challenged these mice and their cognate controls created from the same heterozygote crosses with 104 TCID50 of CA/09 virus. Mice were monitored for morbidity and mortality for 12 days. β6/IFNAR double KO mice did not survive infection (p = 0.0039 compared to β6 KO, Fig 9A) and exhibited similar mortality to WT mice. Further, imaging WT, β6 KO, IFNAR KO, and β6/IFNAR double KO mice infected with CA/09-NLuc reporter virus at 3 and 7 dpi showed that the ββ6 KO mice had reduced viral spread in the lungs as compared to WT controls (S1A and S1B Fig). In contrast, viral spread in β6/IFNAR double KO mice was comparable to WT and IFNAR KO mice, highlighting the importance of type I IFN signaling in limiting viral spread during infection.
However, the ‘activated’ CD11c+CD11b+ macrophage phenotype of the β6 KO crosses remained the same as β6/IFNAR double KO (Fig 9B) suggesting the protective anti-viral phenotype of the β6 KO is type I IFNAR-dependent, but another pathway regulates the phenotype of the resident macrophages in the β6 KO independent of type I IFNs.
We propose a model where during pulmonary homeostasis, β6 integrin, possibly by activation of endogenous TGF-β1, suppresses CD11b expression on alveolar macrophages and type I IFN signaling within the lung microenvironment. However, in the absence of β6 integrin, CD11b expression is increased as is type I IFN signaling inducing a ‘primed’ antiviral state in the absence of infection (Fig 10). Through the use of TGF-β1 ‘rescue’ experiments and β6 KO mice lacking type I IFN signaling, we were able to demonstrate the resistance of β6 KO mice to influenza depended on the loss of TGF-β1 and this elevated type I IFN signaling, while the phenotypic changes in lung macrophages were independent of the type I IFN pathway at least at the level of type I IFN signaling as shown in the β6/IFNAR double KO mice.
Overall, our data argue β6 expression leads to changes in the lung microenvironment that are mediated immediately after birth and include alterations in the AM population, and repressed type I IFN signaling in the lung most likely through a major contribution from TGF-β1. This pathway might lead to selective regulation of pathogen colonization or spread, depending on the pathogen’s sensitivity to the pre-existing interferon milieu. Two caveats to the model proposed in Fig 10 include (i) the fact that β6 may regulate factors other than TGF-β1 that cause the downstream effects, and (ii) that we cannot yet definitively distinguish whether β6 negatively regulates type I IFN signaling, type I IFN production or both pathways. In our hands, assays quantitating type I IFN amounts by ELISA or bead-based assays are variable, lacking sensitivity or reproducibility. Furthermore, the potential increase in type I receptor expression on β6 KO macrophages (Table 1) could cause increased uptake of type I IFNs and not accurately reflect the soluble amounts present in the lung at a given time in homeostasis. Thus, the identification of the point at which β6 (with or without TGF-β1) negatively regulates IFN signaling will require new tools and approaches investigating regulation in both macrophages and epithelial cells.
Mice lacking the β6 integrin are protected from disease caused by influenza infection. Compared to WT controls, β6 KO mice develop less severe ALI, as characterized by edema, inflammatory cytokine expression, and vascular permeability of the lung epithelium. β6 KO mice also have altered baseline homeostasis possessing high baseline CD11b+ CD11c+ macrophages as opposed to CD11b- CD11c+ alveolar macrophages. A reasonable interpretation of these data is that β6 KO mice are ‘primed’ for a more efficient response to respiratory pathogens, due to constitutively increased type I IFN signaling at baseline. Importantly, protection extended beyond pandemic H1N1 influenza virus strain to include other influenza virus strains, Sendai virus, and S. pneumoniae infection. However, protection did not extend to all infections tested. We challenged β6 KO mice with Francisella tularensis Schu4 strain using 25 live organisms via the intranasal route. Schu4-infected β6 KO mice had no differences in survival compared to controls consistent with experiments performed in the Ifnar-/- mice [49] (CB, personal communication) (S8A Fig). Although β6 KO mice were significantly protected from disease caused by H7N9 virus (Fig 1D) they were not protected from the highly pathogenic avian influenza (HPAI) A/Hong Kong/483/1997 H5N1 infection (HK/483, S8B Fig). HPAI H5N1 viruses are unique in their ability to rapidly spread beyond the respiratory tract and cause systemic, including neurological, disease [18] and productively replicate in alveolar macrophages [50]. It is possible that HPAI H5N1 viruses are more resistant to type I IFN, that mice succumb due to direct viral cytopathic effect, or complications of systemic infection. Further studies are needed to understand why β6 KO mice are not protected against HPAI H5N1 infection. Taken together, these data suggest β6-mediated modulation of IFN responses affects responses to multiple respiratory pathogens whereas protective responses depend on IFN signaling.
The airway epithelium is a complex barrier comprised of multiple cell types that acts as an interface between the external environment and the internal lung milieu. It serves three main functions in a healthy lung; providing a tight mechanical barrier that rapidly repairs upon insult, mediating innate immune activity to limit foreign antigen invasion, and initiating an inflammatory response through production of cytokines and chemokines [51]. A variety of experimental models and human studies demonstrate the host response to respiratory infection involves initiation, resolution, and restoration phases, all of which must be tightly regulated to prevent disease [52]. However, little is known about how interactions between epithelial cells, which are often the primary targets of respiratory pathogens, and alveolar macrophages, which play a critical role in controlling infections, and the regulation of these responses. Our data identify a pathway by which alveolar epithelial cells normally suppress the anti-microbial activity of alveolar macrophages through a pathway that may involve local activation of TGF-β and subsequent suppression of IFN signaling. In terms of homeostasis, little is known about how type I IFNs control the status of tissue microenvironments. Most work on type I IFN signaling has been associated with anti-viral responses and the regulation of inflammation over the first 18–24 months of life [16]. The implication of our findings is that the lung can tolerate a homeostatic increase in IFN signaling, which provides an advantage against viral spread, but likely comes with ‘costs’ that are not yet clear. We note a recent study has shown homeostatic IFN-β and IFNAR signaling in the brain has an essential role in suppressing neurodegeneration [53]. Therefore, a more comprehensive understanding of the hitherto unknown tissue-specific homeostatic roles of type I IFN signaling is warranted. During infection, although it would seem that upregulation of β6 would only have a negative impact on the host by limiting the anti-viral response, it could also be important for balancing the inflammatory response by suppressing activated macrophages through a TGF-β-dependent process [2,5,15,45], initiating wound repair, or by recruiting macrophages to the site of infection. Attempts to address the role of TGF-β in influenza pathogenesis by systemic inhibition resulted in lethal infection [54]. Thus, future studies will focus on inhibiting the β6 integrin at different times post-infection to understand its precise role in viral infection and its potential as a therapeutic target.
Why is the lung macrophage population different in the β6 KO mice? The αVβ6 integrin is expressed at low amounts in healthy adult tissue but is rapidly upregulated during development and in response to injury and inflammation [30]. Once expressed, β6 activates latent TGF-β1 by binding to the RGD motif in the latency-associated peptide (LAP) [10,55]. Upon activation, TGF-β1 regulates a variety of genes associated with the immune response and pulmonary fibrosis and is thought to negatively restrict inflammation in macrophages through an unknown mechanism [45]. However, systemic inhibition of all TGF-β isoforms by a broadly neutralizing antibody led to lethal influenza infection [54] suggesting that the role of TGF-β during infection may be complex and depend on both the cellular milieu and TGF-β isoform. Recent work has shown that the pulmonary microenvironment dictates the behavior of alveolar macrophages to a greater extent than the origin of the macrophages [42]. Based on our studies we propose that in the lung microenvironment, upregulation of the β6 integrin leads to localized activation of TGF-β1 that negatively regulates alveolar macrophages leading to decreased type I IFN activity in the microenvironment through an as yet undefined mechanism that will require further molecular investigation. An important extension of this model is that influenza viruses themselves activate latent TGF-β during infection [56] and global inhibition of TGF-β during infection is lethal, suggesting an important role in protection from influenza [54]. Therefore, it will be important to be able to experimentally segregate the host- versus virus-specific TGF-β processing events in mice expressing or lacking β6 integrin.
Overall, loss of the β6 integrin resulted in protection from a variety of respiratory infections including influenza and Sendai viruses, bacterial pneumonia, and viral-bacterial co-infections. β6 KO mice had significantly reduced extent and severity of lung injury and inflammation, reduced collagen deposition despite no difference in viral titer, consistent with studies using β6 integrin function blocking antibodies [57], with virus infection being more restricted to airways instead of infiltrating the deep-lung alveolar spaces. β6 function blocking antibodies have been successfully used to improve outcome in idiopathic pulmonary fibrosis and currently recruiting for phase II human clinical trials [58,59]. While the β6 integrin represents an attractive therapeutic target to combat pulmonary disease, it does play a role in lung homeostasis. While transient inhibition of β6 in the lung results in resistance to infection, complete lack of β6 function for long periods of time could be detrimental. Over long periods of time, β6 KO mice over produce macrophage-derived matrix metalloprotease MMP12, leading to emphysema and destruction of pulmonary tissue [16]. However, for short-term protection against the adverse consequences of potentially lethal pulmonary viral infections, our work identifies the αVβ6 integrin as a potentially attractive therapeutic target for transient intervention in lung viral infection.
All animal work was approved by St Jude Children’s Research Hospital Institutional Animal Care and Use committee (protocol #513). Fransicella challenges were conducted under protocols approved by the NIAID Rocky Mountain Laboratories Animal Care and Use Committee. St Jude is fully accredited by the Association for the Assessment and Accreditation of Laboratory Animal Care International (AAALAC-I) and has an approved Animal Welfare Assurance Statement with the Office of Laboratory Animal Welfare (A3077-01). These guidelines were established by the Institute of Laboratory Animal Resources and approved by the Governing Board of the U.S. National Research Council.
C57BL/6, Ifnb1-YFP, CCR2 KO and CD45.1 mice were obtained from Jackson Laboratories (Bar Harbor, ME). Itgb6-/- mice, generated as described [9] and backcrossed 10 generations onto the C57BL/6 background, were obtained from Dean Sheppard (UCSF) then bred at St Jude Children’s Research Hospital. Littermate controls from heterozygous crosses were used as controls. IFNAR KO mice were obtained from Dr. Laura Knoll (University of Wisconsin). Knockouts were confirmed by PCR using primer sets (CCR2, IFNAR) reported on the Jackson Laboratories website, or as previously described (β6) [16].
A/California/04/2009 (CA/09) H1N1, A/Puerto Rico/8/34 H1N1, A/Anhui/1/2013 H7N9 and A/Hong Kong/483/1991 H5N1 influenza viruses were propagated in 10-day old specific-pathogen-free embroynated chicken eggs and viral titers determined by tissue culture infectious dose 50 (TCID50) on Madin Darby canine kidney cells (MDCK; ATCC CCL-34, Manassas, VA) cells as described [50] and quantitated [60]. Enders strain Sendai virus was propagated as described [61]. D39X S. pneumoniae was grown in C + Y medium to log phase (O.D. 0.4), pelleted, and diluted to the appropriate concentration in phosphate buffered saline pH 7.4 and stored on ice until infection. MDCK cells were cultured in modified Eagle’s medium (MEM) (Corning, Manassas, VA) supplemented with 2 mM glutamine (Gibco, Grand Island, NY) and 10% fetal bovine serum (Atlanta Biologicals, Lawrenceville, GA) at 37°C, 5% CO2.
Deeply anesthetized mice were perfused with 10% neutral buffered formalin, tissues collected at the indicated time post-infection, and embedded in paraffin. H&E and immunohistochemical slides were blinded and scored by PV. Influenza NP staining was performed with anti-NP antibodies (05G, US Biological, Massachusetts, MA) and five separate lung sections quantitated using Aperio ePathology software (Buffalo Grove, IL).
RNA from whole lung homogenates was isolated using a Trizol (Invitrogen) per manufacturer’s protocol. cDNA was synthesized using the SuperScript VILO cDNA synthesis kit (Invitrogen) and the following parameters: 25°C for 10 min, 42°C for 1 hour, and 85°C for 5 min. PCR was run using 2 μl of cDNA to detect Itgb6, Irf7, Ifit1, and Oas1g mRNA using the Quantitect primer/probe assay system (Qiagen) per manufacturer’s protocol. Samples were normalized to GAPDH mRNA as the internal control.
Whole lungs were collected at different times post-infection and homogenized in RIPA buffer (50 mM TrisHCl pH 7.5, 150 mM NaCl, 0.1% SDS, 0.5% sodium deoxycholate, 1% TritonX-100) containing protease and phosphatase inhibitors (ThermoScientific, Rockford, IL). After centrifugation at 15000 rpm at 4°C, protein concentration was determined using the Bradford assay (Pierce, Rockford, IL) and equivalent amounts loaded onto 10% Tris-glycine SDS-PAGE gels (Lonza, Allendale, NJ). After transfer to nitrocellulose, membranes were probed for integrin β6 (0.236 μg/ml, ch2A1 antibody, a kind gift from Biogen Idec, Cambridge, MA), anti-STAT1 (E-23, Santa Cruz, Dallas, TX, sc-346) (0.1 μg/ml), anti-pSTAT1 (D4A7, Cell Signaling, Danvers, MA, 7649S) (1:1000 dilution), anti-IRF3 (D83B9, Cell Signaling, 4947S) (1:1000 dilution), or anti-pIRF3 (4D4G, Cell Signaling, 4302S) (1:1000 dilution) and detected using HRP conjugated secondary goat anti-rabbit or goat anti-mouse antibodies (0.1–0.2 μg/ml, Jackson) and visualized with ECL. Anti-β-actin (AC-15, Sigma, St Louis, MO, A5441) (1 μg/ml) or anti-GAPDH (mAbcam 9484, Abcam, Cambridge, MA, ab9484) (1 μg/ml) served as loading controls.
BALF or whole lung homogenate was collected from PBS control and influenza inoculated mice at 2, 5, 6, or 7 dpi.
Sorted populations of macrophages were administered intratracheally to WT or β6 KO mice. For the AM adoptive transfer, autofluorescent FITC+ F4/80+ lung cells were sorted from donor mice. 200,000 cells were transferred into each recipient mouse in 100 μl PBS.
For transfer of naive PDM, the mouse peritoneal cavity was washed/flushed with 10 ml of PBS, and the F4/80+ population sorted. 500,000 cells were transferred into each recipient mouse in 100 μl PBS. After 7 days, mice were sacrificed and AM populations assessed. Donors and recipients were distinguished by CD45.1 and CD45.2 markers.
BALF or whole lung homogenate was collected from PBS control and influenza inoculated mice at the indicated time post-infection. TNF-α and IL-6 protein expression was determined using the Milliplex Mouse 25-plex Cytokine Detection System mouse cytokine kit (Millipore, Billerica, MA) on a Luminex100 109 reader (Luminex Corp., Austin, TX) according to the manufacturer’s protocol. Cytokine concentration was calculated using a calibration curve obtained in each experiment using the respective recombinant proteins.
Active TGF-β1 protein levels were quantified by ELISA assay (BioLegend) on whole lung homogenates according to manufacturer’s protocol.
Mice were deeply anesthetized with isoflurane and intranasally administered 1 μg of recombinant TGF-β1 (PeproTech, Rocky Hill, NJ) or PBS as previously described [46]. After 18 h, mice were lightly anesthetized and intranasally inoculated with 104 TCID50 CA/09 virus. At 2 dpi, mice were given an additional dose of rTGF-β1 or PBS and monitored for morbidity as described. For homeostasis studies, 1 μg of TGF-β1 was administered every 48 hours for 3 weeks.
RNA from whole lung homogenates was isolated using a Trizol (Invitrogen) per manufacturer’s protocol. cDNA was synthesized using the SuperScript VILO cDNA synthesis kit (Invitrogen) and the following parameters: 25°C for 10 min, 42°C for 1 hour, and 85°C for 5 min. PCR was run using 2 μl of diluted cDNA (1:5 in dH2O) to detect IFNa2/IFNa11 or IFNb mRNA using the Taqman primer/probe assay system (Life Technologies) per manufacturer’s protocol. Samples were normalized to GAPDH mRNA expression as the internal control.
IFN-α (PBL Assay Science, Piscataway Township, NJ) and IFN-β (BioLegend, San Diego, CA) protein levels were quantified by ELISA assay on whole lung homogenates according to manufacturer’s protocol.
Statistical analysis was performed with GraphPad Prism (San Diego, CA) as described in the Fig legends. Statistical significance was defined as p < 0.05.
Microarray data has been deposited in the GEO database with accession number GSE68802.
|
10.1371/journal.pcbi.1004078 | NEMix: Single-cell Nested Effects Models for Probabilistic Pathway Stimulation | Nested effects models have been used successfully for learning subcellular networks from high-dimensional perturbation effects that result from RNA interference (RNAi) experiments. Here, we further develop the basic nested effects model using high-content single-cell imaging data from RNAi screens of cultured cells infected with human rhinovirus. RNAi screens with single-cell readouts are becoming increasingly common, and they often reveal high cell-to-cell variation. As a consequence of this cellular heterogeneity, knock-downs result in variable effects among cells and lead to weak average phenotypes on the cell population level. To address this confounding factor in network inference, we explicitly model the stimulation status of a signaling pathway in individual cells. We extend the framework of nested effects models to probabilistic combinatorial knock-downs and propose NEMix, a nested effects mixture model that accounts for unobserved pathway activation. We analyzed the identifiability of NEMix and developed a parameter inference scheme based on the Expectation Maximization algorithm. In an extensive simulation study, we show that NEMix improves learning of pathway structures over classical NEMs significantly in the presence of hidden pathway stimulation. We applied our model to single-cell imaging data from RNAi screens monitoring human rhinovirus infection, where limited infection efficiency of the assay results in uncertain pathway stimulation. Using a subset of genes with known interactions, we show that the inferred NEMix network has high accuracy and outperforms the classical nested effects model without hidden pathway activity. NEMix is implemented as part of the R/Bioconductor package ‘nem’ and available at www.cbg.ethz.ch/software/NEMix.
| Experiments monitoring individual cells show that cells can behave differently even under same experimental conditions. Summarizing measurements over a population of cells can lead to weak and widely deviating signals, and subsequently applied modeling approaches, like network inference, will suffer from this information loss. Nested effects models, a method tailored to reconstruct signaling networks from high-dimensional read-outs of gene silencing experiments, have so far been only applied on the cell population level. These models assume the pathway under consideration to be activated in all cells. The signal flow is only disrupted, when genes are silenced. However, if this assumption is not met, inference results can be incorrect, because observed effects are interpreted wrongly. We extended nested effects models, to use the power of single-cell resolution data sets. We introduce a new unobserved factor, which describes the pathway activity of single cells. The pathway activity is learned for each cell during network inference. We apply our model to gene silencing screens, investigating human rhino virus infection of single cells from microscopy imaging features. Comparing the learned network to the known KEGG pathway of the genes shows that our method recovers networks significantly better than classical nested effects models without capturing of hidden signaling.
| Network inference benefits substantially from perturbation experiments, such as RNA interference (RNAi) screens. Monitoring high-dimensional effects of gene silencing enables inference of non-transcriptional network structures that cannot be learned on observational data alone [1]. Nested effects models (NEMs) are a class of probabilistic graphical models that aim at learning hierarchical dependencies from such intervention experiments. Upon perturbing nodes in a signaling graph, their connectivity is inferred from the nested structure of observed downstream effects. The concept was first introduced in [2]. Since then, many further additions concerning, for example, parameter inference, structure learning, and data integration, were developed [3, 4]. In addition, dynamic models for time series data have been developed [5–7]. In [5], a first application of dynamic nested effects models to time laps microscopy data has been described, but the model can not handle single-cell data. A Bayesian network representation of NEMs in [8] introduces a probabilistic notation for signal propagation, but in practice the signaling is kept deterministic. In all previous NEM models and applications, the signaling pathway under observation is assumed to be active and the signal flow disrupted by silencing the signaling genes one by one.
In principle, RNAi experiments are a highly informative for learning NEMs. Perturbations are introduced by gene silencing in cells through RNA interference using siRNAs [9, 10]. Effects of the knock-downs are then captured by high-dimensional down-stream observations. The screening data analyzed here, comprises imaging data of thousands of individual cells for genome-wide gene silencing. However, the experiments come at the cost of high noise levels, as well as biological and technical biases, including off-target effects [11, 12]. These confounding factors complicate the analysis and interpretation of the screening results. On the other hand, RNAi screens currently reach very high resolution. Per knock-down, the present data sets comprise about 300 image features for several hundred individual cells, which allows for a very detailed analysis of a knock-down event. However, it has been shown that measurements from individual cells of the same experiment can differ widely, for example, due to local environmental differences [13, 14]. Such variation on the single cell level needs to be accounted for. Otherwise, an ambiguous signal is obtained, when averaging over the cell population of a knock-down.
Here, we specifically investigate single-cell observations of pathogen infection screens [15–17]. The experiments monitor cells with an siRNA knock-down during infection with human rhinovirus (HRV). After siRNA knock-down, the pathogen is added to the cells, and the success of infection as well as many other cellular features are extracted from microscopy images taken of the cells from each experiment [18–20]. The aim is to infer a signaling cascade involved in pathogen entry in to the host cell. However, a challenge in the analysis of data from this experimental setup is that by experimental design even in mock controls (i.e., infection without knock-down) the infection rate is far from complete. In fact, the multiplicity of infection (MOI) of the assay was optimized to reach 30 to 50% infected cells, such that both infection-decreasing and infection-increasing hits can be detected. Which cells in the population finally get infected is, at least to some extent, the result of stochastic effects, since cellular processes can be differently manifested in different cells. The multi-functional nature of proteins, for instance, enables a single host factor to enhance a signaling cascade, and at the same time may antagonize other processes that support or inhibit infection. Obviously, infected cells were reached by a pathogen triggering some signal to get internalized. However, for uninfected cells, it is unknown whether a pathogen actually attempted to infect them, which is crucial for determining the effect that the gene knock-down had on these cells. Wrongly assuming that the pathway is active, even though it is not, can result in conflicting knock-down schemes. In the original NEM setting, individual cell observations are summarized for each signaling gene.
To address the problem of network learning when the activation state of the signaling pathway is unknown we introduce a new model, called NEMix, extending the existing NEM framework in several ways. First, we do not summarize the data across cells, but rather perform network inference using the single-cell observations directly. Furthermore, we model the unknown pathway activation with an additional hidden random variable in the graph of signaling genes. The activation state is then estimated for each individual cell. The pathway activity can be regarded as an additional hidden silencing event in the signaling graph. We introduce a general theoretical framework for probabilistic combinatorial knock-downs in NEMs. We develop our model for the most general case, not making any assumptions about the signal propagation. We have implemented the special case of one hidden variable with probabilistic knock-down, where the remaining network is kept deterministic. For inference of the hidden pathway state, we developed an EM algorithm [21]. This step is repeated for each proposal structure during the network search.
We developed NEMix, a new model based on NEMs, which allows to estimate activity of a pathway in individual cells. A NEM is a graphical model, consisting of two graphs. The transitively closed graph Φ encodes dependencies among signaling gene nodes Ss ∊ 𝓢, which are silenced one by one. The bipartite graph Θ connects a set of observable feature nodes Ee ∊ 𝓔 uniquely to the signaling genes (Fig. 1A). We seek the structure of Φ, i.e., the topology of the signaling pathway, by inferring it from the nested structure of observed effects. For a data set 𝒟 = (dek) of a set of knock-down experiments k ∊ {1, …, K} and observed features e ∊ {1, …, m}, the likelihood function given Φ and θ is
P ( 𝒟 ∣ Φ , θ ) = ∏ e = 1 m ∏ k = 1 K P ( d e k ∣ Φ , θ e = s ) , (1)
where θe = s indicates that feature e is connected to signaling gene s ∊ 𝓢.
The NEMix model consists of the same two graphs Φ and Θ, but has an additional binary hidden variable Z added to the signaling graph Φ. Its connections to the signaling genes, as well as its overall knock-down probability p0 = P(Zkc = 0), are unknown and inferred for each individual cell during the network reconstruction process. Given single cell data 𝒟 = (dekc) with c = 1, …, ck cells in knock-down experiment k, the likelihood function of the NEMix model, given Φ and θ, is
P ( D ∣ Φ , θ ) = ∏ e = 1 m ∏ k = 1 K ∏ c = 1 c k ∑ j ∊ { 0, 1 } p j P ( d e k c ∣ Φ, θ e = s, Z k c = j ). (2)
A detailed derivation of the model and its implementation are given in the Models section. If a signal is activating a pathway, or parts of it, the signal flow is the same as in the NEM. Also the observed knock-down effects for the features Ee are the same. However, when the pathways input signal is inactivated, the knock-down pattern of the features changes (Fig. 1A and B, cells 7 to 15). Not accounting for the pathway disruption can mislead inference of the structure Φ (Fig. 1A, left model).
The connectivity of Z is learned in a greedy fashion during structure inference. For the knock-down probability of the hidden variable, p0, we implemented an EM algorithm, which estimates jointly p0 from each cell’s observation and the connections of observations to signaling genes, θ. In the following, we show improved network inference with NEMix in simulations and then infer networks of high accuracy, from single cell gene silencing experiments.
To test our model, we performed a large simulation study. We generated 30 network structures with 5 signaling genes, randomly sampled from KEGG pathway maps [22] as previously described in [6]. To each network the hidden input signal was attached randomly. The resulting 30 sample networks are shown in supplementary S1 Fig. From each network, we sampled 50 data sets on 300 observed features in the following way. For each gene, we simulated single knock-downs in 200 cells. To the observed features we added another 30 noise features, not attached to any signaling gene. The data sets were generated in the following way. We sampled effects from a normal distribution with mean me = 1 and non-effects from a normal distribution with mean mn = 0. The standard deviation for each experiment was sampled uniformly between 2 and 2.5. We furthermore sampled 200 cells for control experiments. The negative control cells do not show any effects and are therefore drawn from the non-effect distribution. The positive control cells always show effects and hence are drawn from the effect distribution. The whole simulation process was repeated for five different fractions of pathway disruption, p0 ∊ {0, 0.3, 0.5, 0.8, 1}. NEMix inference was restarted for 16 initial networks. Each of them consists of the empty graph Φ plus a unique attachment of Z to the signaling genes. Setting the maximal out-degree of Z to two, there are 16 possible such attachments of Z. This regularization on the edges of Z reduces the search space significantly. During structure search we also imposed this restriction, but additionally allowed transitive edges that had to be added as a consequence of the insertion of any edge connecting Z to a gene (see Models section).
We compared NEMix to two other NEM models and, for a baseline comparison, to a random approach, where network edges are sampled uniformly with probability 1/n, where n∣Φ∣ is the number of signaling nodes. This probability was chosen as it creates networks with approximately the same number of edges as in the original graphs. To assess the impact that pathway disruption has on the cell population level, we ran the simulations on a standard NEM using the log-likelihood model introduced in [23]. For the NEM approach we had to summarize the single cell observations to the gene level. For these gene-level data sets we used p-values of a Wilcoxon test comparing the cell population of a knock-down to the control distribution. From the p-value distributions a Beta-Uniform-Mixture model was estimated. For each feature a density value is calculated from this model, indicating the effect strength of the knock-down. These density values are used as the input data, as previously introduced in [23]. The third approach, called single-cell NEM (sc-NEM). is a NEMix model on individual cell observations, but with fixed p0 = 0, i.e., a single-cell observation-based NEM without considering uncertain pathway activity. For all three models, we applied a uniform prior on the feature attachments θ, and no prior knowledge was added for the network structures Φ. The NEMix parameter p0 was initialized by drawing from a uniform distribution in each EM restart. As NEMix and sc-NEMs infer networks on single-cell observations, we calculated log odds ratios from each observation based on the positive and negative control distributions (see ‘Modeling the effect likelihoods’ in S1 Text). For NEMs and sc-NEMs, we used maximum likelihood estimation to infer θ and in the NEMix it is estimated by in an EM algorithm. Structure learning is performed using a greedy hill climbing algorithm, initialized with an empty network.
Fig. 2A summarizes the overall performance for all methods and the different fractions of pathway signal perturbation p0. We display accuracy of the edge recovery, for varying p0. We also calculated the area under the ROC curve (AUC) based on the edge frequencies of the 50 replicate data sets, which yielded similar results in terms of accuracy (see supplementary S2 Fig). As expected, all methods performed equally well when there is no signal disruption (p0 = 0). However, when p0 is moderate to high, NEMix performs significantly better than the other methods. If the triggering signal is always turned off, performance of all methods drops drastically. Intuitively, this is because in such a special case, all features downstream of Z always show an effect and hence they cannot be used for structure learning. For example, if, in Fig. 1B, Z is inactive for each cell, we could not infer the structure among S2 and S3. In reality though, permanent shut down of the pathway is very unlikely. For the infection screens p0 = 1 would mean that no cell is ever infected. Pathway activity estimates are also of overall high accuracy (Fig. 2B). Although simulation results demonstrate that the performance of learning Z and θ varies, depending on the network structure, the average performance is very good (S3 Fig, S4 Fig, S5 Fig).
Currently, one of the main obstacles for learning larger NEMix models is the fast growing run-time for n > 5 network nodes. Run-time is further increased by a factor of n, when initiating the algorithm with each possible connection of Z to one of the knock-down genes. To assess its performance on larger networks, we ran a reduced simulation study on n = 5, 10, and 15 genes. The setup and results of the study are described in detail in S6 Fig. Larger networks of 15 nodes can still be estimated very well (S6 Fig. A) and estimation of the parameter p0 even improves (S1 Fig. D). However, the average time to estimate a 15-node network was 9.5 hours. This is substantially more than the average 1.9 hours needed for 10-node networks. Thus, in a highly parallelized computing environment, even larger networks can be estimated.
We also assessed the connection of features to the signaling genes in the inferred graph Θ. There can be situations, where attachment of features is equally likely for several signaling genes. In these cases, where no single gene is preferred, we counted a feature as correctly attached if it was connected to any of the signaling genes with equal likelihood. Accuracy of the θ estimates is high (> 80%) for small p0 values and decreases with increasing p0. For small p0, also performance of the sc-NEMs is good, which shows the advantage of learning on the single-cell data level. However, NEMix stands out from the other methods for higher p0. Recovery of noise features, i.e., correct filtering of the additionally added uninformative features, is not strongly affected by the hidden signal (see supplementary S7 Fig). Analyzing individual networks, one again observes high variation in performance (see supplementary S8 Fig).
We applied NEMix in the context of infection signaling, using the RNAi screening data monitoring HRV infection, mentioned in the introduction. Briefly, viruses were added to the siRNA transfected cells and after an incubation time, cells were fixated, stained, and then imaged. Subsequently, 360 cell features were extracted from the 9 images per knock-down experiment using the software CellProfiler [24]. For the whole experimental procedure the protocols of [17] were followed. The HRV assay is rather short with an infection time of only seven hours, resulting in measurements proximal to the infection event. The short time range is advantageous, because it leaves less room for confounding developments in the cells. Furthermore, the used antibody resulted in clean readouts, well to extract from the images.
Before using the data for network inference, we performed two additional filtering steps. For each knock-down, the well is split into 9 images. They are arranged in three rows and three columns. We used only the middle image, because it is of the highest quality. In this way we avoided too many out-of-focus cells, which bias especially the cell texture features. After this filtering step, we had around 200 to 300 cells per knock-down. A second filtering step concerns siRNA off-targets [25]. We sought to avoid confounding by this effect and therefore selected only genes with low predicted off-target effects as described in ‘siRNA filtering for off-targets’ of S1 Text.
We applied NEMix to a small subset of the screened genes, in order to recover a known pathway. We decided on the well-known MAP-Kinase signaling cascade as a proof of principle, for several reasons. First, it has been studied and validated in great detail [26–28], such that the available signaling network from the KEGG database [22] can be used as a reliable source to compare to. Second, the pathway is known to be involved in HVR infection signaling, where it is associated with asthmatic and COPD exacerbation [29–31]. Finally, we observed an enrichment for low off-target siRNAs in this pathway when performing a gene set enrichment analysis [32] (see supplementary S9 Fig). We then selected a small subset of 8 MAP-Kinase pathway genes for analysis based on the derived score for predicted off-target effects. Nodes of KEGG pathways can contain several genes. We selected genes such that they are all assigned to different KEGG nodes using a weighted maximum bipartite matching of low off-target siRNAs and unique KEGG nodes. After gene selection, we inferred networks for the 5 and 8 genes with lowest off-target score.
Like in the simulation study above, we compared the NEMix model to the NEM and the sc-NEM approach. As input data sets, the local effect likelihoods from the selected knock-down gene experiments were computed as follows. As the experiments lack reliable controls, we instead used a random sample of cells from the plate on which the gene was located, assuming that the majority of knock-downs will not have an effect. Like for the simulation study, we derived the cell population effects for the NEM from Wilcoxon tests, comparing the knock-down experiment to the control. From the resulting p-value distributions, effect strengths for the features were estimated using the Beta-Uniform-Mixture model. Log odds ratios for sc-NEMs and NEMix in this case are calculated only based on one control distribution (see Models section). NEMix inference again is repeated for the 16 initial networks of all possible connections of Z with maximal out-degree 2 to the empty graph Φ. Like in the simulation study, p0 was initialized by drawing randomly from a uniform distribution. Again we used uniform priors for θ and imposed no priors for the signaling networks other than the maximal out-degree of Z (plus the transitive edges that need to be added).
The known KEGG network and the inferred results for the top 5 signaling genes are displayed in Fig. 3A-E. Results for the top-8 gene network are given in S10 Fig. To assess robustness of the learned networks, we repeated the inference on 50 bootstrap samples of the original data set. Both networks show high AUC values and even better accuracy (see Table 1). As can be seen from Fig. 3F, network inference was very robust for the top-5 gene network. For the top-8 gene network, performance had a slightly higher variation. Individual plots for sensitivity and specificity are given in supplementary S11 Fig. A, B. Also the estimate of p0 shows only little variation (S11 Fig. C). In all cases, the likelihood score of the known KEGG network is much lower than for the best inferred networks, indicating that under the assumptions of our model, the data and the KEGG database do not perfectly agree. Possible reasons for this observation include our model missing to explain part of the data correctly, the KEGG database being incomplete, and inaccuracies in the data generating process. Nevertheless, the accuracy value of 0.85 for the learned NEMix outperforms all other methods. All edges contained in the learned NEMix models are of high robustness (> 80% for 5 genes, and > 70% for 8 genes). Consensus networks of the bootstrap results are shown in supplementary S12 Fig.
Furthermore, the hidden root Z is attached to the same nodes in both the known KEGG graph and the estimated network for 5 genes. Also the inferred 8 node network connects Z to the same three genes. As genes were selected based on small off-target effects of their targeting siRNAs, they are not necessarily hits for HRV infection. However, of the selected genes EGFR [33], TAB2 [34] and CACNA2D3 [35] have been shown to be involved in this process.
All models have a built-in filter for uninformative features, which has been previously introduced in [36]. A comparison shows that averaged over the bootstrap samples, for all three methods, the set of used features largely agrees (supplementary S13 Fig and S14 Fig). The maximum likelihood attachments of features to the knock-down genes and the null node are shown in supplementary S15 Fig and S16 Fig, together with a detailed description of the different feature types. The inferred signaling disruption of p0 = 0.42 seems rather high. We compared this to the average infection rate in mock experiments, i.e., cells without siRNA knock-down. These resemble cases, where Z can be perturbed but none of the other signaling genes in the network. Mock wells from plates of the 8 genes used here, actually have a much higher percentage of uninfected cells, roughly in the range of 75 to 81%. However, this comparison should be taken with caution since control wells of these screens might have suffered from strong plate location bias, as they were located on the margins of the plate.
As a general observation, NEMix-inferred networks were sparser than those obtained from NEMs, because spurious edges introduced in the latter are correctly explained by hidden pathway activity Z in NEMix. Therefore, NEMix networks have increased specificity, which might come at the cost of some missing true edges. Especially the 8-gene networks inferred by NEM and sc-NEM are much denser than the known KEGG network. A sparse network is beneficial in the sense that it allows to focus on a small set of highly specific edges. For validation experiments, it is desirable to have a low false positive rate in the predicted interactions as usually only very few of these dependencies can be experimentally tested.
RNAi screens are known to be prone to many sources of noise and bias such that their analysis is highly challenging. Here, we have identified one confounding factor, namely heterogeneous signaling pathway activation within a cell population, and incorporated it directly into a novel probabilistic model for pathway reconstruction. To address the problem of unknown activation of signaling pathways during network inference, we have introduced a general framework, building on NEMs, to handle hidden combinatorial knock-downs in a probabilistic manner. With NEMix we provide an implementation for inference under unknown pathway stimulation. For the first time, image features are explicitly used on the single-cell level for NEM inference, acknowledging large cell-to-cell variation. We have demonstrated the advantages of NEMix over current NEMs in simulations and inferred highly accurate networks in a case study on HRV infection. Especially, when the underlying true signaling networks are expected to be sparse, NEMix is beneficial. It removes spurious edges introduced due to confounding factors and therefore reduces the false positive rate, a desired property when it comes to validation of edges.
A limitation of the current model formulation is the assumption of independent single cell observations. In reality, this assumption might not be met as cells can be biased due to their location and neighbors. Removing this bias either by normalization or explicit modeling, as for example in [14], could further improve the model. Furthermore, in the current data sets cells can be in different cell cycle states. Grouping them according their states may remove further biases, but this clustering task is itself very challenging.
Another general limitation of NEMs and NEMix models is that they cannot learn certain pathway features. From static data, NEMs cannot resolve any loop structures by construction. This is a general problem for network inference without time resolved data. Therefore, only performance statements based on comparing transitively closed pathways can be made. The sampled graphs in the simulation are already transitively closed and since the transitive closure is a feature inherent to all the models we compare, it should not influence the ranking based on their performance. Before comparing a network to the corresponding KEGG pathway, we also built its transitive closure. This fact should be considered when interpreting the inferred models. For example, the model does not allow for distinguishing a feed forward loop from a sequential cascade; however, the hierarchical order of genes in the network would remain the same, and this piece of information does already provide considerable insight into the biological processes. The way we have assessed performance here puts particular emphasis on this hierarchical structure of the network nodes.
Further improvements could be achieved during data preparation. Image segmentation is not always perfect and might introduce technical biases into data sets, adding more confounding factors. If data is not curated carefully, we risk to capture technical biases with the additional hidden variable in NEMix models. Another interesting aspect of the data sets deserving a more thorough analysis, is the nature of the image features themselves. Here, readouts have been used to infer the graph of signaling genes. However, one could investigate in more detail how features are grouped when attaching them to the signaling genes. Some features might not contribute useful information and could be filtered in advance, others might be redundant. Future projects could use the output of NEMix models and seek for biological interpretation of feature correlations.
In case of cell infection screens, infection efficacy was an obvious factor that needed to be addressed. However, the same idea could be applied to other sources of noise. For example, transfection efficacy of the knock-downs could be considered. Quality and efficacy of a knock-down can be quantified by mRNA levels (qPCR) or protein level (western blot analysis) of a gene. However, for high-throughput assays, such confirmation is not available for most gene knock-downs. In order to account for different siRNA transfection efficacies further hidden variables could be introduced. In contrast to the global Z variable introduced here, hidden knock-down rates would then be estimated for each gene individually. As a consequence, the complexity of the problem would increase substantially. Instead of one parameter, n (number of genes) parameters would have to be estimated. Furthermore, knock-down probabilities could only be estimated from a fraction of the observations (e.g., cells under the specific knock-down). Another drawback is that the increased number of hidden variables gives rise to identifiability problems when estimating infection efficacy in combination with the knock-down rates. For example, if the hidden variable Z was only attached to one signaling gene, effects of Z and a failed transfection could not be distinguished. Although extending the NEMix model to this situation would be an interesting future project, we believe that problems in the transfection process play an overall minor role. For the current experiments, KIF11 siRNAs (cell killers) were used to control transfection quality on the plate level. For the plates containing the cells used in our analysis, these controls show very high penetrance, i.e., out of an average of 2000 cells per well, on average only 7% of cells survive in these wells. Although this test does not make a statement about the efficacy of individual siRNAs, it ensures the general functioning of the transfection process. Additionally, the library vendor claims the knock-down efficacies achieved with their smart-pool siRNAs to be in the range of 70–95%. This proportion is a result of many possible sources of imperfect gene silencing, including non-transfected cells and off-target effects. Given the above facts in combination with our off-target filtering strategy, we are convinced that the analyzed data are of high quality.
We tried to minimize the general problem of confounding siRNA off-targets by considering only genes targeted by siRNAs with low predicted off-target effects. This selection step helps to achieve reasonably unbiased results with our model, but it also limits the gene sets we can analyze. Ideally, we want to be able to select any gene of interest. This scenario calls for models that can correct the off-target effects on the single-cell level. A potential solution to this issue could be delivered by NEMs directly. We could still learn the networks based on siRNA knock-downs directly, but handle the signal propagation differently. With NEMix it is already possible to use each siRNA as a combinatorial knock-down. In reality however, individual genes are knocked-down to different degrees by an siRNA. In a NEM, this would mean to split up the silencing signal of an siRNA into partial knock-downs of several genes. Then, signal propagation would have to be formulated in a fully probabilistic fashion and NEMs would have to be reformulated such that their nodes do not have binary states anymore. Further developing NEMix, by integrating the above mentioned shortcomings, will make the models more powerful for future network reconstruction tasks.
Especially in the light of single cell data sets, which show large heterogeneity among individual observations, our approach is beneficial. Such data sets are becoming more and more available, and they reveal that the high cell-to-cell variation has severe consequences when summarizing such heterogeneous observations. On the population level, the signal is potentially confounded as it is only contained in part of the observations. NEMix uses the full power of single-cell experiments, as it is applied on the single-cell level directly, avoiding any data averaging. Only at this data resolution, the heterogeneity within a cell population can be accounted for and it becomes possible to investigate potentially confounding factors, such as, for example, pathway activity. NEMix is the first NEM-based method with additional unknown components in the signaling graph Φ. It is capable of inferring these missing data and provides an estimate for the fraction of signal disruption. We find such ambiguous signaling in RNAi infection screens and we have demonstrated that NEMix can improve network inference substantially by accounting for the confounding factor.
A NEM, as introduced in [2], aims to infer the hidden dependency structure among a set of n binary signaling variables 𝓢 from the nested structure of m observed effect variables 𝓔 (features). It therefore consists of two directed graphs, one describing the dependencies among the signaling genes and one connecting the features to the genes.
The binary adjacency matrix of signaling genes is denoted Φ = (ϕks), with ϕks = 1 if gene k propagates its effects to gene s and using the convention Φk, k = 1, for all k. The signaling graph Φ is thus always transitively closed. If a gene is silenced, the effect is propagated deterministically along the edges of Φ. The connection of features 𝓔 to the genes 𝓢 is given by parameters θe, where θe = s indicates that feature e is linked to gene s. For a gene k and a feature e, a NEM predicts an effect of k on e if there is a gene s such that ϕks = 1 (i.e., k and s are connected), and θe = s (i.e., s has an effect on e). The observed data are denoted D = (dek), where each dek is the measurement of feature e under perturbation of k (Fig. 1A).
Given an external signal which affects one or more of the signaling genes, each of them will have a binary signaling state. The state value is 0 if the signaling is interrupted, i.e., does not reach the node, and 1 if the signal reaches the node, i.e., the natural state of a stimulated pathway.
For inferring the structure Φ among the signaling genes, we consider its posterior
P ( Φ ∣ D ) = P ( D ∣ Φ ) P ( Φ ) P ( D ), (3)
where the marginal likelihood P(D∣Φ) can be obtained by integrating out the connections of features to the genes,
P ( D ∣ Φ ) = ∫ θ P ( D ∣ Φ, θ ) P ( θ ∣ Φ ) d θ, (4)
with prior distribution P(θ∣Φ). In the absence of further knowledge, the prior is usually set to the uniform distribution. Given the network structure and assuming conditional independence of the parameters θe and of the silencing experiments k, the marginal likelihood becomes
P ( D ∣ Φ ) = ∏ e = 1 m ∑ s = 1 n ∏ k = 1 K P ( d e k ∣ Φ, θ e = s ) P ( θ e = s ). (5)
The local effect likelihoods P(dek∣Φ, θ) denote the probability of observing an effect in feature e under knock-down of gene k. They can usually be pre-computed from the data and different approaches have been proposed [2, 23, 36]. For the results presented below, log-odds ratios as introduced in [36] were used (see ‘Modeling the effect likelihoods’ in S1 Text for details).
We first define the NEMix model and then derive it in detail. A NEMix consists of a nested effects model with effects graph Θ and an extended signaling graph Φ. The signaling graph Φ describes the dependency structure among the signaling genes and has an additional binary hidden variable Z indicating pathway activity. Z is a root of Φ, i.e., it can be connected to any of its nodes and does not have any direct connections to features in θ. The silencing probability of Z is denoted by p0 and is a priory not known. For a set knock-down experiments k ∊ {1, …, K}, with single cell observations c ∊ {1, …, ck} of signaling gens s ∊ {1, …, n} and features e ∊ {1, …, m}, the marginal likelihood of a NEMix is
P ( D ∣ Φ ) = ∏ e = 1 m ∑ s = 1 n P ( θ e = s ) ∏ k = 1 K ∏ c = 1 c k ∑ j ∊ { 0, 1 } p j P ( d e k c ∣ Φ, θ e = s, Z k c = j ), (6)
where pj = P(Zk = j).
Structure learning is performed using a greedy heuristic to find an optimal network. Similar to the NEM procedure described in [3], edges are incrementally added if the likelihood is increased (see ‘Structure learning’ in S1 Text). In addition, our approach is restricted to structures without incoming edges into the hidden root Z. We initialize the algorithm with a set of initial networks. These consist of the empty graph and one edge connecting Z to one of the knock-down genes. Additionally, we limit the out-degree of Z to two. Here, by out-degree we mean only the non-transitive edges. We still allow the insertion of transitive edges from Z to any signaling gene, which has to be added in order to fulfill the transitivity requirement. This regularization reduces the search space and prevents that too many dependencies between genes are explained by Z alone.
As for classic NEMs, network structure scoring involves the marginal likelihood. For the NEMix model, P(𝒟∣Φ) cannot be optimized analytically. Marginalization over the feature attachments is omitted in our extended model. Instead, we estimate θ jointly with p during model inference. To do so, we approximate the marginal likelihood (10) by the expectation of the complete data log-likelihood
P ( D, Z ∣ Φ, θ, p ) = ∏ k = 1 K ∏ c = 1 c k ∏ j ∊ { 0, 1 } p j ∏ e = 1 m P ( d e k c ∣ Φ, θ e = s, Z k c = j ) Z k c ( j ), (20)
with respect to Z, where θ and p0 need to be efficiently estimated. For this task we have developed an EM algorithm. A derivation of the expected hidden log-likelihood and the maximum likelihood estimates is given in ‘Estimating the hidden signal’ of S1 Text. When starting the EM algorithm, p0 is initialized with a random draw from the uniform distribution and for θ we use a uniform initial configuration.
The NEMix model is included as part of the R/Bioconductor package NEM as an additional inference type. It is invoked by calling the package’s main function NEM(data, inference = ‘NEM.greedy’, control) and choosing the inference type control$type = ‘NEMix’. (See ‘NEMix implementation in NEM package’ in S1 Text for more detailed instructions on the implementation and usage of NEMix in R). To record run-times of NEMix model estimation, simulations were run without any parallelization on a 1.7GHz Intel i7 machine. Only one starting configuration was used, and EM iterations were performed using three restarts to avoid local optima that are globally suboptimal. For realistic data sets of 300 features and 200 cells per knock-down, NEMix estimation took on average nine minutes for 5-gene networks, with an average of 13 iteration steps until convergence of the EM algorithm. For the 8-gene network, the average run-time was 66 minutes, while the average number of iterations per EM round remained 13 also for these larger networks. The longer run-times of NEMix models as compared to NEMs are primarily due to the hidden data estimation. Each structure scored once in a NEM inference, needs to be scored 40 times on average during NEMix estimation. In addition, the input data sets are roughly 200 times larger.
|
10.1371/journal.pntd.0007131 | Discovery of Leptospira spp. seroreactive peptides using ORFeome phage display | Leptospirosis is the most common zoonotic disease worldwide. The diagnostic performance of a serological test for human leptospirosis is mainly influenced by the antigen used in the test assay. An ideal serological test should cover all serovars of pathogenic leptospires with high sensitivity and specificity and use reagents that are relatively inexpensive to produce and can be used in tropical climates. Peptide-based tests fulfil at least the latter two requirements, and ORFeome phage display has been successfully used to identify immunogenic peptides from other pathogens.
Two ORFeome phage display libraries of the entire Leptospira spp. genomes from five local strains isolated in Malaysia and seven WHO reference strains were constructed. Subsequently, 18 unique Leptospira peptides were identified in a screen using a pool of sera from patients with acute leptospirosis. Five of these were validated by titration ELISA using different pools of patient or control sera. The diagnostic performance of these five peptides was then assessed against 16 individual sera from patients with acute leptospirosis and 16 healthy donors and was compared to that of two recombinant reference proteins from L. interrogans. This analysis revealed two peptides (SIR16-D1 and SIR16-H1) from the local isolates with good accuracy for the detection of acute leptospirosis (area under the ROC curve: 0.86 and 0.78, respectively; sensitivity: 0.88 and 0.94; specificity: 0.81 and 0.69), which was close to that of the reference proteins LipL32 and Loa22 (area under the ROC curve: 0.91 and 0.80; sensitivity: 0.94 and 0.81; specificity: 0.75 and 0.75).
This analysis lends further support for using ORFeome phage display to identify pathogen-associated immunogenic peptides, and it suggests that this technique holds promise for the development of peptide-based diagnostics for leptospirosis and, possibly, of vaccines against this pathogen.
| Leptospirosis is an infectious disease that is transmitted from animals to humans. It is associated with a broad range of clinical presentations, and diagnostic tests with high diagnostic accuracy are required in order to enable accurate diagnosis. Leptospirosis is diagnosed by detecting DNA of the pathogen or antibodies against it in patients’ blood; the latter are preferred in resource limited regions, and diagnostics based on peptides (small fragments of proteins) are advantageous because they are inexpensive to produce and more stable in hot climates than full-length proteins. We used a technique called open reading frame phage display to identify peptides from Leptospira spp. that could be used to detect antibodies against them in human blood. In this method, the pathogen’s genome is fragmented, the corresponding peptides displayed on the surfaces of phages (viruses that infect bacteria), and the peptides that bind most strongly to the patients’ antibodies are then selected by screening. Using this method, we identified 2 leptospiral peptides that accurately identified antibodies against Leptospira spp. in sera from patients with leptospirosis. These results are encouraging because they demonstrate that ORFeome phage display may be a powerful tool to develop better diagnostics for leptospirosis for use in less developed areas.
| Leptospirosis is the most common and geographically widespread zoonotic disease worldwide. It is caused by pathogenic strains of the spirochete Leptospira spp. Humans are accidental hosts who get infected through direct contact with body fluids of carrier animals or with contaminated water and soil [1]
Human leptospirosis is a major problem in countries of tropical and, less often, subtropical climates. Numerous outbreaks have been reported in association with rainy seasons, floods, and recreational and sports activities [2]. Globally, leptospirosis is estimated to occur at a frequency of approximately 1.03 million cases per year with a mortality of 5.7% [3]. In Malaysia, incidence is estimated to be 1–10 cases per 100,000 individuals per year with case fatality rates around 10% [4]. Leptospirosis was designated a notifiable disease in 2011, and reported incidence increased subsequently, as illustrated by a reported incidence for 2015 of 17.5 cases per 100,000 individuals, 86 outbreaks, and case fatality rates of 0.6% [5].
The microscopic agglutination test (MAT) is the serological ‘gold standard’ for diagnosis of leptospirosis. Unfortunately, this test is not only time-consuming and laborious, but it also can only be performed by a reference laboratory with a collection of serovars endemic in the region. To overcome the drawbacks of the MAT, numerous serological assays have been developed that do not depend on the availability of viable leptospires, particularly IgM ELISA tests based on either whole cell extracts or recombinant proteins [6–8]. Other serological approaches that have been described were agglutination, dipstick, and lateral flow assays [8]. Among these serological tests, in-house Leptospira ELISA using whole cell extracts have achieved the highest specificity and sensitivity [9]. Nevertheless, this method requires the standardized preparation of whole cell extracts from serovars most relevant to the geographical region [10].
Phage display is mainly used for antibody generation [11], but also for the identification of pathogen-associated immunogenic proteins; for the latter, it is based on using fragmented pathogen genomes or cDNA fragments to allow high-throughput screening for novel potential antigens [12–16]. Immunogenic proteins are often identified using 2D-PAGE of proteins from cultivated bacteria followed by mass spectrometry. However, phage display offers clear advantages, notably better performance for the identification of immunogenic proteins below 10 kD, low abundance proteins, and proteins only expressed in host-pathogen interaction [17, 18]. ORFeome phage display is an improved variant of the phage display technique in which the library quality of fragmented genomes or cDNA fragments is improved by a unique open reading frame enrichment step [19–22].
Diagnostics for use in hot, resource-limiting environments should be inexpensive to produce and stable at ambient temperature. Peptide antigens fulfil both requirements; besides, they offer the added advantage that they can be synthesized in biotinylated form, thus facilitating their use in streptavidin-based diagnostics.
In the present study, we have used ORFeome phage display to identify L. interrogans peptides that are immunogenic in humans and can be used to develop relatively inexpensive diagnostics tools for this challenging pathogen.
The study protocol NMRR-14-1687-23346(IIR) involved the use of patients’ sera obtained for clinically indicated diagnostics to be used for research purposes and was approved by the Medical Research & Ethics Committee, Ministry of Health Malaysia, Malaysia. All human biosamples were anonymized.
Sera from 34 patients who presented in the year 2015 with symptoms suggestive of acute leptospirosis were obtained from the Public Health Laboratory of Kota Bharu and Kota Kinabalu, Malaysia. The sera were obtained from hospitalized patients after an average of three to seven days of illness and had MAT titers of 1:800, which was consistent with the acute phase of leptospirosis. Infection with other pathogens, including Dengue virus, was excluded by the respective diagnostics as suggested by clinical history and other laboratory findings. Dengue fever is the most common infectious disease in this region and it may cause false positive results in Leptospira spp. ELISA tests due to cross-reactivity. All samples were therefore tested for the Dengue virus NS1 protein and anti-Dengue IgM and IgG reactivity and were thus ensured to be negative.
The patients presented with fever and one or more of these signs and symptoms: headache, myalgia, arthralgia, conjunctival injection, anuria or oliguria and/or proteinuria, jaundice, pulmonary and/or intestinal hemorrhage, cardiac arrhythmia or failure, skin rash, and gastrointestinal symptoms such as nausea, vomiting, abdominal pain, and diarrhea.
Sera from 18 patients were pooled into two groups based on their reactivity to Malaysian strains (n = 8) or WHO reference strains (n = 10). Two sets of Leptospira-negative control sera were used. The control pool used in S2 Fig was obtained from 7 healthy adult volunteers from Germany with no travel history to Leptospira-endemic countries. The other 16 patient sera were used for the titration ELISAs in the ORFeome procedure. The individual control sera for ELISA validation shown in Fig 2A were obtained from 16 healthy adults of Caucasian origin who participated in an unrelated epidemiological study [23]. Absence of leptospiral seroreactivity in the control sera was confirmed by in-house ELISA with leptospiral culture antigens.
Two libraries of Leptospira spp. were constructed using genomic DNA. The first library consisted of five strains of L. interrogans isolated from leptospirosis patients in Malaysia between 2014 and 2015. The second library consisted of seven WHO reference strains which were obtained from the Leptospirosis Reference Centre (also known as OIE Reference Laboratory for Leptospirosis, Amsterdam Medical Centre, Amsterdam). The strains are listed in Table 1. Strains from both groups were cultured in Ellinghausen-McCullough-Johnson-Harris (EMJH) medium at 30°C for 7–10 days at 250 rpm. Genomic DNA was isolated from pellets of 5 mL culture centrifuged at 8000 x g for 30 minutes (min), using the QiaAmp DNA Mini Kit according to the manufacturer’s instructions (Qiagen, Hilden, Germany). The extracted DNA for each library was mixed and amplified with the illustra™ Ready-To-Go GenomiPhi V3 DNA amplification kit (GE Healthcare) according to the manufacturer’s instructions. Twenty μg of DNA from each mixed and amplified genomic library were fragmented by sonication upon extraction. Subsequently, the DNA was concentrated using Amicon Ultra 0.5 mL centrifugal filters with a cut-off of 30 kDa.
DNA fragments with sizes from 100 to 800 bp were extracted from an agarose gel and the DNA ends were repaired with the Fast DNA End Repair Kit (Thermo Scientific) according to the manufacturer’s instructions. 1.4 μg of fragmented DNA were then ligated into 1.4 μg of the PmeI-digested pHORF3 vector [20] and subsequently transformed into E. coli TOP10F’ (Invitrogen) by electroporation. Colony PCR was performed in some of the resulting clones to determine the insert rate of ligation.
The library was packaged using Hyperphage [24, 25] as described before [19, 20]. By packaging the genomic DNA library with Hyperphage, ORFs are enriched and the resulting oligopeptides are presented on the phage particles for panning. The E. coli XL1-Blue MRF’ containing the library was inoculated into 400 mL 2x YT-GA medium (2x yeast-tryptone broth supplemented with 0.1 M glucose and 100 μg/mL ampicillin) to an OD600 <0.1 and grown at 37°C, 250 RPM until OD600 ≈0.5. At this point, the culture was infected with Hyperphage (MOI 1:20) for 30 min at 37°C without shaking, and then 30 min under 250 RPM.
The culture was then centrifuged, suspended in 400 mL 2x YT-AK medium (2x YT containing 100 μg/mL ampicillin and 50 μg/mL kanamycin), and phage particles were produced at 30°C and 250 rpm overnight. Cells were then centrifuged for 20 min at 10,000 x g, and phage particles in the supernatant were precipitated with 1/5 volume of polyethylene glycol (PEG)/NaCl solution (20% w/v PEG 6000), 2.5 M NaCl) for 3 hours (h) on ice with gentle shaking. Phage particles were then pelleted for 1 h at 10,000 x g and suspended in 10 mL phage dilution buffer (10 mM TrisHCl pH 7.5, 20 mM NaCl, 2 mM EDTA). Remaining bacteria were pelleted by an additional centrifugation step of 10 min at 20,000 x g, and the solution was then filtered through a 0.45 μm filter to remove residual bacteria. The filtrate was again precipitated with 1/5 PEG/NaCl for 1 h and then centrifuged for 30 min at 20,000 x g. Pellets were suspended in 1 mL phage dilution buffer and residual bacteria removed by centrifugation for 1 min at 16,000 x g. The final supernatant containing the oligopeptide presenting phages was stored at 4°C. Phage titers were determined as described previously [26].
To check library quality, a number of random E. coli colonies were analyzed by colony PCR and sequenced after packaging with Hyperphage. Therefore, the primers MHLacZPro_f (5'-GGCTCGTATGTTGTGTGG-3') and MHgIII_r (5'-GGAAAGACGACAAAACTTTAG-3') were used with the following PCR protocol: 98°C 30 s, 98°C 10 s, 56°C 20 s (35 cycles), 72°C 60 s and a final extension of 72°C for 2 min. The DNA was separated and analyzed by gel electrophoresis (Qiaxcel Advanced). Additionally, plasmid DNA was sequenced with the primers used for colony PCR to verify the correct inserts and the ORF enrichment after packaging.
Two wells of a Maxisorp™ 96 well microplate were coated with 150 μL of goat IgG directed against human IgG, IgA and IgM Fc (Dianova 109-005-064, 2 mg/mL, Lot No. 113036) diluted in 1:500 phosphate buffered saline (PBS) and another six wells were coated with 5 x 1010 CFU Hyperphage in PBS and incubated at 4°C overnight. Subsequently, the coating solutions were removed and the wells were blocked for 30 min with PBS with 0.1% Tween, supplemented with 2% (w/v) milk powder (2% MPBST).
A pool of twelve patients’ sera was diluted 1:100 in 2% MPBST and pre incubated twice in the Hyperphage coated wells for 1 h to eliminate IgG binding to helper phage. After pre-incubation, the serum pool was incubated for 1.5 h in the wells coated with goat anti-human IgG, A, M antibody. The Leptospira spp. phage library (corresponding to 1.1 x 1010 CFU) was mixed 1:3 with 2% MPBS-T and incubated in the wells with the pooled leptospirosis patients’ sera for 1.5 h. Unbound phage and phage with low affinity were removed by stringent washing steps. Three panning rounds were performed, and the wells were washed twice after each step, with one additional wash for the second panning round. After washing, bound phage particles were eluted with 200 μL of 10 μg/mL trypsin in PBS for 30 min at 37°C. Eluted phages of both wells were combined and 10 μL were used for titration. The remaining 390 μL were used to infect 20 mL of an E. coli TOP10F’ culture grown to an OD600 of 0.5. The cells were incubated for 30 min at 37°C and harvested by centrifugation for 10 min at 3250 x g. The pellet was suspended in 250 μL 2xYT-GA. The bacterial suspension was plated onto 15 cm 2xYT-GA agar plates and incubated overnight at 37°C. Colonies were swept off with 5 mL 2xYT-GA medium, then 50 mL of 2xYT-GA medium was inoculated with the bacterial suspension to an OD600 of <0.1 and grown to an OD600 of 0.5 at 37°C and 250 rpm. For infection, 5 mL of the bacterial culture (approx. 2.5 x 109 cells) was mixed with Hyperphage infected at an MOI of 1:20 resulting in 5 x 1010 CFU Hyperphage. The suspension was incubated at 37°C for 30 min without shaking and another 30 min at 37°C and 250 rpm. To remove glucose, which inhibits phage expression, the infected cells were harvested by centrifugation for 10 min at 3220 x g. The remaining pellet was resuspended in 30 mL 2x YT-AK and incubated at 30°C and 250 rpm overnight for phage production. The bacterial cells were then pelleted by centrifugation for 20 min at 3220 x g and the remaining supernatant was used to precipitate phage particles with PEG/NaCl (20% (w/v) PEG 6000, 2.5 M NaCl). Thirty mL of supernatant were separated and incubated for 1 h with 6 mL PEG/NaCl on ice with slight shaking on a rocker, followed by centrifugation at 6000 x g for 1 h at 4°C. The phage pellet was resuspended in 500 μL phage dilution buffer (10 mM TrisHCl pH 7.5, 20 mM NaCl, 2 mM EDTA), centrifuged in a microcentrifuge at 16,100 x g for 1 min and the supernatant was used for further panning rounds. For the 2nd and 3rd panning rounds, 150 μL of the amplified phage was used. Eluted phage particles from the 3rd panning round were used for titration without further amplification. Single colonies were then used for single oligopeptide phage production.
Single oligopeptide phage clones were produced by inoculating 175 μL 2x YT-GA medium with single colonies from the titration plate in a polypropylene 96-well U-bottom plate (Greiner bio-one). The cultures were incubated at 37°C and 500 rpm shaking overnight.
From this plate, 10 μL were used to inoculate another 165 μL 2xYT-GA medium per well, which was incubated at 37°C and 800 rpm for 2 h. Subsequently, the bacteria were infected with 5 x 109 cfu Hyperphage and incubated for 30 min at 37°C without shaking and 30 min at 37°C and 800 rpm. The bacteria were pelleted by centrifugation at 3220 x g for 10 min and the pellets were resuspended in 175 μL/well 2x YT-AK and incubated overnight at 30°C and 800 rpm. The produced phage in the supernatant were transferred to another plate and precipitated with 1/5 volume of PEG/NaCl solution for 1 h at 4°C. Next, precipitated phage particles were pelleted by centrifugation at 3220 x g for 1 h and the pellets dissolved in 150 μL PBS. Remaining bacterial cells were separated by another centrifugation step and the phage-containing supernatants stored in a new plate at 4°C and used for screening ELISA.
Two types of ELISA were performed for selection of oligopeptide phage clones. First, a screening ELISA was performed on each genomic library. Oligopeptide phage particles were captured by a monoclonal mouse anti-M13 (B62-FE2, Progen) antibody for screening. For this, 100 μL of a 250 ng/mL solution of antibody in PBS were coated overnight at 4°C and subsequently blocked with 2% MPBST. The wells were washed after each incubation step three times with 300 μL PBST. One hundred μL of the monoclonal phage clones were added to each well and incubated for 2 h at 4°C. 100 μL of pooled patients’ sera reactive to Malaysian strains were added to library 1 while pooled sera reactive to WHO strains were added to library 2. All sera were diluted in 2% MPBST supplemented with 10% E. coli TOP10F’ lysate and 1 x 1010 CFU/mL Hyperphage. The dilutions were incubated at RT for 2 h prior to use in the ELISA. Then, the dilutions were added onto the captured phage particles for 1.5 h and detected via a goat anti-human IgG, A, M antibody conjugated to horseradish peroxidase (HRP) (1:20,000) for 1.5 h. Visualization was achieved by adding 100 μL TMB (3,3’,5,5’-tetramethylbenzidine) solution and the reaction was stopped with 100 μL 1 N sulfuric acid. A SUNRISE microtiter plate reader (Tecan, Crailsheim, Germany) was used to measure absorbance at 450 nm and subtract scattered light at 620 nm.
A second titration ELISA was performed on the selected unique oligopeptide clones from the screening ELISA. The method was the same except this time clones were tested for reactivity against each of three aforementioned serum pools. The sera were serially diluted 2-fold from 1:100 to 1:102,400 in 2% MPBST supplemented with 10% E. coli TOP10F’ lysate and 1 x 1010 CFU/mL Hyperphage.
The lipL32 and loa22 genes were amplified from L. interrogans serovar Copenhageni genomic DNA. Phusion DNA Polymerase (Thermo Scientific F-530L) was used to amplify full-length genes according to the following protocol: 98°C 30 seconds (s), 98°C 10 s, annealing temperature primer dependent 20 s, 72°C 20 s, 30 cycles, 72°C 10 min. The amplified genes were digested with Nde1 and Not1 and the resulting fragments resolved by 1% agarose electrophoresis, purified from the gel with NucleoSpin Gel and PCR Clean-Up kit (Macherey-Nagel 740609.250), ligated into the Nde1/Not1 digested vector pET21a(+) and the ligation product subsequently transformed into E. coli BLR (-DE3). Finally, positive clones were identified by colony PCR and confirmed by sequencing.
For protein expression, 200 mL 2xYT-GA medium were inoculated with 10 mL overnight culture and incubated at 37°C and 120 rpm in a baffled flask to an OD600 of 0.6. Expression was induced with a final concentration of 1 mM IPTG for 6 h at 30°C, followed by centrifugation at 3,000 x g for 20 min for cell harvesting. Cells were then suspended in His-tag binding buffer pH 8 with urea (50 mM Na2HPO4, 100 mM NaCl, 10 mM imidazol, 8 M urea) and incubated for 1 h under over-head rotation, followed by sonication (6 cycles of 10 s 50% power, 10s incubation on ice, Sonotrode MS72, Bandelin). Subsequently 0.5 mL Ni-NTA agarose slurry (Qiagen 30210) was added to the disrupted cell solution and incubated for 1 h under over-head rotation. Then, the solution was loaded onto a polypropylene column. The agarose settled by gravity flow and was washed with 10 mM, 30 mM and 50 mM imidazole (50 mM Na2HPO4, 300 mM NaCl, 8 M urea, pH 8). Elution was achieved with 3 x 1.25 mL PBS pH 7.4 supplemented with 100 mM EDTA and 8 M urea.
The proteins were prepared for analysis by 12% SDS-PAGE by heating 0.5 μg of protein sample mixed with 5x Lane Marker Reducing Sample Buffer (Thermo Scientific 39000) at 95°C for 5 min. PageRuler Plus Prestained Ladder (Thermo Scientific 26619) and Spectra Multicolor Low Range Protein Ladder (Thermo Scientific 26628) were used as size marker. The samples were stacked for 10 min at 60 V, followed by separation for 60 min at 110 V. The gels were then stained with Coomassie Brilliant Blue R250 solution dye. The protein bands migrated at ~32 kDa and ~22 kDa.
Leptospira spp. whole-cell antigen for the whole-cell in house ELISA was prepared using the supernatant of live Leptospira spp. cultures. The antigen was prepared and coated to microtiter plates essentially as described by [27].
To confirm the interaction of each isolated peptide with the sera, 200 ng of synthetic peptide (Peps4LS GmbH, Heidelberg, Germany) was diluted in 100 μL of PBS and coated to a high binding 96-well microtiter plate (Greiner-bio one) and incubated at 4°C overnight. Blocking was performed with 2% MPBST for 30 min. Then, 16 sera from patients with acute leptospirosis (MAT titer, 1:800; reactive towards endemic serovars in Malaysia i.e. Australis, Autumnalis, Bataviae, Canicola, Celledoni, Copenhageni, Djasiman, Gryppotyphosa, Hardjobovis, Hardjoprajitno, Icterohaemorrhagiae, Javanica, Lai, Patoc, Terengganu, Sarawak) and 16 sera from healthy donors were serially diluted 2-fold from 1:100 to 1:102,400 in 2% MPBST and added into the wells for 1.5 h. After incubation, the wells were washed three times with 300 μL PBST. Bound IgM was detected with mouse IgG anti-human IgM (CH2)-HRPO, MinX none 100 μg; product no. AFC-5349-2, Dianova) diluted 1:20,000 in 2% MPBST, by incubation for 1.5 h at RT. The reaction was developed with TMB solution, stopped with sulfuric acid, and the plates were read at 450 nm as described before.
Similar ELISA steps were repeated with control proteins (rLipL32 and rLoa22) and Leptospira antigens. For ELISA of control proteins, the sera were diluted and titrated from 1:100 to 1:102,400 in 2% MPBST supplemented with 10% E. coli TOP10F’ lysate and 1 x 1010 CFU / mL Hyperphage and incubated for 2 h prior to use, as done in the ELISA of oligopeptide phage clones described above.
Antigen/antibody ELISA signals in control and disease groups were non-normally distributed, and statistical significance of between-group differences was therefore assessed with the Wilcoxon rank sum (Mann-Whitney U) test [28]. To evaluate discriminatory biomarker potential, a logistic regression model fitted using Bayesian generalized linear models [29] was used to calculate the area under the receiver operating characteristic (ROC) curve (AUC). The AUC and corresponding confidence intervals (CI) values were estimated using the cross-validation procedure based on 1000 bootstrap samples as described before [30, 31]. In addition, multiple logistic regression was used to evaluate the classification performance of each combination of peptides / proteins.
Genomic DNA of Leptospira spp. was fragmented by sonication and cloned into pHORF3, resulting in libraries with 3.0 x 107 (Malaysian local strains library, i.e. library I) and 2.2 x 107 (WHO strains library, i.e. library II) independent clones. The insert rates and sizes were analyzed by colony PCR and sequencing. Thirteen randomly picked clones of each library were used for colony PCR and an additional seven per library were picked for sequencing. The average insert size was 250 bp and 290 bp for library I and library II, respectively. All libraries had insert rates of more than 80%.
Both genomic libraries were then packaged with Hyperphage for the selection of open reading frames. The packaged libraries were checked for the number of inserts by colony PCR and for correct in-frame inserts by sequencing. The cloned libraries had an in-frame insert ratio of approximately 50 to 55% and phage titers of 2.9 x1010 and 4.0 x 1010 cfu/mL. The average DNA fragments were shorter compared to the initial size, i.e. 160 bp and 120 bp for library I and II, respectively.
The pooled sera from patients with acute leptospirosis were used as polyclonal antibody source in the panning rounds and screening. Individual clones were picked after the second and third panning rounds during which interaction partners with low affinities were removed, thus selecting the best interaction partners. The results of this panning are summarized in S1 Fig.
From both libraries combined, this resulted in the selection of 92 oligopeptide phage clones to be screened by ELISA. Of these, 35 had signals two-fold higher than the negative control and were analyzed further by sequencing (Fig 1). Eighteen of these 35 clones were identified as unique sequences and matched Leptospira spp. sequences according to BLAST analysis [32]. Their encoded amino acid sequences were translated using Expasy [33] and the corresponding Leptospira spp. proteins identified using BLAST (Table 2).
After three rounds of panning, screening ELISA were performed on phage clones derived from both libraries. Monoclonal phage clones displaying oligopeptides were captured in the wells of a microtiter plate using a monoclonal anti-M13 antibody and screened for reactivity with pooled sera from 18 patients with acute leptospirosis or from 10 healthy donors and detected with a goat anti-human HRP conjugate (Fig 1).
The BLAST results of 18 unique oligopeptide phage clones corresponded to amino acid sequences of peptides consisting of 13–80 amino acids. Six clones corresponded to hypothetical proteins of unknown function; of these, the most frequent one was LEP1GSC042_0155, which was identical in five clones. Altogether, 18/35 clones (51%) corresponded to hypothetical proteins, which agrees with the report that around 40% of genes of L. interrogans, borgpetersenii and biflexia encode proteins of unknown function [34]. The peptide fragments of known proteins included various enzymes, transporter proteins, and outer membrane protein.
The selected 18 oligopeptide phage clones produced as monoclonal phage were tested in ELISA using serial dilutions of pooled sera reactive with Malaysian strains, WHO strains and from healthy donors as described above. The peptides of the five most promising phage clones were then synthesized as biotinylated peptides and tested for immunoreactivity against 16 positive and 16 negative individual sera which were not included in the aforementioned two serum pools. The two most abundant proteins in pathogenic Leptospira spp., i.e. LipL32 [35] and Loa22 [36], were used as controls.
Based on the signal to noise ratio in the screening ELISA and the BLAST analysis, five peptides (SIR16-A1, SIR16-C1, SIR16-D1, SIR16-E6, SIR16-H1; summarized in Table 3) were selected for further validation. Of note, all five selected peptides were derived from the Malaysian genomic library. A validation test was carried out by a titration ELISA (1:200) using individual sera from 16 patients with acute leptospirosis and 16 healthy donors (Fig 2). The aforementioned two recombinant leptospiral reference proteins, rLipL32 and rLoa22, were included for comparison. The results were also compared to Leptospira culture supernatant antigen used for in-house ELISA. Two peptides (SIR16-D1, SIR16-H1), the two reference proteins, and the leptospiral antigen reacted significantly more strongly with the patient sera than with the control sera, indicating high immunoreactivity (Fig 2A). In addition, signal strength was comparable among the two peptides and reference proteins.
ROC curve analysis revealed that the two peptides (SIR16-H1 and SIR16-D1) had AUCs close to or greater than 0.8 and thus demonstrated potential as diagnostic biomarkers to differentiate between acute leptospirosis and controls (Fig 2B).
Logistic regression was then used to evaluate whether combinations of the peptides with each other and / or with the two recombinant reference proteins would improve classification. First, using a simulation by resampling from the existing dataset [37] we made a posteriori power analysis for the logistic regression model, which includes four predictors and has the highest AUC. For the 32 samples used in the analysis, we detected a power of 98.9% at a significance level 0.001. Among the 5 peptides, a combination with better classification than the best single peptide could not be identified (Table 4, row 3). This was unexpected, as the reactivity with the individual sera correlated only weakly among the peptides, suggesting that there would be diagnostic synergy (S3 Fig). However, when combining the 5 peptides with the two reference proteins, a classifier consisting of the two best peptides, rLipL32 and the peptide SIR16-A1 was identified that possessed near perfect (AUC, 0.98) discrimination between patient and control sera (Table 4, rows 4 and 6). When combining only the two best peptides with the reference proteins, classification was somewhat less accurate (AUC 0.93 vs. 0.98; compare rows 12 and 14 with 4 and 6), demonstrating the added value of including peptide SIR16-A1. Inspection of the data then revealed that SIR16-A1 had a lower reactivity with two of the control sera, likely explaining the observed improved classification in this multiple regression model.
ORFeome phage display has been proven to be a successful method for selection of immunogenic peptides to be used for diagnostic purposes. It has previously been used successfully to identify novel biomarkers from Salmonella typhimurium, Mycoplasma pneumoniae and Neisseria gonorrhea [14, 20, 38, 39]. Regarding Leptospira, it has been used successfully to identify (1) host proteins that interact with LipL32 [40], (2) LigB protein acting as heparin binding protein [41], (3) adhesin activity of Leptospira lipoprotein [42], and (4) mimotopes from monoclonal antibodies specific for Leptospira spp. [43]. Our study is the first to use phage display to identify immunogenic Leptospira antigens from Leptospira spp. genomes.
The findings of this study are of particular interest in that they indicate that ORFeome phage display can be used to identify novel peptides for development of leptospiral diagnostics, an approach that promises to be superior to protein or cell extract based methods for use in tropical and resource-limited settings for the reasons outlined in the Introduction. For instance, an ELISA with directly immobilized peptide would constitute a simple peptide-based point-of-care diagnostics for resource-limited settings [44–47]. A very simple antibody-antigen diagnostic based on specially treated paper to immobilize the antigen and a drinking straw as an incubation chamber has been developed for use in resource-limited setting [48] and might constitute an attractive basis for diagnostics based on peptides identified by ORFeome phage display such as ours.
Even though the two identified peptides demonstrated good diagnostic performance, even higher accuracy would be desirable for clinical application. Combinations of these peptides with proteins of immunodominant properties, such as LipL32 may result in better coverage of pathogenic serovars. However, adding recombinant proteins to the assay would obviate the clear advantages of using peptide-based assays. Thus, to preserve the “peptide only” aspect of a new diagnostic, it will be important to assess the diagnostic value of the two dominant antigenic epitopes of LipL32 [49] in peptide form in order to assess whether combining them with our peptides would be useful. In addition, future work should include additional screens to identify other immunoreactive peptides that could synergize diagnostically with SIR16-D1 and/or SIR16-H1 or be superior by themselves.
Since our goal was to identify peptides for the diagnosis of leptospirosis in the acute phase, we evaluated the peptides for reactivity against IgM only. All patient sera were obtained from patients who presented to the health care system after an average of three to five days of illness symptoms. IgM is the first agglutinating antibody to develop 5 to 14 days after exposure to infection and diagnostically meaningful IgG levels appear 1–3 weeks later [9]. It would now be interesting to test whether these peptides are also reactive with IgG and could therefore be used for seroepidemiological surveys, in addition to diagnostics in the acute phase.
It came as a surprise that all five selected peptides were derived from the Malaysian genomic library, but none from the WHO reference genomic library. This is probably because the Malaysian strains had undergone a smaller number of passages in culture than the WHO strains, thus preserving antigen profiles more closely resembling natural infection. This observation has important implications for future work, as it clearly suggests that early passage strains would constitute a better source for ORFeome phage display libraries than extensively subcultured reference strains.
Peptide SIR16-H1 corresponded to a predicted protein of unknown function. In contrast, peptide SIR16-D1 turned out to be a fragment of Glucose-1-phosphate cytidylyltransferase, also known as CDP-glucose pyrophosphorylase, the product of the rfbF gene [50–52]. This protein is found in 10 pathogenic leptospiral species and also in some other bacteria. Even though this is an intracellular protein, it is quite feasible that it does lead to a humoral immune response as it might become exposed to the immune system during lysis of leptospirae, during phagocytosis by antigen-presenting cells, or even by being secreted from live leptospirae in the sense of a “moonlighting protein” [53].
This study was limited by the number of sera selected for ELISA and also by the prevalence of serovars in two endemic regions in Malaysia, i.e. Kota Bharu and Kota Kinabalu. Besides, the control group was recruited from healthy donors from a non-endemic region. This is because individuals who had been exposed to leptospirosis in endemic region can produce antibodies from the memory pool, leading to background reactivity and false positive results. In fact, it was reported that healthy individuals in high endemic region have a 15% prevalence of positive anti-Leptospira antibodies detected by MAT [54]. Evaluation of the peptides with sera from patients infected with other known tropical infection diseases such as Dengue Fever, Chikugunya, typhoid etc. should be included, as it is important to assess their practicality and specificity as a leptospirosis diagnostic assay in populations exposed to pathogens that may cause serological cross-reactivity. A more comprehensive study involving sera from patients and healthy individuals from various endemic and non-endemic countries should be included in the future, as the present study is based on patient sera from a limited geographical region in Malaysia only and serological responses to field isolates from other geographic areas may differ. All the samples used here were obtained from hospitalized leptospirosis patients, but the immune response may be different in mild presentations. As there are various serovars causing leptospirosis worldwide, we suggest to apply ORFeome phage display screening to genomically more diverse isolates and to human sera collections from various endemic regions for a more universal selection and characterization of the antigen repertoire.
In summary, we report the first study of seroreactive peptides identified by a phage display approach using a combination of endemic Leptospira spp. in Malaysia. The synthetic peptides SIR16-D1 and SIR16-H1 showed good potential for the discrimination of acute phase leptospirosis and healthy patients and can form the basis for the development of peptide-based diagnostics for use in resource-limited settings and hot climates.
|
10.1371/journal.pntd.0003568 | Gastrointestinal Infections and Diarrheal Disease in Ghanaian Infants and Children: An Outpatient Case-Control Study | Diarrheal diseases are among the most frequent causes of morbidity and mortality in children worldwide, especially in resource-poor areas. This case-control study assessed the associations between gastrointestinal infections and diarrhea in children from rural Ghana.
Stool samples were collected from 548 children with diarrhea and from 686 without gastrointestinal symptoms visiting a hospital from 2007–2008. Samples were analyzed by microscopy and molecular methods.
The organisms most frequently detected in symptomatic cases were Giardia lamblia, Shigella spp./ enteroinvasive Escherichia coli (EIEC), and Campylobacter jejuni. Infections with rotavirus (adjusted odds ratio [aOR] = 8.4; 95% confidence interval [CI]: 4.3–16.6), C. parvum/hominis (aOR = 2.7; 95% CI: 1.4–5.2) and norovirus (aOR = 2.0; 95%CI: 1.3–3.0) showed the strongest association with diarrhea. The highest attributable fractions (AF) for diarrhea were estimated for rotavirus (AF = 14.3%; 95% CI: 10.9–17.5%), Shigella spp./EIEC (AF = 10.5%; 95% CI: 3.5–17.1%), and norovirus (AF = 8.2%; 95% CI 3.2–12.9%). Co-infections occurred frequently and most infections presented themselves independently of other infections. However, infections with E. dispar, C. jejuni, and norovirus were observed more often in the presence of G. lamblia.
Diarrheal diseases in children from a rural area in sub-Saharan Africa are mainly due to infections with rotavirus, Shigella spp./EIEC, and norovirus. These associations are strongly age-dependent, which should be considered when diagnosing causes of diarrhea. The presented results are informative for both clinicians treating gastrointestinal infections as well as public health experts designing control programs against diarrheal diseases.
| Gastrointestinal infections are frequent in many low-income countries. However, their role in diarrheal diseases is still under discussion. Many epidemiological studies focus on individuals with diarrheal symptoms only, ignoring the fact that infections may progress asymptomatically as well. In order to identify infectious agents associated with diarrhea it is imperative to consider cases without symptoms as a control group. We conducted a case-control study, including 548 children with diarrhea and 651 children without gastrointestinal symptoms in order to untangle the role of gastrointestinal infections in diarrheal disease. As shown in other studies infections with rotavirus, Shigella spp./EIEC and norovirus are responsible for the main diarrhea burden. Co-infections are frequently observed in our study group and some organisms occur more frequently in the presence of a second one. Especially Giardia lamblia, which is not associated with diarrhea, is more often observed along with Campylobacter jejuni and norovirus, which are responsible for a high number of diarrheal episodes. This may be of particular interest since G. lamblia is, with a frequency of 40% within the study group, the most prevalent organism observed. Furthermore, the high number of co-infections challenged the identification of causative pathogens since diagnosing a particular isolate may not rule out the effect of another potentially infectious agent in diarrheal disease. We observed a strong effect of age on the course of an infection, which may guide clinicians when diagnosing causes of diarrhea.
| Diarrheal diseases are the second leading cause of childhood mortality worldwide. In 2010, diarrhea was responsible for 0.8 million deaths of children below the age of five years, accounting for 10.5% of all deaths within that age group [1]. Mortality and morbidity patterns differ across geographical regions, with 78% of all pediatric diarrhea-associated deaths occurring in the African and South-East Asian World Health Organization (WHO) Regions [2].
The etiology of diarrhea is often not completely understood, especially in developing countries, including those in sub-Saharan Africa. Knowledge of the distribution and impact of infectious agents in diarrheal diseases is crucial in guiding empirical medical treatment and in designing prevention programs. However, many studies on the epidemiology of gastrointestinal infections are restricted to only patients with diarrhea, ignoring the possibility that infections may progress asymptomatically or even influence one another. This may be of particular importance in areas in which certain infectious agents are endemic, which would results in a high probability of ongoing infections after the development of partial immunity and/or tolerance. Thus, an adequate control group is essential to determine the pathogenicity of infectious agents, their fractions attributable to gastrointestinal symptoms (GIS), and the age-dependent association of infectious agents with GIS [3,4].
This hospital-based case-control study in a rural area of Ghana was designed to analyze gastrointestinal infections in children with and without diarrhea. The aims of this study were (i) to identify the causative pathogens linked to diarrhea, (ii) to describe their pathogenicity and contribution to the burden of diarrhea, and (iii) to analyze the frequency and interactive effects of co-infections.
Data were collected at the Agogo Presbyterian Hospital (APH), a district hospital with approximately 250 patient beds located in the Asante Akim North municipality in Ghana. Among other facilities, it has a children’s Outpatient Department (OPD) and a pediatric ward. Asante Akim North municipal area has a population of approximately 142,400 inhabitants, spread over an area of 1,160 square kilometers. The region has a tropical climate and is mainly covered by secondary rain forest and cultivated land [5]. Falciparum malaria is highly endemic in this area [6] and HIV is, with a prevalence of 3.0% in pregnant women in 2009, at a stabile state [7].
This case-control study included children aged up to 13 years who visited APH between June 2007 and October 2008. Stool samples were collected from children with diarrhea attending the hospital’s OPD. Diarrhea was defined as at least three episodes of loose stools within the previous 24 hours. Therefore, a stool container was handed to guardians of cases and controls to collect a sample. In case a child could not provide stool at the OPD, parents were asked to return a sample to the hospital within a day after collection. Assistance was provided if needed (3). The laboratory personnel was required to confirm loose stool consistency of the collected samples. Throughout the study period, each day stool samples were also collected from children who visited the hospital OPD without diarrhea and vomiting, again with the laboratory personnel’s confirmation that stool samples were of solid consistency. Children with ongoing diarrheal disease, defined as repeated hospital visits because of diarrhea within a 6-week period, were excluded from the analysis. Data per individual were used to describe the study groups. Data per hospital visit were used for further analyses.
The Committee on Human Research, Publications and Ethics, School of Medical Science, Kwame Nkrumah University of Science and Technology, Kumasi, Ghana, approved the study design and the informed consent procedure. All participants were informed of the study’s purpose and procedures. Written informed consent was obtained from the parents or the legal guardian on behalf of the study children prior to study enrolment. Non-participation had no effect on the medical treatment provided.
Stool samples were refrigerated (4°C) immediately after collection and transported within a day to the laboratories at the Kumasi Centre for Collaborative Research in Tropical Medicine (KCCR). Sample transport took about 1.5 hours and a cool-box was used to maintain the cold chain. Upon arrival stool samples were aliquoted (3 x 0.2 mg) and frozen at -20°C while the reaming sample was processed for further analyses. Microscopy of blood and stool samples took place at the KCCR laboratory and DNA extractions, polymerase chain reaction (PCR) assays and ELISA-tests were conducted at the Bernhard Nocht Institute for Tropical Medicine (BNITM) in Germany. Therefore, frozen stool samples were transported on dry ice to the BNITM maintaining a temperature of -20°C (3). All samples were handled, stored and transported according to good laboratory practice.
Sodium acetate-acetic acid-formalin (SAF) solution was added to the sample to preserve parasites. Following formalin-ether concentration of protozoan cysts and trophozoites, as well as helminth eggs, fractions of the concentrated material were stained with iodine or by the modified acid-fast method, the latter to improve detection of Coccidia species, such as Cryptosporidium, Cyclospora, and Cystoisospora [8,9]. The samples were then viewed under a light microscope.
For molecular detection, DNA was extracted from the frozen stool samples using QIAamp DNA Stool Kits (Qiagen, Hilden, Germany). Specific sequences were amplified by PCR to identify the following organisms: Campylobacter jejuni [10], Cryptosporidium parvum/hominis [11], Cyclospora cayetanensis [12,13], Entamoeba dispar [14,15], Entamoeba histolytica [14,15], Giardia lamblia [11], norovirus [16], Salmonella enterica [10], Shigella spp./enteroinvasive Escherichia coli (EIEC) [10], and Yersinia enterocolitica [10]. Diagnostic PCR assays give the Ct value as a semi-quantitative result and a cut-off of 35 cycles was applied to determine a positive test result. As Shigella spp. and EIEC both possess the ipaH gene, these organisms cannot be distinguished by PCR [10]. Further, the PCRs applied did not allow differentiation between typhoidal and non-typhoidal Salmonella or between norovirus genotypes. Rotavirus was identified by an enzyme-linked immunosorbent assay (ELISA; Ridascreen).
Blood samples were obtained from all participants by finger prick. Thick and thin smears were prepared from these blood samples, which were stained and examined by microscopy. Malaria was defined as any asexual parasitaemia with body temperature >38°C.
All PCR reactions, as well as malaria and stool microscopy, were regularly evaluated by external and internal quality assessments. The study personal was trained to adhere to the standard operational procedures for each laboratory method.
A sample size of about 500 children per group was estimated to identify isolates in 5% of the controls and 10% of the cases, considering an alpha-level of 5% and a power of 80%. Categorical variables are reported as frequencies and percentages, whereas continuous variables are reported as means ± standard deviations (SDs) or as medians with interquartile ranges (IQRs). Missing values were excluded from analyses, thus the denominators for some comparisons differ.
Because microscopy has decreased sensitivity and specificity in diagnosing diarrheal samples [8,17], all statistical comparisons were based on PCR- or ELISA-based diagnoses. The associations between diarrhea and gastrointestinal infections were determined by calculating odds ratios (OR) and 95% confidence intervals (CI). Subjects were stratified to show effects within categories of a third variable to assess and account for confounding or effect modification. Mantel-Haenszel adjusted ORs (aOR) were calculated from the stratified analyses. The attributable fractions (AF) on the diarrhea burden, defined as the proportion of diarrhea attributable to a certain pathogen, were calculated as described [18], from logistic regression estimates, including dummy variables, for age categories.
Heterogeneity in the occurrence of co-infections was assessed by comparing the probability of organism A in the presence of organism B over the probability of A in the absence of B. These associations were determined by calculating the risk ratio (RR), using the formula RR = P(A|B = 1)/P(A|B = 0), in which a value of about one indicates independence and a value different from one indicating dependence. RRs were calculated for the total study group and for cases and controls separately. Organisms diagnosed by PCR or ELISA and detected in more than 5% of stool samples in the respective study groups were included in this calculation to ensure a sufficient number of co-infections. Age-adjusted RRs (aRR) were calculated to account for the age-dependence of infections. Age was categorized into the groups 0–<1, 1–<2, 2–<5 and 5–15 years to analyze age specific infection dynamics.
All data analyses were performed with STATA 12 (StataCorp LP, College Station, USA).
In total, 1,234 patient visits made by 1,168 children were included in the analysis. The majority of children visited the hospital once (n = 1,109; 94.9%), 52 (4.5%) visited twice and 7 (0.6%) three times. Fifty-seven (4.6%) patients were admitted to the children’s ward; the other children were treated at the OPD. Girls were slightly under-represented (n = 536; 45.9%). The median age of the attendees was 33 months (IQR: 15–70 months). Stratification by age showed that, at the time of visits, 227 (18.4%) children were aged 0 to <1 year, 266 (21.6%) were aged 1 to <2 years, 375 (30.4%) were aged 2 to <5 years, and 366 (29.7%) were aged 5–13 years. Malaria was diagnosed during 236 (20.7%) visits. Differences shown in measles and yellow fever vaccination status were not observed in an age-stratified comparison.
Diarrhea was present in 548 (44.4%) cases (case visits), but absent in 686 (55.6%) instances (control visits). Fever was the most frequent disease symptom and more often observed in controls than in cases [n = 568 (82.8%) vs. n = 397 (72.5%)]. Acute malnourishment was observed in 57 (12.0%) case and 35 (9.4%) control visits., 218 (39.8%) cases suffered from vomiting. The proportion of children with diarrhea decreased gradually with age. The median ages of children with and without diarrhea were 18 months (IQR: 9–36 months) and 57 months (IQR: 26–93 months), respectively (Table 1).
Potentially pathogenic organisms as well as facultative and non-pathogenic parasites were detected 1,843 times in 915 (79.5%) stool samples. The most frequent infections were with G. lamblia (n = 470; 38.1%), Shigella spp./EIEC (n = 336; 27.2%), C. jejuni (n = 242; 19.6%), Blastocystis hominis (n = 144; 14.3%), and norovirus (n = 139; 11.3%). Sex-dependent differences were not observed. All parasites and protozoa, apart from C. parvum/hominis, tended to be less frequently observed in diarrheal compared to non-diarrheal samples. Cyclospora cayetanensis, E. histolytica, and Yersinia spp. were not detected in any of the stool samples tested (Table 2).
Fig 1 shows the age-stratified proportions and the median ages of case and control children infected with particular organisms. In children with diarrhea, rotavirus, norovirus, and C. parvum/hominis were most frequently observed in younger infants, with infected individuals having median ages of 12 months (IQR: 8–23 months), 13 months (IQR 8–22 months), and 14 months (IQR: 10–20 months), respectively. By contrast, control children with these organisms were older, with median ages of 30 months (IQR: 16–45 months), 40 months (IQR: 20–82 months), and 23 months (IQR: 18–28 months), respectively. For all other infections the median age in children with diarrhea ranged from 19 months (IQR: 11–36 months) to 68 months (IQR: 35–123 months). Infants below 6 months accounted for 95 (8.1%) hospital visits, with 16 stool samples (16.9%) diagnosed with rotavirus and 13 (13.7%) with norovirus. All virus-infected children, except one diagnosed with norovirus, had diarrhea. Within this age group other organisms were not detected in more than six stool samples (6.3%).
Crude analyses showed that the strongest positive associations with diarrhea were for infections with rotavirus (OR = 11.9; 95% CI: 6.2–24.9), C. parvum/hominis (OR = 4.3; 95% CI: 2.3–8.6), and norovirus (OR = 2.6; 95% CI: 1.8–3.9). Inverse associations were found for G. lamblia and E. dispar suggesting that more infections were diagnosed in controls than in cases. However, this effect was attenuated after age stratification (Table 3). Age stratification revealed varying associations of most infections with diarrhea. For example, associations with S. enterica or Shigella spp./EIEC increased with age, and associations with norovirus were highest in the youngest age group (OR = 4.7; 95% CI: 1.4–24.6) and lower in older children. Associations between rotavirus and diarrhea were strong amongst all age groups, although the frequency of infections decreased with age. Stratified estimates for C. parvum/hominis were lower than the crude OR, indicating that age confounded effect estimates. No associations between sex (being female) and diarrhea (OR = 0.9; 95% CI: 0.7–1.1) or rainy season and diarrhea (OR = 1.1; 95% CI: 0.9–1.4) were observed. Also, in logistic regression models these factors did not alter the association between infections and diarrheal symptoms.
Infections manifesting at younger ages tended to be more strongly associated with diarrhea. For example, the highest ORs were observed for rotavirus, C. parvum/hominis and norovirus, infectious agents most frequently diagnosed in the younger case age groups, with median ages of 12 months (IQR: 8–23 months), 14 months (IQR: 10–20 months), and 12 months (IQR: 8–23 months), respectively. By contrast, E. dispar and G. lamblia were not positively associated with diarrhea and were more frequently diagnosed in older cases, with median ages around 38 months (IQR: 35–57 months) and 28 months (IQR: 16–46 months), respectively.
The highest AFs (proportion of diarrheal symptoms attributable to a certain organism) were observed for rotavirus (AF = 14.3%; 95% CI: 10.9–17.5%), Shigella spp./EIEC (AF = 10.5%; 95% CI: 3.5–17.1%), and norovirus (AF = 8.2%; 95% CI: 3.2–12.9%), whereas all other infections had AFs of about 5% and lower (Fig 2).
Probabilities for the occurrence of distinct pairs of infectious agents were calculated for organisms detected by PCR or ELISA in more than 5% of the stool samples (Table 4). Considering the total study group, most of the infections occurred independently of other organisms. However, E. dispar, C. jejuni, and norovirus were observed more often in the presence of G. lamblia, showing rate ratios of 1.6 (95% CI: 1.3–1.9), 1.3 (95% CI: 1.2–1.6), and 1.3 (95% CI: 1.1–1.6), respectively. These, estimates were comparable between cases and controls.
The most important cause of diarrheal disease was rotavirus, with both the highest AF and the largest risk for diarrhea across all age groups. Frequency of rotavirus infections decreased with age, but its association with diarrhea was nearly constant throughout all age groups. Similarly, a multi-center study performed in seven resource-poor countries found that rotavirus was the leading cause of diarrhea in infants, with age-dependent AFs between 16% and 28% [3]. Furthermore, 96% of children in a Mexican birth cohort were infected with rotavirus at least once by the age of 2 years. Rotavirus infections conferred protection against re-infection, resulting in less frequent and less severe manifestations in older children [19].
The proportion of norovirus infections among cases and controls was, with 16.6% and 6.8%, respectively, comparable to figures from other high-mortality developing countries, where 14% (CI: 11–16) and 7% (CI: -2–16) are reported, respectively [20]. In children with diarrhea the frequency of norovirus infections was similar to that of rotavirus. However, compared to rotavirus the amount of norovirus infections was higher in asymptomatic (control) individuals, most probably due to increasing pathogen tolerance and limited sterilizing immunity. A transmission model showed that, in highly endemic settings, protection against severe norovirus gastroenteritis could be acquired early in life, resulting in frequent asymptomatic re-infections [21]. Likewise, our study found a significant association between norovirus infection and diarrhea in infants, whereas older children were more likely to be asymptomatic carriers.
In industrialized countries, cryptosporidiosis is primarily an opportunistic infection in HIV/AIDS patients and a major cause of water-born outbreaks reported from several countries [22]. In West Africa, however, cryptosporidiosis is the cause of diarrhea in 4.9% to 14.7% of immunocompetent children, depending on age and geographical location [3,23]. Similarly, our results showed that C. parvum/hominis infections were strongly associated with diarrhea throughout all age groups and was present in more than 10% of symptomatic children below the age of 2 years. Asymptomatic cryptosporidium carriers were not observed in this age group and rarely seen in older children. A review of cryptosporidiosis in sub-Saharan Africa reported the same age distribution, with a peak amongst children aged 6–12 months. Apparently, infection can occur throughout childhood, but symptoms become less severe with age [24].
Shigella spp./EIEC was, after G. lamblia, the second most frequent pathogen identified in children with diarrhea, increasingly occurring in older children. However, comparing our findings with other studies is challenging because our test was PCR based. Traditionally, diagnosing shigellosis relies on culturing techniques, which selectively isolates the pathogen, followed by a biochemical identification of one of the four Shigella species [25]. Introducing PCRs techniques has their benefits but the gene sequence used for the diagnosis is also carried by enteroinvasive Escherichia coli (EIEC) [10]. Technically, our study cannot differentiate between Shigellosis and EIEC, which, in addition to the greater sensitivity of the PCR based test, might also be mirrored by the higher numbers of infections presented. Thus, despite its moderate association with diarrhea, the high prevalence of this pathogen group led to the second highest AF observed in our study.
Generally, the used PCR methods have a higher test sensitivity compared to conventional culture methods. This improves the ability to diagnose organisms in a stool sample. However, a drawback seems to be an increase in asymptomatic detections overall [26]. Since this affects diagnostics in both cases and controls the calculated ORs should not be affected. However, the estimated disease prevalence, and consequently the estimated AF, might be overestimated. Further studies using quantitative approaches [27] are needed to establish and improve diagnostic analyses for gastrointestinal diseases in low- and middle-income countries.
The burden of G. lamblia infections was quite high. Interestingly, the frequency of G. lamblia infections was lower in children with diarrhea than in asymptomatic carriers. A systematic review of the impact of G. lamblia on diarrhea highlighted that, although most studies show no or inverse effects, some studies report positive associations in children aged around 1 year, presumably as a response to initial G. lamblia infections [28]. However, the statistical power of the current study did not allow to disentangle such age-effects.
In both cases as well as controls high numbers of multiple infections were observed. Most co-infections were identified as statistically expected, although G. lamblia was more often found together with E. dispar, C. jejuni, and norovirus. G. lamblia has been reported to induce apoptosis of epithelial cells leading to increased epithelial permeability [29]. Further, G. lamblia was found to secrete proteins capable of impairing the innate immune response [30]. Alternatively, co-infections may be due to shared transmission routes. However, since most gastrointestinal organisms are transmitted via the fecal-oral route, it is unlikely that this alone explains the association among co-infections. Interestingly, a recent pooled case-control study from Ecuador also identified mechanistic interactions for diarrhea symptoms between rotavirus and G. lamblia as well as between rotavirus and Escherichia coli [31]. In vitro models have indicated that rotavirus may foster the adhesion, invasion, and multiplication of bacteria in enteric cells, mechanisms that may explain these synergistic effects [32–34]. Generally, the role of co-infections in diarrheal diseases deserves more attention in order to identify the associations between infections as well as interactions with GIS. Furthermore, strategies to identify causative pathogens in the presence of multiple infections are needed since diagnosing a particular isolate may not rule out other potential infectious causes in diarrheal disease.
The study presented here has several limitations, therefore the results should be interpreted with caution. Cases as well as controls were selected at a hospital OPD, thus the control group does not consist of healthy individuals. These are children seeking help for other health conditions that may increase the risk for gastro-intestinal infections. We have little background information on the total eligible study group, i.e., children that visited the OPD during the study period, from which we selected cases and controls. Thus, we cannot judge how well characteristics of cases and controls match. Table 1 highlights differences between cases and controls. For example, falciparum malaria is more frequently observed in controls, because controls need an alternative reason to attend the hospital, which is malaria in some attendees. Controls are also more likely to have a full vaccination schedule, which can be explained by age differences as well as due to possible differences in socio-economic status. In the analyses these factors cannot be controlled for, however, we believe that they do not act as a confounder. Even though some factors are associated with gastrointestinal infections, they are not associated with diarrheal symptoms, which would be needed to qualify as a confounder. Controls had to be diarrhea free at the point of study enrollment, yet GIS before enrollment were not assessed. Thus, controls could be carriers of pathogens if infections occurred before study enrollment. For instance, norovirus can be found in stool for up to 60 days after infection [35]. In case some controls are pathogen carriers due to recent infections study results would underestimate the actual diarrheal association. Two preconditions need to be fulfilled to generalize AFs from case-control data: (i) the case-control selection must be representative of the source population and (ii) the OR must be a robust estimator of the RR. Cases in this study were recruited from children in the OPD, making this a select group of patients seeking professional care, thereby representing individuals with moderate to severe diarrheal disease. Considering our case-control sampling approach, the OR would approximate the RR only if the rare disease assumption is fulfilled. This, however, was applicable to all infections studied, especially not for Shigella spp./EIEC, C. jejuni, and norovirus. In these cases, the OR is likely to overestimate the true RR, resulting in a higher AF. Several possible infectious causes of diarrhea were not detected by these methods, including adenovirus [36] enteropathogenic Escherichia coli [37] and enterotoxigenic Escherichia coli [37,38].
The AFs express the proportion of diarrheal disease that would be reduced if an organism could be removed. This measure is highly relevant to public health concerns since it demonstrates the potential effects of disease prevention and control as well as empirical disease treatment measures. In particular, it highlights the potential roles of vaccinations against rotavirus and norovirus in sub-Saharan Africa, as well as of water purification, sanitation, and hygiene measures; effective options that can reduce the burden of diarrheal diseases [39].
|
10.1371/journal.ppat.1003952 | AvrBsT Acetylates Arabidopsis ACIP1, a Protein that Associates with Microtubules and Is Required for Immunity | Bacterial pathogens of plant and animals share a homologous group of virulence factors, referred to as the YopJ effector family, which are translocated by the type III secretion (T3S) system into host cells during infection. Recent work indicates that some of these effectors encode acetyltransferases that suppress host immunity. The YopJ-like protein AvrBsT is known to activate effector-triggered immunity (ETI) in Arabidopsis thaliana Pi-0 plants; however, the nature of its enzymatic activity and host target(s) has remained elusive. Here we report that AvrBsT possesses acetyltransferase activity and acetylates ACIP1 (for ACETYLATED INTERACTING PROTEIN1), an unknown protein from Arabidopsis. Genetic studies revealed that Arabidopsis ACIP family members are required for both pathogen-associated molecular pattern (PAMP)-triggered immunity and AvrBsT-triggered ETI during Pseudomonas syringae pathovar tomato DC3000 (Pst DC3000) infection. Microscopy studies revealed that ACIP1 is associated with punctae on the cell cortex and some of these punctae co-localize with microtubules. These structures were dramatically altered during infection. Pst DC3000 or Pst DC3000 AvrRpt2 infection triggered the formation of numerous, small ACIP1 punctae and rods. By contrast, Pst DC3000 AvrBsT infection primarily triggered the formation of large GFP-ACIP1 aggregates, in an acetyltransferase-dependent manner. Our data reveal that members of the ACIP family are new components of the defense machinery required for anti-bacterial immunity. They also suggest that AvrBsT-dependent acetylation in planta alters ACIP1's defense function, which is linked to the activation of ETI.
| How host disease resistance pathways are activated in response to pathogens remains a fundamental question in host-pathogen interactions. In this work, we used the Pseudomonas-Arabidopsis pathosystem to study how the AvrBsT effector activates plant immune signaling. AvrBsT belongs to the YopJ effector family, a group of virulence proteins shared by bacterial pathogens of plants and animals. Bacteria inject these effectors into plant or animal host cells to promote pathogenesis. Recent biochemical studies show that several members of the YopJ family encode acetyltransferases that acetylate host proteins to suppress immune signaling. How the immune system specifically recognizes this family of effectors and/or monitors host acetylation is poorly understood. In this work, we provide biochemical evidence that AvrBsT is an acetyltransferase. We also report the identification and characterization of ACIP1, an Arabidopsis protein of unknown function that is an AvrBsT substrate. We provide evidence that ACIP1 is required for plant immunity and its association with microtubules changes during infection. Moreover, our work suggests that AvrBsT acetyltransferase in planta leads to dramatic changes in ACIP1 localization, which coincides with the activation of strong defense responses. This study highlights an important link between ACIP1 and the microtubule network during anti-bacterial immunity.
| It is well established that bacterial pathogens utilize type III secretion (T3S) systems to translocate virulence factors (referred to as T3S effectors) into eukaryotic hosts to modulate immune signaling during infection [1]. The T3S effector proteome reflects the coevolution of specific host-pathogen interactions as well as microbe-microbe interactions within a given environment. Few T3S effector homologs are conserved among bacterial pathogens that colonize plant or animals hosts. One exception is the YopJ effector family, which is shared by a number of bacterial species in different genera (e.g. Yersinia, Salmonella, Vibrio, Pseudomonas, Xanthomonas, and Sinorhizobium) [2].
The YopJ effector family is named after the archetypal protein YopJ, first identified in Yersinia pseudotuberculosis [3]. These effectors belong to the C55 peptidase family because they share putative structural folds characteristic of cysteine proteases and contain the conserved catalytic triad – His, Glu and Cys [4]. Mutation of this catalytic triad destroyed effector-triggered phenotypes in host cells [5], providing the first clue that enzyme activity is critical for the virulence of the YopJ effector family. Biochemical studies revealed however that YopJ has potent acetyltransferase activity [6]. In subsequent work, several effectors from this family were shown to have acetyltransferase activity important for host-pathogen interactions, including VopA from Vibrio parahemeolyticus [7], AvrA from Salmonella typhimurium [8], PopP2 from Ralstonia solanacearum [9], and HopZ1a from Pseudomonas syringae [10]. These data indicate that a predominant virulence activity for the YopJ effector family is the post-translational acetylation of host proteins.
Resistance to YopJ-like effectors (i.e. AvrBsT, AvrRxv, AvrXv4, HopZ1a, and PopP2) has been reported in several plant hosts [11]–[13]; however, only two disease resistance (R) proteins have been characterized to date [14], [15]. Arabidopsis RRS1-R (for RESISTANCE TO RALSTONIA SOLANACEARUM1) is a Toll-IL-1-receptor-nucleotide binding site-leucine rich repeat-WRKY motif (TIR-NBS-LRR-WRKY)-type R protein that recognizes the PopP2 effector from Ralstonia solanacearum [14]. RRS1-R directly interacts with PopP2 in the plant nucleus [16]. Arabidopsis ZAR1 (for HOPZ ACTIVATED RESISTANCE1) is a coiled-coil (CC)-NBS-LRR-type disease R protein that recognizes the HopZ1a effector from Pseudomonas syringae and activates immune signaling that is distinct from most R protein pathways and independent of salicylic acid [15]. Neither RRS1-R nor ZAR1 were reported to be acetylated by the corresponding acetyltransferase [9], [10] suggesting that acetylation of other plant targets is required for recognition and/or initiation of defense signaling by these R proteins. Interestingly, a recent study revealed that HopZ1a acetylates the Arabidopsis ZED1 (for HOPZ-ETI DEFICIENT1), a pseudokinase that is required for ZAR1-mediated immunity [17]. ZED1 is proposed to act as a decoy in a ZAR1 defense complex.
Notably in mammals, YopJ acetylation suppresses innate immune signaling by exclusively targeting kinases in mitogen-activated protein kinase (MAPK) and/or NF–κB pathways. For example, YopJ catalyzes the O-acetylation of Ser or Thr residues in the activation loop of MAPKK6 [6], MEK2 [18], inhibitor of kappa B kinase [18], and MAP3K transforming growth factor β-activated kinase 1 (TAK1) [19]. Similarly in flies, AvrA inhibits c-Jun N-terminal kinase signaling by O-acetylation of the Thr residue in the activation loop of the MAPKK JNK-K [8].
In plants, a direct link between YopJ-like effector acetylation and suppression of disease resistance has not been made. HopZ1a was reported to acetylate tubulin in vitro, suggesting that the plant cytoskeleton may be disrupted during infection [10]. Consistent with this hypothesis, P. syringae pathovar tomato strain DC3000 (Pst DC3000) infection reduced microtubule density in a HopZ1a catalytic-dependent manner [10]. Interestingly, the mammalian tubulin acetyltransferase TAT1 acetylates Lys40 in α-tubulin (Nε-acetylation) [20], [21] and this modification is commonly found in less dynamic microtubules. The type of tubulin acetylation mediated by HopZ1a in planta has not yet been reported.
In previous work, we exploited the use of the Pseudomonas-Arabidopsis pathosystem to elucidate the biochemical function of the AvrBsT effector from Xanthomonas euvesicatoria. AvrBsT was engineered to be delivered into plant cells by Pst DC3000's T3S system [22] because Arabidopsis is not a host for X. euvesicatoria. Two Arabidopsis ecotypes were identified that differentially respond to Pst DC3000 AvrBsT infection. The Col-0 ecotype is susceptible to Pst DC3000 AvrBsT infection whereas the Pi-0 ecotype is resistant. Pi-0 resistance is due to a recessive, loss of function mutation in SOBER1 (for SUPPRESSOR OF AVRBST-ELICITED RESISTANCE1). SOBER1 encodes a α/β-hydrolase that negatively regulates the accumulation of phosphatidic acid (PA) triggered by AvrBsT activity during bacterial infection [23]. High PA levels in Pst DC3000 AvrBsT-infected Pi-0 leaves correlate with ETI-like defense responses [22], [23]. These data suggest that AvrBsT interferes with lipid homeostasis during infection and that this interference induces strong immune responses in the absence of SOBER1 activity.
Given that PA is a multifunctional stress signal [24], we hypothesized that AvrBsT-triggered PA bursts may directly lead to the local activation of defense signaling. Moreover, we hypothesized that AvrBsT host targets may be linked to the generation or perception of lipid signals during AvrBsT-triggered immunity. To begin to test these hypotheses, we sought to identify AvrBsT interacting proteins from Arabidopsis and elucidate their function(s) in the Pi-0 sober1-1 background [22]. Importantly, the availability of putative host substrates also enabled us to determine if AvrBsT possesses acetyltransferase activity, as reported for other effectors in the YopJ family [6], [9], [10].
Here we report that AvrBsT has acetyltransferase activity. We provide evidence that AvrBsT-dependent trans-acetylation activity is required for the activation of ETI in Arabidopsis Pi-0 leaves and that AvrBsT trans-acetylates Arabidopsis ACIP1 (for ACETYLATED INTERACTING PROTEIN1). ACIP1 is an unknown protein that localizes to punctae on the cell cortex and some of these punctae co-localize with cortical microtubules. We provide evidence that ACIP1 is a new component of the defense machinery required for anti-bacterial immunity. These data support the model that AvrBsT-dependent acetylation in planta alters ACIP1's defense function, which is linked to the activation of ETI.
To identify potential AvrBsT-interacting proteins in Arabidopsis, we performed a yeast two-hybrid screen using the GAL4 DNA-binding domain (BD) fused to AvrBsT (i.e. BD-AvrBsT) and an Arabidopsis cDNA library fused to the GAL4 activation domain (AD). We screened ∼7 million primary yeast transformants and isolated 11 independent clones with a candidate cDNA encoded by At3g09980 (Figure 1A and S1A). Given that AvrBsT is predicted to encode an acetyltransferase, we named the At3g09980-encoded protein ACIP1, for putative acetylated-interacting protein 1 (Figure 1A).
ACIP1 is predicted to encode a protein with 178 amino acids and molecular weight of ∼20.6 kDa. ACIP1's only distinguishing feature is that it is predicted to be a small, α-helical protein [25] that contains the widely conserved domain of unknown function, DUF662 [26]. It was first identified as a tubulin-binding protein [27]. ACIP1 belongs to a small Arabidopsis protein family containing six ACIP-like isoforms (ACIP-L1 to ACIP-L6, Figure S1A). ACIP-L4 and its wheat ortholog TaSRG are required for salt tolerance [28], although their biochemical function(s) are not known. ACIP1 shares 79% identity and 87% similarity with ACIP-L1, the closest isoform. A tree for the Arabidopsis ACIP protein family is shown in Figure S1B. None of the ACIP-like isoforms were isolated in the primary AvrBsT interaction screen.
A candidate yeast interaction screen comparing AvrBsT binding to ACIP1 or the six ACIP-like isoforms revealed that AvrBsT strongly interacts with ACIP1 but only weakly interacts with ACIP-L1 on selection media containing 1 mM 3-AT (Figure S1C,D). In the presence of 5 mM 3-AT, AvrBsT only interacted with ACIP1 (data not shown). Taken together, these data suggest that AvrBsT preferentially binds to ACIP1 in yeast.
Next, we used GST pull-down assays to independently monitor the physical association of AvrBsT and ACIP1 in vitro. Recombinant GST and GST-AvrBsT were expressed in E. coli and then purified using glutathione sepharose. Purified GST-AvrBsT migrated as a doublet in protein gels, suggesting that proteolysis of the full-length polypeptide likely occurred during extraction and/or affinity purification. His-tagged ACIP1 was expressed in E. coli and soluble protein extracts were incubated with the GST or GST-AvrBsT in a standard GST pull-down assay. His6-ACIP1 was affinity purified by GST-AvrBsT but not GST (Figure 1B). These findings are in agreement with the yeast two-hybrid data and provide additional evidence that AvrBsT interacts with Arabidopsis ACIP1.
We attempted to verify AvrBsT-ACIP1 physical interaction in planta; however, the assays were not successful. Transient or inducible expression of AvrBsT in Arabidopsis Pi-0 leaves or Nicotiana benthamiana leaves results in localized cell death. It was difficult to obtain reproducible, conclusive binding data under these cellular conditions.
AvrBsT belongs to the YopJ family of T3S effector proteins, some of which have been shown to exhibit acetyltransferase activity [6], [9], [10]. To ascertain if AvrBsT acetylates ACIP1, we first sought to determine if AvrBsT possesses auto-acetylation activity in vitro. Recombinant wild-type GST-AvrBsT, GST (negative control) and GST-HopZ1a (positive control) [10] were over-expressed in E. coli and then purified using glutathione sepharose. Purified proteins were incubated with 14C-acetyl-coenzyme A (acetyl-CoA) ±100 nM inositol hexakisphosphate (IP6) for 30 minutes at room temperature and then separated by SDS-PAGE analysis followed by autoradiography. IP6 is a eukaryotic cofactor that stimulates the acetyltransferase activity of effectors in the YopJ family [9], [10], [29].
Auto-acetylation of GST-AvrBsT was detected in the presence of IP6 but not its absence (Figure 2A and S2). As expected, similar IP6-dependent activation and auto-acetylation of GST-HopZ1a was observed, and GST was not modified (Figure 2A and S2). Mutation of the conserved catalytic Cys residue (C222) or His residue (H154) to Ala inactivated AvrBsT-dependent acetyltransferase activity but did not affect protein expression levels (Figure 2A). By contrast, mutation of the conserved Lys residue (K282) (Figure S3A) to Arg, which has been shown to be an auto-acetylation site for some effectors in the YopJ family [9], [10], did not affect AvrBsT's acetylation state or protein accumulation (Figure 2A). The auto-acetylation activity of GST-AvrBsT(K282R) was comparable to that of wild-type GST-AvrBsT in reactions with varying concentrations of enzyme (Figure S3B). All GST-AvrBsT protein (wild type and mutant) analyzed migrated as a doublet and both of these species were auto-acetylated (Figure 2). Taken together, these data indicate that AvrBsT possesses auto-acetylation activity in vitro that is dependent on the conserved catalytic residues H154 and C222, but this activity is independent of K282.
Next, we tested if AvrBsT directly acetylates ACIP1 using similar reaction conditions to those described above. Wild-type GST-AvrBsT activity resulted in auto-acetylation of the enzyme and trans-acetylation of GST-ACIP1 (Figure 2B), whereas the catalytic core mutants GST-AvrBsT(C222A) or GST-AvrBsT(H154A) exhibited neither activities (Figure 2B). Although the GST-AvrBsT(K282R) mutant possessed auto-acetylation activity, trans-acetylation of ACIP1 was not detected under the same reaction conditions (Figure 2B). Importantly, mutation of C222 or K282 did not disrupt AvrBsT binding to ACIP1 in vitro (Figure S4).
To gain insight to the specificity of acetyltransferases in the YopJ effector family, we determined if HopZ1a could acetylate AvrBsT's substrate ACIP1. Conversely, we determined if AvrBsT could acetylate HopZ1a's substrate tubulin [10]. Incubation of GST-HopZ1a with GST-ACIP1 did not result in detectable acetylation of ACIP1 (Figure 2C). Moreover, neither HopZ1a nor other members of the HopZ family could physically associate with ACIP1 in targeted yeast two-hybrid screens (Figure S5A, B). Similarly, we could not detect AvrBsT-dependent acetylation of tubulin in vitro (Figure S5C) or direct physical interaction between AvrBsT and tubulin in yeast (Figure S5A,B). These data suggest that AvrBsT and HopZ1a possess distinct substrate specificity.
We assessed the biological activity of the AvrBsT(K282R) mutant in Arabidopsis Pi-0 leaves, given that mutation of the analogous Lys residue in PopP2 and HopZ1a inhibits effector auto-acetylation activity and effector-dependent phenotypes in planta [9], [10]. Bacterial growth curve analyses showed that the K282R mutation attenuated the ability of AvrBsT to activate defense in Pi-0, similar to that observed for the H154A mutation in the catalytic core (Figure 3A). Furthermore, Pst DC3000 expressing AvrBsT(K282R) did not elicit HR in Pi-0 leaves (Figure 3B) despite stable protein expression (Figure S3C). These data indicate that the K282R mutation affects AvrBsT's trans-acetylation activity in vitro (Figure 2B) and its defense eliciting activity in planta (Figure 3). Moreover, these data indicate that the auto-acetylation activity of AvrBsT(K282R) is not sufficient to activate ETI in Arabidopsis.
Given that nothing was known about ACIP1 function, we first sought to elucidate its potential role in immunity. Previously we showed that the Pi-0 ecotype of Arabidopsis is resistant to Pst DC3000 expressing AvrBsT, whereas the Col-0 ecotype is susceptible [22]. Interestingly, ACIP1 mRNA abundance was significantly reduced at 3 and 6 HPI in Pi-0 (Figure S6A) and Col-0 (data not shown) leaves inoculated with a 2×108 cells/mL suspension of Pst DC3000 or Pst DC3000 AvrBsT compared to leaves inoculated with mock solution of 1 mM MgCl2. By contrast, endogenous ACIP1 protein levels appeared to remain constant (Figure S6B). These data suggest that ACIP1 may be transcriptionally or post-transcriptionally regulated during pathogen attack and potentially linked to PTI and/or ETI.
To explore this further, we first analyzed the growth of virulent Pst DC3000 in a homozygous Col-0 acip1 null mutant (SALK_028810) to determine if ACIP1 is required to limit pathogen growth. Pst DC3000 grew equally well in wild-type Col-0 and acip1 mutant leaves (data not shown). Similar results were observed when the acip1 null allele was crossed into the Pi-0 background (data not shown). We speculated that the lack of a bacterial growth phenotype in the Col-0 acip1 and Pi-0 acip1 mutants may be due to genetic redundancy since ACIP1 belongs to a small gene family in Arabidopsis (Figure S1A,B). Since the nucleotide sequences between ACIP1 and ACIP-like genes are highly conserved (Figure S7A), we engineered RNAi lines to target multiple ACIP family members in attempt to uncover an immune phenotype linked to this gene family. Notably, we silenced the ACIP gene family in the Pi-0 background to be able to monitor both PTI and ETI, considering that AvrBsT induces ETI in the Pi-0 ecotype but not the Col-0 ecotype [22]. A 365-bp hairpin ACIP binary construct (hp-ACIP) was designed using the ACIP1 gene, which included the most conserved region shared by the entire gene family (Figure S7A,B), and then it was transformed into Pi-0 plants. Five independent transgenic RNAi lines were characterized. The hp-ACIP construct significantly reduced the mRNA levels for 4 of the 7 family members (i.e. ACIP1, ACIP-L1, ACIP-L2, and ACIP-L3) in two T2 ACIP RNAi lines (i.e. lines 1 and 29; Figure S7C). Of these 4 genes, ACIP1 mRNA was the most abundant transcript in 4-week old Pi-0 leaves (Figure S7D), suggesting that it may be the major isoform expressed in leaves.
To monitor ACIP1 protein expression in leaves, we generated rabbit polyclonal antisera using recombinant ACIP1-His6 protein purified from E. coli. The resulting antisera recognized multiple, recombinant purified ACIP isoforms with distinct molecular weights by immunoblot analysis (data not shown). However in wild-type Pi-0 leaf extracts, the antisera only detected a single 20 kDa protein band (Figure 4A, inset). Three of the isoforms have predicted molecular weights in this range: ACIP1 = 20.5 kDa, ACIP-L1 = 20.2 kDa, and ACIP-L3 = 20.9 kDa. The 20 kDa protein band was not detected in the two ACIP RNAi lines (Figure 4A, inset) suggesting that ACIP1, ACIP-L1 and/or ACIP-L2 protein accumulation was significantly reduced.
Bacterial growth curves were then performed using a 1×105 cells/mL suspension of Pst DC3000 expressing AvrBsT and the two Pi-0 ACIP RNAi transgenic lines to determine if ACIP expression is required for AvrBsT-triggered ETI. The phenotypes of the ACIP-silenced lines were compared with an unsilenced Pi-0 control plant (Figure 4).
We found that the titer of Pst DC3000 AvrBsT was significantly higher in infected Pi-0 ACIP RNAi leaves compared to that in wild-type Pi-0 leaves (Figure 4A). Notably, the Pi-0 ACIP RNAi leaves were also more susceptible to Pst DC3000. These data suggested that the silenced ACIP isoforms might function in PTI as well as ETI.
To confirm that AvrBsT-triggered ETI is impaired in the RNAi lines, we performed HR and electrolyte leakage assays in leaves challenged with a high titer (3×108 cells/mL) of Pst DC3000 AvrBsT. ETI in the Pi-0 ACIP RNAi lines was delayed but not fully inhibited (Figure 4B). In control Pi-0 leaves, AvrBsT-dependent HR was visible at 9 HPI in many leaves and by 10 HPI, 22/25 leaves exhibited HR. By contrast, HR was not observed in similarly inoculated RNAi leaves at 9 HPI; however, HR started to develop at 10 HPI in 14/25 leaves for line 1 and 12/25 leaves for line 29. Consistent with these findings, electrolyte leakage was significantly reduced in the Pst DC3000 AvrBsT-inoculated Pi-0 ACIP RNAi leaves relative to the inoculated Pi-0 leaves at 10 HPI (Figure 4C). These data suggest that multiple ACIP isoforms are required for AvrBsT-triggered ETI symptoms in Pi-0.
The Pi-0 ACIP RNAi lines were also examined for their ability to mount ETI in response to two other Pseudomonas effectors – AvrB and AvrRpt2 [30], [31]. As observed for AvrBsT, HR symptom development was slower in the RNAi lines infected with a high titer of Pst DC3000 AvrB or Pst DC3000 AvrRpt2 (data not shown). Subsequent bacterial growth curve analyses revealed that the Pi-0 ACIP RNAi lines were more susceptible to both Pst DC3000 AvrB and Pst DC3000 AvrRpt2 (Figure S8). These data suggest that the ACIP isoforms play a general role in ETI and are not specific to defense responses elicited by AvrBsT.
Given that the RNAi lines were also more susceptible to Pst DC3000, we next examined the potential role of the ACIP family in PTI. First, we analyzed the responsiveness of the Pi-0 ACIP RNAi lines to the PAMP elicitor flg22 (Figure 5). Perception of flg22 by the PRR FLS2 results in the production of reactive oxygen species (ROS) [32], one the first measurable PTI responses, followed by changes in PTI gene induction [33]. Flg22-induced ROS production was significantly reduced in both Pi-0 ACIP RNAi lines (Figure 5A). Similarly, flg22-induced mRNA accumulation for WRKY22 and WRKY29, two genes encoding transcription factors that positively regulate PTI, was significantly reduced at 3 hr post-treatment in both Pi-0 ACIP RNAi lines (Figure 5B). Consistently, the RNAi line 29 exhibited the least responsiveness to flg22 elicitation (Figure 5A,B).
We also examined the responsiveness of the Pi-0 ACIP RNAi lines to Pst DC3000 ΔhrcU, a Pseudomonas strain known to elicit PTI. Pst DC3000 ΔhrcU lacks a functional T3S apparatus [34] and does not suppress PTI because T3S effectors are not secreted. Leaves were infected with a 1×105 cells/mL suspension of bacteria and titers were determined at 4 DPI. Pi-0 ACIP RNAi leaves were significantly more susceptible to Pst DC3000 ΔhrcU (Figure 5C). Consistent with these findings, accumulation of WRKY22 and WRKY29 mRNA was significantly reduced at 6 HPI in Pi-0 ACIP RNAi leaves compared to wild-type Pi-0 leaves inoculated with a high titer (2×108 cells/mL suspension) of Pst DC3000 ΔhrcU (Figure 5D). Taken together, these data suggest that a subset of the ACIP family (i.e. ACIP1, ACIP-L1, ACIP-L2, and ACIP-L3) collectively contribute to anti-bacterial immunity in Arabidopsis.
To begin to address ACIP1's function, we examined ACIP1 protein localization in Arabidopsis seedlings and mature plants. We generated homozygous transgenic Pi-0 plants expressing a GFP-ACIP1 protein fusion under the control of the native ACIP1 promoter (i.e. PACIP1::GFP-ACIP1). Using confocal microscopy, we observed a low level of GFP-ACIP1 fluorescence in 4-day old etiolated seedlings and juvenile leaves. Little or no detectable fluorescence was observed in mature, senescing leaves. In hypocotyl epidermal cells, GFP-ACIP1 was predominantly found as punctae at the cell cortex. A portion of these punctae was aligned, forming transverse cable-like structures (Figure 6A). ACIP1's subcellular localization pattern partially resembled that of cytoskeletal structures. Unlike TaSRG, the ACIP-L4 ortholog, GFP-ACIP1 was not observed in the plant nucleus, indicating that ACIP1 localization is distinct from this predicted transcription factor [28].
Next, we applied drugs to disrupt the cytoskeleton to determine if ACIP1 co-localizes with actin and/or microtubules. Treatment of the Pi-0 PACIP1::GFP-ACIP1 seedlings with oryzalin, a microtubule depolymerizing agent, disrupted the GFP-ACIP1 cable-like structures and caused the formation of numerous GFP-ACIP1 punctae throughout the cell (Figure 6B). By contrast, the actin depolymerizing agent latrunculin B did not appear to significantly disrupt these cables (Figure 6B).
To show ACIP co-localization with microtubules, we transformed the Pi-0 PACIP1::GFP-ACIP1 lines with P35S::mCHERRY-TUA5, a fluorescently tagged isoform of α-tubulin. A large portion of the GFP-ACIP1 punctae co-localized with mCHERRY-TUA5 microtubules (Figure 6A). Some of the cortical GFP-ACIP1 punctae were not associated with microtubules. Inspection of the literature revealed that ACIP1 was identified in the Arabidopsis proteome that co-purified with tubulin by affinity chromatography [27]. We did not detect a direct interaction between ACIP1 and TUA5 in a targeted yeast two-hybrid assay (Figure S5A,B). It is possible that ACIP1 association with tubulin might be indirect or via a weak electrostatic interaction. Or, ACIP1 might interact with another isoform of tubulin. Collectively, our findings indicate that GFP-ACIP1 signal forms punctae on the cell cortex and some of these punctae co-localize with the cortical microtubule network.
We speculated that AvrBsT-binding to and acetylation of ACIP1 might interfere with ACIP1's stability and/or localization within plant leaves during pathogen infection. We did not detect by protein gel blot analysis any differences in endogenous ACIP1 protein abundance or mobility using extracts isolated from Pi-0 leaves infected with Pst DC3000 or Pst DC3000 AvrBsT (Figure S6B). However, we did notice that GFP-ACIP1 localization in 4-week old Pi-0 PACIP1::GFP-ACIP1 leaves was dramatically altered in response to both Pst DC3000 and Pst DC3000 AvrBsT (Figure 7). Unlike the signal in young hypocotyls, GFP-ACIP1 fluorescence at the cortex of epidermal cells in 4-week old leaves inoculated with the 1 mM MgCl2 mock control was diffuse and faint. This signal was difficult to capture in the image projection and varied among plants. By contrast, GFP-ACIP1 punctae were observed at or near the cell periphery of these cells (Figure 7A). Infection with Pst DC3000 for 6 h led to the formation of rod-shaped GFP-ACIP1 structures of various lengths (Figure 7B), which were difficult to detect in the 1 mM MgCl2 mock control (Figure 7A). The GFP-ACIP1 rods were also detected in response to Pst DC3000 ΔhrcU (Figure 7C), indicating that these structures are associated with PTI in a T3S effector-independent manner. Strikingly, Pst DC3000 AvrBsT infection for 6 h led to the formation of large, bright GFP-ACIP1 aggregates and fewer rod-like structures (Figure 7D). This localization pattern was dependent on AvrBsT's catalytic activity. Pst DC3000 AvrBsT (H154A) infection resulted in a GFP-ACIP1 signal (Figure 7E) similar to that induced by Pst DC3000 alone (Figure 7B). Similarly, Pst DC3000 AvrBsT(K282R) infection led to the formation of GFP-ACIP1 rods (Figure 7F), but not large aggregates.
To determine if GFP-ACIP1 aggregates are generated specifically by AvrBsT and/or in response to PA production, we tested the phenotype of Pst DC3000 AvrRpt2. AvrRpt2 elicits ETI in Pi-0 leaves [23], which is dependent on PA production [23], [35]. Infection with Pst DC3000 AvrRpt2 induced the formation of GFP-ACIP1 rods (Figure 7G), similar to those formed in response to Pst DC3000 alone, Pst DC3000 AvrBsT(H154A), and Pst DC3000 AvrBsT(K282R) (Figure 7B,E,F). However, ACIP1 aggregates were not observed in response to Pst DC3000 AvrRpt2. Pi-0 PACIP1::GFP-ACIP1 leaves were inoculated with 50 µM PA and then imaged the leaves 1.5 hr later. Exogenous PA triggered the formation of several GFP-ACIP1 rods but only a few punctae (Figure S8C), whereas the mock control containing 0.2% DMSO did not (Figure S8D). These data suggest that PA exposure is sufficient to promote the formation of ACIP1 punctae and rods, but not the formation of ACIP1 aggregates. Moreover, they indicate that ACIP1 aggregate formation is a specific phenotype linked to AvrBsT acetyltransferase activity in planta.
Pathogen-dependent acetylation of host targets has emerged as a key virulence strategy to alter eukaryotic defense responses. The study of several YopJ and YopJ-like effectors in animals and flies indicates that O-acetylation of Ser/Thr residues or Nε-acetylation of Lys residues in the activation loop of kinases in innate immune pathways directly interferes with residue phosphorylation or ATP binding, respectively [6]–[8], [18]. Both scenarios prevent the activation of kinases that are required to mediate innate immune signal transduction. In plants, the mechanism(s) by which YopJ-like effector acetylation of host substrates modulates immune signaling is less clear.
Based on this study, we propose that the YopJ-like effector AvrBsT acetylates Arabidopsis ACIP1. The role of ACIP1 in planta is not known; however, it is predicted to be a small α-helical protein [25]. ACIP1 emerged from an Arabidopsis screen looking for tubulin-binding proteins [27], suggesting that it might be a microtubule-associated protein. Our microscopy studies of Arabidopsis Pi-0 lines expressing GFP-ACIP1 revealed that ACIP1 is localized to punctae on the cell cortex and some of these punctae co-localize with the cortical microtubule network (Figure 6). These data are consistent with ACIP1 being a part of the tubulin proteome [27].
Importantly, we discovered that GFP-ACIP1 organization and accumulation changed significantly during bacterial infection (Figure 7). Numerous small GFP-ACIP1 punctae and rod-like structures formed throughout the cell in response to Pst DC3000 infection. These structures also formed during Pst DC3000 ΔhrcU infection, indicating that changes in ACIP1 localization are coincident with PTI. Strikingly, Pst DC3000 AvrBsT infection, but not Pst DC3000 AvrRpt2 infection, dramatically altered GFP-ACIP1 localization. AvrBsT activity triggered the accumulation of large GFP-ACIP1 aggregates throughout the plant cell. The aggregates did not appear to be aligned like those observed in leaves infected with Pst DC3000 or uninfected hypocotyl epidermal cells (Figure 6). Interestingly, a mutation (K282R) that disrupted AvrBsT's ability to acetylate ACIP1 in vitro (Figure 2B) also prevented the formation of GFP-ACIP1 aggregates (Figure 7) and activation of ETI during Pst DC3000 AvrBsT infection (Figure 3). Taken together, these data suggest the model that AvrBsT acetyltransferase activity in planta uniquely alters ACIP1's localization, which is linked to AvrBsT-dependent activation of ETI.
The nature of the large GFP-ACIP1 aggregates and their function during pathogen infection remains to be determined. Given the requirement for ACIP1 for both PTI (Figure 5) and the formation of ACIP1 punctae and rods during Pst DC3000 infection (Figure 7), we speculate that ACIP1 association with microtubules and/or the cell cortex is important for plant immunity. We also speculate that AvrBsT acetyltransferase activity either directly or indirectly alters ACIP1 association with microtubules. Association of ACIP1 with microtubules may play a direct role in microtubule organization, or it may be involved in microtubule-dependent processes such as vesicle and protein trafficking. Alternatively, ACIP1 may simply use microtubules to position itself and its interacting proteins at the cell cortex, where plant cells first encounter injected bacterial effectors. Our ACIP RNAi plants (silenced for ACIP1, ACIP-L1, ACIP-L2, and ACIP-L3, Figure S7) did not show cell shape or cell growth phenotypes, which are caused by microtubule cytoskeleton defects, suggesting that four ACIP family members do not regulate microtubule cytoskeleton structure. Future functional studies will test if ACIP1 and/or other isoforms expressed in leaves play a role in suppressing bacterial growth by regulating microtubule-dependent trafficking or by regulating other processes at the plasma membrane or cell cortex.
Notably, AvrBsT catalysis in Arabidopsis Pi-0 leaves leads to the accumulation of PA, a lipid signal associated with plant adaptation to biotic and abiotic stress [24]. Elevated PA levels in Pi-0 leaves inoculated with Pst DC3000 AvrBsT correlate with ETI [22], [23]. ACIP1 is required for AvrBsT-triggered ETI (Figure 4); however, the causal relationship between changes in PA production and ACIP1 localization in response to AvrBsT acetyltransferase activity is not clear. Furthermore, it is not known if PA is required to alter ACIP1 localization and/or function. Exogenous PA treatment (Figure S8C) or infection with Pst DC3000 AvrRpt2 (Figure 7G), which triggers a PA burst during ETI [35], induced the formation of small ACIP1 punctae and rods of various sizes, but large ACIP1 aggregates similar to those formed in response to AvrBsT-dependent catalysis (Figure 7D) were not observed. These data further highlight the specificity for AvrBsT in inducing ACIP1 aggregation during infection. They also suggest that a threshold concentration of PA or local production of PA relative to ACIP1 might be required to trigger the formation of large ACIP1 aggregates in planta.
PA is known to play a critical role in the regulation of cytoskeletal dynamics [36]. Recent data suggests that PA alters the microtubule network by directly binding to cytoskeletal components, including tubulin and the microtubule bundling protein MAP65-1 [37], [38]. Interestingly, elevated PA resulting from salt stress recruits Arabidopsis MAP65-1 to the membrane and enhances its ability to stabilize microtubules, which promotes cell survival [38]. How PA directly alters the microtubule network during ETI is not known. The link between PA, ACIP1, and the microtubule network during pathogen infection established in this study suggests that PA might regulate ACIP1 complex formation and/or association with microtubules.
Interestingly, the HopZ1a acetyltransferase was recently shown to disrupt plant cortical microtubule arrays and secretion during bacterial infection [10]. In this case, P. syringae HopZ1a infection led to reduced microtubule density, suggesting that HopZ1a acetylation in planta affects the stability or nucleation of microtubules. Acetylation of mammalian EB1, a microtubule-associated protein, which promotes microtubule assembly, was recently shown to compromise EB1 binding to other microtubule plus-end tracking proteins [39]. HopZ1a binds and acetylates tubulin in vitro. Whether or not HopZ1a modifies tubulin and/or affects microtubule properties (i.e. assembly, disassembly, and/or stability) during infection remains to be determined.
In terms of acetylation, our data suggests that AvrBsT trans-acetylation activity, not auto-acetylation activity, triggers ETI in Pi-0 leaves (Figure 3). Mutation of Lys 282 to Arg in AvrBsT, a conserved residue found in YopJ and YopJ-like effectors [9], did not affect AvrBsT's auto-acetyltransferase activity in vitro (Figure 2A), although it inhibited its ability to trans-acetylate ACIP1 (Figure 2B). We speculate that K282 is required for enzyme-substrate interactions, although acetyl-CoA docking or direct acetylation [9] is also possible. Importantly, Pst DC3000 expressing the AvrBsT(K282R) mutant failed to trigger ACIP1 aggregates (Figure 7) and elicit host resistance (Figure 3). These data suggest that acetylation is linked to changes in ACIP1 function and immunity. Whether or not acetylation of ACIP1 is directly linked to punctae formation, localization with microtubules, PA production and/or the activation of AvrBsT-triggered ETI awaits characterization of ACIP1's acetylation status in planta. Since acetylation can increase the electronegativity of proteins, it has the potential to disrupt ACIP1 interactions with the negatively charged microtubule lattice. Electrostatic interactions have been shown to play a significant role in microtubule binding of motor proteins and microtubule-associated proteins that often possess domains enriched in positively charged residues [40]–[42]. Future mapping of the ACIP1 microtubule-interaction domain in relation to residues acetylated by AvrBsT will allow us to test the functional significance of ACIP1 acetylation in planta.
A growing number of plant targets have been identified for YopJ-like effectors, questioning the specificity of these enzymes as acetyltransferases versus binding partners in immune complexes. Our work indicates that there is selectivity between AvrBsT and HopZ1a in vitro. AvrBsT acetylates ACIP1 whereas HopZ1a acetylates tubulin. In addition to ACIP1, AvrBsT has been recently shown to bind to arginine decarboxylase (ADC1), an enzyme proposed to mediate polyamine and γ-aminobutyric acid metabolism and impact cell death responses [43], and SNF-1 related kinase (SnRK1), a putative regulator of sugar metabolism [44]. Post-translational acetylation of these plant proteins has not yet been reported. The fact that a number of metabolic enzymes are normally regulated by acetylation warrants further investigation [45].
Similarly, HopZ1a appears to have multiple plant targets. In addition to tubulin, HopZ1a was shown to acetylate Arabidopsis ZED1, a pseudokinase required for HopZ1a-dependent ETI [17], and Arabidopsis jasmonate (JA) ZIM-domain proteins required to repress JA signaling during PTI [46]. HopZ1a was also shown to bind and destabilize an enzyme involved in isoflavonoid biosynthesis, 2-hydroxyisoflavanone dehydratase (HID1), by an unknown mechanism [47]. The diverse nature of these targets suggests that HopZ1a is a promiscuous enzyme capable of altering defense signaling at multiple nodes.
It is intriguing that HopZ2, the closest YopJ-like homolog to AvrBsT [48], was found to directly interact with Arabidopsis MLO2 in planta [49]. Arabidopsis mlo2-7 mutants are compromised for HopZ2-dependent virulence, further supporting the role of MLO2 as a negative regulator of immunity [49]. MLO2 is a plasma membrane protein of unknown function that interferes with vesicular trafficking mediated by the syntaxin PEN1 [50], [51]. It is too early to tell if there is a common theme for YopJ-like targets in plant cells. However, the identification of ACIP1, tubulin, and MLO2 as host targets suggests that some YopJ-like effectors might have undergone specialization to interfere with the trafficking function of the microtubule cytoskeleton in infected cells.
Does AvrBsT target ACIP1 or an ACIP1 complex to suppress immunity? This question has been difficult to answer because we have yet to detect an AvrBsT virulence phenotype in Arabidopsis during bacterial infection. This is not so surprising given the potential functional redundancy between AvrBsT and the suite of T3S effectors in Pst DC3000. Overexpression of AvrBsT in transgenic Arabidopsis lines however was recently shown to enhance susceptibility to Pst DC3000 [52]. In solanaceous plants, AvrBsT is known to suppress PTI in tomato [53] and ETI in pepper [44] during Xanthomonas infection. The study of the ACIP1 ortholog in tomato may provide insight to AvrBsT virulence, by specifically addressing how AvrBsT acetyltransferase activity interferes with ACIP1's function during PTI and/or ETI.
In summary, the study of AvrBsT-triggered defense responses in Arabidopsis Pi-0 plants has led to the identification of ACIP1, a member of a new protein family required for PTI and ETI. We demonstrate that the expression of four Arabidopsis ACIP isoforms (ACIP1, ACIP-L1, ACIP-L2, and ACIP-L3) is required for proper execution of PTI in response to Pst DC3000 (Figure 5) and ETI in response to Pst DC3000 expressing AvrBsT, AvrB or AvrRpt2 (Figure 4, S8). In addition, we show that AvrBsT is an acetyltransferase and provide evidence that acetyltransferase activity plays an important role in altering ACIP1 localization within the plant cell during infection and the activation of ETI. This study highlights an important link between ACIP1 and the microtubule network during plant defense.
Escherichia coli DH5 alpha and Agrobacterium tumefaciens strain GV3101 were grown on Luria agar medium at 37 and 28°C, respectively. Pseudomonas syringae pathovar tomato (Pst) DC3000 strains were grown on nutrient yeast glycerol agar (NYGA) [54] at 28°C. E. coli antibiotic selection was 100 µg/mL carbenicillin and/or 50 µg/mL kanamycin. A. tumefaciens antibiotic selection was 100 µg/mL rifampicin, 50 µg/mL kanamycin, and/or 30 µg/mL gentamicin. Pst antibiotic selection was 100 µg/mL rifampicin, and/or 50 µg/mL kanamycin.
Arabidopsis thaliana Col-0 and Pi-0 ecotypes were grown in growth chambers (22°C, 60% RH, 125 µE/m2/s fluorescent illumination) on an 8-h light/16-h dark cycle. Plants were transformed using the floral dip method [55].
Standard DNA cloning methods [56], PCR, and Gateway technology (Invitrogen) were used for plasmid construction. All primer sequences are listed in Table S1. For GST-AvrBsT, avrBsT (wild type, H154A, C222A, and K282R) was amplified by PCR, cloned into pJET1.2, and then sub-cloned into pGEX-5X-3 using BamHI and XhoI. For Gateway constructions, amplified PCR products (i.e. avrBsT, hopZ1a, hopZ1b, hopZ2, hopZ3, ACIP1, ACIP-like genes (ACIP-L1 to ACIP-L6), and TUA5) were cloned into pCR8 to create donor plasmids. The respective donor plasmids were recombined into: 1) pGADT7 to create AD-gene fusions and pGBKT7 or pXDGATcy86 to create BD-gene fusions for two-hybrid analysis; 2) pDEST15 for GST-fusions; and/or 3) pDEST17 for His6-fusions. For avrBsT mutagenesis, QuikChange Site –Directed Mutagenesis kit (Stratagene) was performed with pCR8(avrBsT) and PfuUltra II Fusion HS DNA polymerase (Agilent).
Yeast strain AH109 carrying pXDGATcy86(avrBsT) was transformed with pAD-GAL4-2.1 containing the Horwitz and MA cDNA library isolated from A. thaliana inflorescence meristem, floral meristem, and floral buds (obtained from TAIR). Approximately 7 million transformants were screened and interaction with At3g09980 cDNA was confirmed.
Yeast cells were resuspended in lysis buffer (1.85 M NaOH and 7% 2-mercaptoethanol) and then proteins were precipitated in 10% trichloroacetic acid. Protein pellets were washed in 1 M Tris and then resuspended in 8 M urea sample buffer.
Protein was extracted from plant cells as described [34], separated by SDS-PAGE, transferred to nitrocellulose, and then detected by ECL or ECL plus (GE Healthcare) using anti-ACIP1, anti-HA (Covance), anti-Myc (Covance), anti-His (Qiagen), or anti-GST (Santa Cruz) sera and horseradish peroxidase conjugated secondary antibodies (Bio-Rad). Membranes were stained with Ponceau S to control for loading.
Recombinant His6-ACIP1 was expressed in E. coli BL21 tRNA cells and purified using Ni-NTA agarose under denaturing conditions following manufacturer's protocol (Qiagen). Polyclonal antisera were raised in rabbits using the purified His-ACIP1 protein (Covance).
GST or GST-AvrBsT were expressed in E. coli BL21-CodonPlus(DE3) cells (Stratagene). Cells were lysed in buffer (1X PBS, pH 8, 1% Triton X-100, 0.1% 2-mercaptoethanol, and 1 mM phenylmethylsulfonyl fluoride (Sigma Aldrich)) with a sonicator (Branson). GST and GST-AvrBsT supernatants were incubated with 30 µL of pre-equilibrated Glutathione Sepharose 4B (GE Healthcare) for 1 h at 4°C with rotation. Sepharose beads were recovered by centrifugation and then washed with buffer for 5 min at 4°C with rotation. GST or GST-AvrBsT (WT, C222A, or K282R) bound beads were incubated with soluble E. coli lysates containing His6-ACIP1 for 2 h at 4°C with rotation. The beads were washed with buffer (50 mM TrisHCl, pH 7.5, 150 mM NaCl, 10 mM MgCl2, 0.1% Triton X-100, and 0.1% 2-mercaptoethanol) three times. Protein bound to the beads was separated by SDS-PAGE and analyzed by immunoblot analysis. Anti-GST and anti-His sera were used to detect GST-AvrBsT and His6-ACIP1.
Purified recombinant GST-tagged proteins (1 µg each) were incubated with 100 nM inositol hexakisphosphate (IP6) (Santa Cruz), 0.4 µCi 14C-acetyl-CoA (Perkin Elmer) in 50 mM TrisHCl pH 8 and 1 mM DTT for 30 min at RT. Urea sample buffer was added to stop the reactions. Proteins assayed included: GST, GST-AvrBsT, GST-AvrBsT(C222A), GST-AvrBsT(H154A), GST-AvrBsT(K282R), GST-HopZ1a, and GST-ACIP1. Proteins were separated in a 10% SDS-PAGE gel, stained with Coomassie blue, transferred to blotting paper, dried, treated with EN3HANCE (Perkin Elmer), and then exposed to film for 2–3 weeks at 80°C.
A 365 bp region of ACIP1 was PCR amplified using primer set JG616/JG617, and the product was cloned into pKANNIBAL to create pKANNIBAL(hp-ACIP). The NotI fragment was then subcloned into pART27 [57], creating pAR27(hp-ACIP). The resulting plasmid was then transformed into A. thaliana ecotypes Col-0 and Pi-0. Transformants were analyzed by quantitative RT-PCR to measure ACIP isoform mRNA levels using primer sets listed in Table S1.
Fully expanded leaves of 4- to 5-week-old plants were used for bacterial inoculations. A suspension of bacterial cells (Pst DC3000, Pst DC3000 AvrBsT, Pst DC3000 AvrB, Pst DC3000 AvrRpt2, or Pst DC3000 ΔhrcU; 3×108 cells/mL for HR and 1×105 cells/mL for growth curves) was infiltrated into the extracellular space of fully expanded leaves using a 1-mL syringe. For HR, plants were incubated at RT under lights and phenotypes were recorded 9–12 HPI. For growth curves, plants were incubated at high humidity in a growth chamber for 4 d. Leaf tissue was collected at 0–4 DPI, ground in 1 mM MgCl2, diluted and then plated on NYGA plates containing appropriate antibiotics and cycloheximide (50 µg/mL) in triplicate to determine bacterial load. Four plants were used and the experiment was repeated at least three times. The average bacterial titer ± SD is reported.
Three fully expanded leaves of 4- to 5-week-old plants (n = 4) were inoculated with a 3×108 cells/mL suspension of Pst DC3000 (vector) or Pst DC3000 (AvrBsT). Ten HPI, three leaf discs (10 mm diameter) per plant were floated in 20 mL of water in petri dishes for 5 min and then transferred to a test tube containing 3 mL of water. Tubes were incubated for 1 h at RT with shaking. Conductivity of the solution was measured with an EC meter (Spectrum Technologies) before and after boiling for 30 min [58]. Percent electrolyte leakage was calculated as conductivity before boiling/conductivity after boiling ×100. Assay was repeated at least three times.
Three leaf discs (5 mm diameter) from the youngest fully expanded leaves of a 4-week-old plant (n = 9–18) were incubated in water in a 96-well plate (one leaf disc per well) for 24 h. To measure ROS, leaf discs were treated with ± flg22 (100 nM) in 10 µg/mL horseradish peroxidase and 100 µM Luminol (Sigma), and then luminescence was immediately measured with a 1420 Multilabel Counter (PerkinElmer) [32]. Relative luminescence units (RLU) are reported. Assay was repeated at least three times.
Total RNA was isolated from uninfected or infected leaves using Trizol reagent (Invitrogen) according to manufacturer's instructions. For infection, leaves were inoculated with 1 mM MgCl2 or bacterial strains (2×108 cells/mL in 1 mM MgCl2) and then one leaf from three plants was harvested, pooled, and total RNA was extracted. 2.5 µg of RNA were used for cDNA synthesis. Quantitative RT-PCR was performed using the cDNA and gene-specific primers (Table S1). Each cDNA was amplified by real-time PCR using SensiFAST SYBR Kit (Bioline) and the MJ Opticon 2 instrument (Bio-Rad). UBQ5 or ACTIN8 expression was used to normalize the expression value in each sample and relative expression values were determined against the average value of buffer or bacterially infected sample using the comparative Ct method (2−ΔΔCt).
To monitor ACIP1 protein expression and localization, the promoter region (1.5-kb upstream of start) was fused with GFP-ACIP1 in the backbone of pMDC43 to create pMDC43(PACIP1::GFP-ACIP1). The resulting plasmid was transformed into A. thaliana Pi-0 plants. Transgenic PACIP1::GFP-ACIP1 lines were then transformed with P35S::mCHERRY-HA-TUA5. This plasmid was construct by modifying pEarleygate 104 [59]. YFP was substituted with mCHERRY and Arabidopsis TUA5 genomic coding region was inserted after mCHERRY. Localization of GFP-ACIP1 and mCHERRY-TUA5 in 5-day old dark grown hypocotyls was determined using a Leica TCS SP5 confocal microscope (Leica Microsystems) with Leica LAS AF software and a Leica spinning disc confocal microscope with the Yokogawa CSUX-M1 confocal scanner. Seedlings were treated with 10 µM oryzalin in MeOH for 8 hr at RT or 1 µM latrunculin B in DMSO for 4 hr at RT and then imaged. For infection, Pi-0 PACIP1::GFP-ACIP1 leaves were inoculated with 1 mM MgCl2 or bacterial strains (3×108 cells/mL in 1 mM MgCl2) for 6 h. For exogenous PA treatment, Pi-0 PACIP1::GFP-ACIP1 leaves were inoculated with 50 µM PA in 0.2% DMSO or 0.2% DMSO. Images were analyzed using ImageJ [60].
Sequence data from this article can be found in the Arabidopsis Genome Initiative or GenBank/EMBL databases under the following accession numbers: At3G09980 (ACIP1), At5G03660 (ACIP-L1), At2G36410 (ACIP-L2), At3G52920 (ACIP-L3), At2G27740 (ACIP-L4), At3G52900 (ACIP-L5) and At2G36355 (ACIP-L6).
|
10.1371/journal.ppat.1001035 | The Transcription Factor Rbf1 Is the Master Regulator for b-Mating Type Controlled Pathogenic Development in Ustilago maydis | In the phytopathogenic basidiomycete Ustilago maydis, sexual and pathogenic development are tightly connected and controlled by the heterodimeric bE/bW transcription factor complex encoded by the b-mating type locus. The formation of the active bE/bW heterodimer leads to the formation of filaments, induces a G2 cell cycle arrest, and triggers pathogenicity. Here, we identify a set of 345 bE/bW responsive genes which show altered expression during these developmental changes; several of these genes are associated with cell cycle coordination, morphogenesis and pathogenicity. 90% of the genes that show altered expression upon bE/bW-activation require the zinc finger transcription factor Rbf1, one of the few factors directly regulated by the bE/bW heterodimer. Rbf1 is a novel master regulator in a multilayered network of transcription factors that facilitates the complex regulatory traits of sexual and pathogenic development.
| The basidiomycetous fungus Ustilago maydis is the causal agent of the smut disease on corn. The fungus exhibits two different life-styles, a saprophytic and a pathogenic stage, where it grows yeast-like by budding, or as filamentous, dikaryotic hyphae, respectively. The switch between these two stages is controlled by a heterodimeric transcription factor, bE/bW, which is encoded by the b-mating type locus. We have now identified the genes that are regulated in response to the activation of bE/bW, following the b-mediated developmental change, and address their contribution to the altered morphology and pathogenic development. Interestingly, most of the b-responsive genes are not regulated directly by the bE/bW proteins, but require the action of a second transcription factor, Rbf1, which is induced by bE/bW. Rbf1 defines a novel master regulator as a central component in a multilayered network of different, hierarchically ordered transcription factors that facilitate the complex regulatory traits to coordinate morphology as well as sexual and pathogenic development.
| In a wide range of fungi, complex developmental traits such as cell identity, morphogenesis and sexual development are controlled by mating type loci [1], [2], [3]. In the smut fungi, a group of plant pathogens, these traits also include the ability to infect their host plants. In Ustilago maydis, a smut fungus that infects maize, it is the b-mating type locus that is critical for both sexual as well as for pathogenic development. Similar to other smuts, U. maydis exhibits a dimorphic life cycle. The haploid, cigar-shaped cells, called sporidia, multiply by yeast-like budding, and the dikaryon, which is formed upon the fusion of two compatible sporidia, grows as a filament. This switch in cell morphology is accompanied by an alteration of the life-style. While the sporidia are apathogenic and grow strictly saprophytic, the filament is biotrophic, i.e. it depends on the living tissue of its host plant maize for further development. Initially, the dikaryotic hypha consists of a long tip cell with the accumulated cytoplasm; the succeeding, older parts consist of “empty” cells that are separated by regularly spaced septae. Cell division is stalled until the hypha has penetrated the cuticula of a corn plant, and only then a “true” filament with multiple septated compartments is formed. Upon plant invasion, hyphae traverse the plant without harming the cells and without an apparent host defense response. After several days, the fungus induces plant tumors, coinciding with a massive proliferation of fungal hyphae [for review, see 4].
In order to fuse and to form the pathogenic filament, the two sporidia must carry different alleles both of the biallelic a- and of the multiallelic b-mating type locus. The a-locus encodes a pheromone/receptor system required for cell sensing, initiation of filamentous conjugation tubes, and cell fusion. After fusion, the crucial step for the initiation of the pathogenic phase is the formation of a heterodimeric complex of two homeodomain proteins, bE and bW, which are encoded by the b-mating type. This bE/bW complex is formed only when the two proteins are derived from different b-alleles, and is sufficient to initiate the switch from budding to filamentous growth. Concomitantly, activation of b leads to a cell cycle arrest that is only released after host plant infection. It has been shown conclusively that the bE/bW complex is sufficient to initiate the pathogenic development, as exemplified by haploid “solopathogenic” strains that harbor different alleles of bE and bW and that are capable to infect plants without a mating partner [5]. Thus, it is conceivable that genes regulated by the bE/bW heterodimer are involved in (1) the establishment of the biotrophic phase, (2) cell cycle regulation and (3) the dimorphic transition from budding to the polarized growth of the filament. However, until now, only four b-regulated genes have been identified with impact on these processes, three of which are required during the very early infection stages. biz1 encodes a zinc finger transcription factor that is involved in the G2 cell cycle arrest preceding plant penetration as well as in the induction of appressoria, specific infection structures at the tip of penetrating hyphae [6]. The mitogen-activated protein (MAP) kinase Kpp6 is required for the subsequent step: U. maydis strains harboring a non-activatable kpp6 allele still form appressoria, but are defective in the penetration of the plant cuticula [7]. After plant penetration, the clp1 gene is required for further proliferation of dikaryotic filaments in planta. clp1 mutant strains still penetrate the plant cuticula, but development is stalled prior the first mitotic division; in addition, mutant strains do not form clamps, a structure that ensures the proper distribution of nuclei in the dikaryotic hyphae [8]. Interestingly, the induced expression of clp1 strongly interferes with the b-dependent induction of several of the genes regulated by the bE/bW-heterodimer, indicating that Clp1 may modulate the activity of the bE/bW complex. And finally, the b-dependently expressed cyclin Pcl12 is involved in the polarized growth of the b-dependent filament, but is dispensable for pathogenic development [9].
The bE/bW heterodimer binds to a conserved sequence motif, the b-binding sequence (bbs) that has been identified in the b-dependently induced lga2-gene [10]. Out of the 20 b-dependent genes identified so far, only two additional genes were found to harbor the bbs-motif: the above mentioned clp1 gene, and frb52, a gene with unknown function [11]. As the majority of b-controlled genes is obviously not directly regulated by bE/bW, it appears likely that the bE/bW heterodimer triggers a regulatory cascade with a limited number of direct targets genes. Thus, these “class I” genes should encompass regulators that trigger the regulation of the larger number of indirect, “class II” b targets. It was proposed that these regulators play pivotal roles either in all (as master regulator) or distinct (as relay) b-dependent processes.
Here, we employed U. maydis strains that harbor inducible combinations of the bE and bW genes [11] and DNA array technology to investigate the b-dependent processes in a time-resolved manner. Our analysis provides insight in the complex interconnection of cell cycle regulation during the dimorphic switch and highlights the specific characteristics of the “pathogenic” hyphae. Most important, we identify the zinc-finger transcription factor Rbf1 as a novel master regulator that is required for all b-dependent processes.
In order to identify genes regulated by the bE/bW heterodimer, we performed microarray experiments with custom Affymetrix arrays (MPIUstilagoA) covering 5823 of the predicted 6786 U. maydis genes. Changes in gene expression were monitored during a 12-h time course (with samples taken at 1h, 2h, 3h, 5h, 12h) using the haploid U. maydis strains AB31 and AB33 that harbor the bE1 and bW2 genes under the control of the arabinose-inducible crg1 promoter and the nitrate-inducible nar1 promoter, respectively [11]. Induction of bE1/bW2 in these strains results in a filament that resembles the infectious hypha formed after fusion of compatible sporidia [11]. Strains AB32 and AB34, which harbor the incompatible bE2 and bW2 combination, were used as controls. Expression of bE and bW genes was induced by a shift from glucose- to arabinose (AB31 and AB32) or from glutamine- to nitrate containing media (AB33 and AB34). The expression profiles after b-induction in AB31 and AB33 were similar, but not identical. Firstly, the use of different media had an effect on gene expression, and, secondly, the use of the crg1 promoter resulted in gene expression values that were two- to fivefold higher when compared with nar1-driven gene expression (Suppl. Fig. S1). To account for expression changes caused by the medium shift, we considered changes only as relevant when the expression for a particular gene was altered significantly in both AB31 and AB33 in at least one time point (change in expression ≥2, adjusted p-value ≤0.01). Using these criteria, 206 genes were induced and 139 were repressed in response to b-induction (Fig. 1; Suppl. Table S1). Within this list, all genes with a significant b-dependent regulation identified in previous studies were present, emphasizing the validity of the global approach and the quality of our data set (Suppl. Table S2).
From the 345 b-regulated genes, a total of 239 were functionally classified using the Blast2Go tool [12]. Using enrichment analysis, we did not observe a significant over-representation of b-induced genes in any of the Gene Ontology (GO) categories (http://www.geneontology.org). However, for the b-down-regulated genes, we observed a significant enrichment of the GO categories “Cell Cycle” (GO:0007049; 29 genes), “Chromosome” (GO:0005694; 25 genes) and “DNA metabolic process” (GO:0006259; 19 genes), “Cytoskeleton” (GO0005856; 16 genes) and “Microtuble cytoskeleton” (GO:0015630; 9 genes) (Suppl. Table S3).
The induction of the active bE1/bW2-heterodimer leads to a G2 cell cycle arrest, and in accordance with this observation we found cln1, clb1 and clb2, which encode a G1-type cyclin and two B-type cyclins [13], [14] among the down-regulated genes (−29.9-fold, −7.7-fold and −2.6-fold, respectively; Suppl. Table S1). cln1 and clb1 are involved in G1 to S transition, clb1 and clb2 in the G2 to M transition; thus, it is expected that these genes are poorly expressed in cells that are arrested in G2. For clb1 it has been shown previously that the repression leads to a G2 cell cycle arrest [14]; thus, the observed low expression of this cyclin may trigger the b-induced G2 arrest. Additionally, we find um03928, encoding a homologue of the S. pombe Cdr2 protein, as 40.7 fold down-regulated. In S. pombe, Cdr2 functions as a mitotic inducer via the Wee1 kinase and is required for G2/M transition [15]. In U. maydis, Wee1 has been shown to be a central regulator for G2/M transition [16]; however, a function for the Cdr2 homologue um03928 has not been assigned yet. Another level of complexity may be achieved via the up-regulation of the Cdk5 associated Pho80 Cyclin Like protein Pcl12 (49.1-fold, Suppl. Table S1). Induced expression of pcl12 leads to a G2 cell cycle arrest, and promotes filamentous growth [C. Pothiratana and J. Kämper, unpublished; 9]. Thus, the b-induced cell cycle arrest may be realized via the synchronized regulation of independent pathways. In line with the cell cycle arrest, we observe the repression of genes involved in DNA-replication and nucleotide metabolism, as, for example, um01008, encoding the catalytic subunit of DNA polymerase epsilon (3.6-fold down-regulated at 12h), or um06402, encoding a DNA replication licensing factor (3,2-fold down-regulated at 12 h; Suppl. Table S1, FunCat DNA).
Several of the b-regulated genes can be attributed to the morphological switch from budding- to filamentous growth. A total of 20 genes with a potential function in cell wall synthesis or modification was found to be induced, starting 3 h after b-induction, which coincides with the onset of filamentation; five additional genes were repressed (Suppl. Table S1, FunCat: CW). These genes encode for chitin synthases as well as for exochitinases, chitin deacetylases, and exo- and endoglucanases, indicating that the cell wall composition is altered during the switch from sporidia to hyphae.
The largest “functional” group (74 genes) encodes for potentially secreted proteins. 34 of them have no ascribed function, and of these 15 are specific for U. maydis. Such secreted proteins are candidates for effectors that may play a role in the establishment of the biotrophic interaction (Suppl. Table S1, secreted).
To identify b-dependent genes important for pathogenic development, we focused initially on genes whose expression was “strictly” dependent on the presence of the bE/bW heterodimer, i.e. genes that showed only basal expression levels in strains AB32 and AB34 and showed a more than 10-fold induction upon expression of an active bE1/bW2-heterodimer. None of the 53 genes that fulfilled these criteria showed a significant similarity to known pathogenicity factors. Potential exceptions were dik6 and dkh6, which encode two related seven trans-membrane (7TM) domain proteins. 7TM proteins represent an extended protein family in M. grisea that is discussed to function in plant/pathogen interactions [17]. However, neither the single, nor the double deletion of the two genes had an impact on pathogenic development (Suppl. Table S1, G. Weinzierl and J. Kämper, unpublished). In total, we deleted 30 of the 53 strictly b-dependent genes in the haploid, solopathogenic strain SG200; in addition, nine genes have been analyzed in the course of previous studies. 35 of the 39 deletion strains did not show altered virulence when assayed in plant infection assays.
However, the individual deletion of each of the five genes encoding proteins with potential regulatory functions affected pathogenic development or filamentous growth (Suppl. Table S1). Among these genes was clp1 (um02438), which has been identified in the course of this study and has been shown to be required for pathogenic development and in planta proliferation [8]. The biz1 gene (um02549) encodes a C2H2 zinc finger transcription factor that is required for pathogenic development and efficient appressoria formation [6]. In addition, we could show that the deletion of two genes encoding potential homeodomain transcription factors (um12024 and um04928, termed hdp1 and hdp2) impaired filamentous growth or led to loss of pathogenicity, respectively; the detailed characterization of these two genes will be published elsewhere. Here we will focus on the analysis of um03172, encoding a potential C2H2 zinc finger transcription factor.
Due to the initially observed phenotype (see below), the U. maydis gene um03172 was termed rbf1 (regulator of b-filament). According to our microarray analysis, rbf1 expression was strongly induced early after b-induction (Fig. 2A). Significant expression was detected already 1h after b-induction in AB33 (13.6-fold induction), and expression peaked at 2h to 3 h (176.3-fold in AB33 at 2h and 297.4-fold in AB31 at 3 h, respectively; Fig. 2A). In the control strains AB32 and AB34 rbf1 expression was not detectable. The b-dependent expression of rbf1 was confirmed by qRT-PCR using strains AB31 and AB32 (Fig. 2B). Within the rbf1 promoter, we identified three motifs with similarities to the previously identified b-binding sequences (bbs) (Fig. 2C). We used an AB31 derivative expressing the bE1 protein fused to a triple HA-tag (AB31bE1:3xHA) for quantitative chromatin immunoprecipitation analysis (qChIP). Induction of bE1:3xHA/bW2 genes in this strain led to filamentous growth (see below), demonstrating that the bE1:3HA protein is functional (data not shown). In a qChIP analysis with bE1:3xHA, a significant enrichment (P = 5.7 10−5, Students t-test) was observed for the PCR amplicon covering the bbs-motif located at position −1377, when compared to a amplicon covering a region further upstream in the rbf1 promoter (Fig. 2D). The bbs−1377-motif shares also the highest sequence similarity with the previously described bbs-motifs (Fig. 2C). When the rbf1 gene with a promoter fragment deleted for bbs−1377 was used for transformation of a strain deleted for rbf1, the rbf1 deletion phenotype could not be complemented (Fig. 3D, Table 1, see below), demonstrating that bbs−1377 is required for expression of rbf1. The early induction of rbf1 upon b-activation, the presence of a conserved bbs-motif which is bound by the bE/bW heterodimer in vivo, and the requirement of this bbs-motif for rbf1-function strongly suggest that rbf1 is a direct target of the bE/bW heterodimer.
The cDNA-copy of rbf1 was obtained by RACE and revealed four introns when compared to the genomic locus. The predicted open reading frame encodes a protein of 404 amino acids (aa) with an N-terminal C2H2 zinc finger domain (aa 18 to 131), a putative nuclear localization sequence (RHRR, aa 95 to 98) within the zinc finger domain and a C-terminal glutamine-rich sequence (aa 365 to 373) (Fig. 2E). To determine the localization of Rbf1, the open reading frame was fused to a triple eGFP gene and integrated into strain AB31 via homologous recombination, thereby replacing the native rbf1 gene. Fluorescence microscopy of the resulting strain AB31rbf1:3eGFP (UMS63) revealed a nuclear localization of the functional Rbf1-3xGFP fusion protein upon induction of the bE/bW heterodimer (Fig. 2F), fostering the assumption that rbf1 encodes a C2H2 zinc finger transcription factor.
To investigate the biological function of rbf1, the gene was deleted in the haploid solopathogenic strain SG200 (a1mfa2bE1bW2) and in the haploid U. maydis wild-type strains FB1 (a1b1) and FB2 (a2b2), producing strains SG200Δrbf1 (UMS20), FB1Δrbf1 (UMS49) and FB2Δrbf1 (UMS51), respectively. In all strains, the deletion of rbf1 did not cause any obvious phenotype in haploid sporidia growing in axenic culture, and the growth rate was not altered in different minimal or complete media (data not shown). However, when the compatible strains FB1Δrbf1 and FB2Δrbf1 were crossed on charcoal containing plates, only very short filaments were observed at the edge of the forming colonies, while the cross of FB1 or FB2 resulted in fuzzy white colonies indicative for the formation of the filamentous dikaryon (Fig. 3A). Similarly, only scarce filament formation was observed in SG200Δrbf1 (Fig. 3B). Since SG200 cells undergo the dimorphic switch without the need of a mating partner on charcoal containing media, we can exclude that the drastically reduced filamentation is caused by a defect in cell-cell fusion. Treatment of FB1Δrbf1 cells with synthetic a2 pheromone resulted in the formation of conjugation tubes which were indistinguishable from those produced by wild-type FB1 cells, indicating that deletion of rbf1 does not affect polarized growth per se (Fig. 3C). Transformation of SG200Δrbf1 with plasmid pRbf1 harboring the rbf1 gene and 3kb of 5′sequence restored the fuzzy colony appearance; three independent transformants (SG200Δrbf1 ip::rbf1) were indistinguishable from the SG200 wild type strain (Fig. 3 D). However, the rbf1 deletion phenotype was not complemented when plasmid pRbf1Δbbs−1377, in which the bbs-motif at position −1377 in the rbf1 promoter was deleted, was used for transformation (SG200Δrbf1 ip::rbf1Δbbs−1377) (Fig. 3 D).
To assess the role of rbf1 during pathogenic development, seven days old maize plants were inoculated with SG200Δrbf1, or with a mixture of FB1Δrbf1 and FB2Δrbf1, and scored for tumor formation. Seven days post inoculation (dpi) 95% and 94% of the plants inoculated with SG200 and a mixture of FB1 and FB2, respectively, had developed tumors. In contrast, inoculation with the respective Δrbf1 mutants resulted in the complete absence of infection symptoms (Table 1). As expected, transformation of SG200Δrbf1 with pRbf1 (SG200Δrbf1 ip::rbf1) restored pathogenicity, while transformation with pRbf1Δbbs−1377 (SG200Δrbf1 ip::rbf1Δbbs−1377) did not (Table 1). To determine at which stage of pathogenic development the rbf1 mutant strains were blocked, fungal hyphae were stained with calcofluor at 2 dpi. Microscopic observation revealed that the Δrbf1-strains formed filaments on the leaf surface (Fig. 3E), however, we did not observe any hyphae within plant cells. To assess whether SG200Δrbf1 was able to form appressoria, we co-inoculated plants with a mixture of SG200 and SG200Δrbf1 strains, each expressing either cytoplasmatically localized CFP or YFP to distinguish the strains. In the combinations SG200-CFP/SG200Δrbf1-YFP and SG200-YFP/SG200Δrbf1-CFP, we counted 57 and 60 appressoria for the SG200 strains on the leaf surface, respectively. By contrast, we were unable to detect any appressoria formation for the SG200Δrbf1 strains in the same surface areas. Thus, the observed pathogenicity defect of Δrbf1 strains results from the inability to form appressoria and to penetrate the plant cuticle.
To get a more detailed view on the role of rbf1 during b-dependent filament formation, we deleted the gene in strain AB31. More than 90% of the cells had switched to filamentous growth 12h after b-gene induction in AB31, while in AB31Δrbf1 (UMS25) no filament formation was observed (Fig. 4A). Upon induction of bE1/bW2 in AB31 the cells stop to divide; in contrast, in AB31Δrbf1 cells continued to grow by budding (Fig. 4A and 4C), indicating that rbf1 is required for both filamentous growth as well as for the b-dependent cell cycle arrest. FACS analysis of AB31 cells revealed an accumulation of cells containing a 2C DNA content upon b-induction, indicative for the b-induced G2-cell cycle arrest. In AB31Δrbf1, however, the distribution of cells with 1C and 2C DNA content was comparable to the wild-type strain FB2, corroborating the requirement of rbf1 for the b-induced cell cycle arrest (Fig. 4B).
To dissect b-dependent and rbf1-dependent processes, we constructed strain CP27 (a2Δb::Pcrg1:rbf1), an FB2 derivative in which the b-locus was replaced by a copy of rbf1 under control of the arabinose-inducible crg1 promoter. Induction of rbf1 in CP27 resulted in the formation of filamentous cells that were indistinguishable from b-induced filaments: the cells contained single nuclei (Fig. 4A) and stopped to divide (Fig. 4C). FACS analysis revealed that rbf1 induction in CP27 leads to a G2 cell cycle arrest (Fig. 4B) analogous to that observed after b-induction. In summary, our results demonstrate that rbf1 is required for b-dependent filament formation and G2 cell cycle arrest and, in addition, sufficient to induce these developmental steps in the absence of an active bE/bW-heterodimer.
To analyze the connection between b- and rbf1-mediated gene-regulation in more detail, we performed DNA-array analysis. b-dependent genes for which rbf1 is required for expression were identified by comparing the transcriptional profile of strains AB31Δrbf1 and AB31 at 3h, 5h and 12h after b-induction. Induced expression (5h) of rbf1 in strain CP27 (a2Δb::Pcrg1:rbf1) was used to identify genes for which rbf1 is sufficient for regulation. 189 (91.7%) out of the 206 previously identified b-induced genes showed no significant changes in expression after b-induction in strain AB31Δrbf1 (Fig. 5; Suppl. Table S4). From the remaining 17 genes, 11 showed comparable expression levels after b-induction in AB31 and AB31Δrbf1, and 6 genes showed significant, but reduced expression levels in AB31Δrbf1. The 11 genes that showed no altered b-dependent expression upon rbf1 deletion did not respond to rbf1 induction in strain CP27; we consider these genes to be regulated only by b, and not by Rbf1 (“only b”, Fig. 5, Suppl. Table S4). With the exception of um00027, all these genes harbor sequence motifs with similarities to the b-binding site within their promoter sequences. In addition, they are all up-regulated early upon b-induction, indicating that these genes are most likely direct targets of the bE/bW heterodimer. The six genes that show a significant, but reduced b-responsive expression in AB31Δrbf1 all respond to rbf1 induction in CP27. Four of the genes harbor b-binding sites in their promoter regions; apparently, these genes may be regulated directly via b and, in addition, independently via rbf1 (“rbf1 OR b sufficient”, Fig. 5, Suppl. Table S4).
For a large fraction (46%) of the b-dependent genes rbf1 is both sufficient as well as required for expression; for these genes, deletion of rbf1 abolishes the b-dependent induction, and they respond to rbf1-induction in CP27. It is likely that the regulation of these genes occurs by a b-mediated regulatory cascade via Rbf1 as a central regulator (“rbf1 required AND sufficient”, Fig. 5, Suppl. Table S4). Expression of the remaining 102 genes was dependent on rbf1, however, no significant induction was detected 5h after rbf1 induction in CP27. Notably, 63 of these genes were late b-induced (12h after b-induction in AB31), and additional 22 genes were only weakly induced (less than 3-fold), or only transiently induced 5h after b-induction in AB31. It is well possible that these genes respond to rbf1 only after prolonged rbf1 induction (>5h). For 16 genes, we observed a significant b-dependent induction, no induction in AB31Δrbf1, and no rbf1-dependent induction in CP27. Thus, we have to assume that for the regulation of these genes the action of both b and rbf1 is required (“rbf1 AND b required”, Fig. 5, Suppl. Table S4).
An analogous scenario was found for the b-dependently repressed genes: of the 139 b-dependently repressed genes, the repression was abrogated for 129 (92.8%) genes in AB31Δrbf1. For a total of 69 genes rbf1 was both required and sufficient for repression. Formally, the b-repressed genes can be grouped equivalent to the b-induced genes (b only; rbf1 AND b; rbf1 OR b; only rbf1; Fig. 5, Suppl. Table S4).
To assess whether the rbf1-dependent gene expression involves the binding of Rbf1, we dissected the promoter of dik6, one of the rbf1 responsive genes, by means of qChIP analysis. We used an AB31 derivative where the rbf1 gene was replaced by a rbf1-3xHA fusion (AB31rbf1:3xHA). Induction of bE1/bW2 in this strain triggers the expression of the Rbf1-3xHA fusion protein, which results in filamentous growth, demonstrating that the Rbf1-3xHA fusion protein is functional (Data not shown). qChIP analysis was performed via a set of 9 overlapping amplicons spanning 930 bp of the dik6 promoter (Fig. 6A); as controls, we used an amplicon upstream of the potential promoter (−1703 to −1829 with respect to the ATG) and an amplicon within the dik6 ORF (Fig. 6A). With the exception of an amplicon spanning the region from −9 to −157, all amplicons within the promoter showed significant (P<0.001) differences in enrichment when compared to the control amplicon located within the ORF. The amplicons with the highest enrichment were found to span the region from −825 to −422 (Fig. 6B) The functional analysis of the dik6 promoter by means of promoter-GFP fusions revealed that Rbf1-induced GFP expression levels declined when the dik6 promoter was truncated from 816 to 638 bp, while a 298 bp fragment was not sufficient to mediate expression (Fig. 6A,C). Internal promoter deletions corresponding to the amplicons used for the qChip analysis revealed that deletion of the dik6 promoter region from −825 to −680 (corresponding to amplicon 3) led to reduced Rbf1-dependent induction, while the deletion of the promoter region from −601 to −500 (corresponding to amplicon 5) abolished expression completely (Fig. 6A,C). In summary, our data indicate that the dik6 promoter has at least one binding site for Rbf1 that is required for Rbf1-mediated dik6-expression.
Previously, it was shown that rbf1 is induced when haploid cells are treated with compatible pheromone [18]. Since both bE as well as bW are also induced upon pheromone treatment [19], we asked whether Rbf1 might be required for pheromone-dependent expression of the b genes. However, real time qRT-PCR analysis revealed no difference in the abundance of bE and bW transcripts in the strains FB1 and FB1Δrbf1 (UMS49; a1b1Δrbf1) upon treatment (75 min) with synthetic a2 pheromone (see Suppl. Fig. S2). Thus, we can exclude the possibility that Rbf1 is required for the pheromone-dependent b-induction.
In summary, our data identify Rbf1 as the central regulatory switch within the b-dependent regulatory cascade, which is not only required for the regulation of the majority of the b-dependent genes, but also indispensable for all b-mediated developmental processes.
The switch from saprophytic to biotrophic growth of U. maydis requires a meticulous coordination of different processes, such as cell cycle control, the change to polarized growth, and, most interestingly, the onset of a program facilitating plant invasion and colonization. The top-most control instance for these processes is the b-mating type locus; it has been conclusively shown that compatible b-alleles are both required and sufficient for pathogenic development [5]. The bE/bW-heterodimer also controls polarized cell growth and induces a G2 cell cycle arrest, but not exclusively, since both can be triggered as well via the pheromone/receptor system encoded by the a-mating type locus [20]. Necessarily, the a and b loci are cross-controlled: activation of the a-pathway leads to induction of the bE and bW genes via Prf1, and the formation of an active bE/bW-heterodimer after cell fusion leads to a down-regulation of the a-pathway [19], [21].
Since a direct binding of the bE/bW heterodimer to promoters of the plethora of genes associated with b-dependent processes appeared unlikely, we have favored a model that places b on top of regulatory proteins (relays) mediating the regulation of further downstream targets. We have now identified the C2H2 zinc finger transcription factor Rbf1 as a central key player within this regulatory network.
The fast induction of rbf1 upon b-activation, the binding of the bE/bW-heterodimer to a defined b-binding site in the rbf1-promoter region as well as the requirement of this binding site for rbf1 function define rbf1 as a direct target of the bE/bW-heterodimer. Deletion of rbf1 abolishes all b-mediated processes, and induction of rbf1 leads to filamentation and a G2 cell cycle arrest analogous to that observed upon b-induction. In addition, we could show that rbf1 is required for regulation of the far majority of b-regulated genes, and, for a large fraction, also sufficient. Thus, we consider Rbf1 as a key master regulator whose action is sufficient to induce an entire complex developmental pathway. Despite of the essential function of Rbf1 within the b-regulatory cascade, we consider it unlikely that rbf1 alone is sufficient to trigger pathogenic development of U. maydis, because clp1, which was shown to be required for pathogenicity [8], is induced directly by b and independently from rbf1. We were not able to address this question experimentally, since transformants with a constitutively expressed rbf1 gene were not viable, most probably as a result of the rbf1 induced cell cycle arrest.
In fungi, only few master regulators of pathogenic development have been identified yet. In Candida albicans, WOR1 is the master regulator of white to opaque switching [22], and the C. neoformans Gat201 [23] is a key regulator of melanin production and capsule formation. The C. neoformans Cir1 transcriptional regulator integrates the sensing of iron with the expression of virulence factors, with signalling pathways influencing virulence, and with growth at elevated temperature [24], [25]. WOR1 and Gat201 are required (and sufficient) for the initiation of specific programs that are tightly linked to fungal pathogenesis. In contrast, nearly all of the genes regulated by b require rbf1 for their expression, and it is not possible to assign specific, common functions to the rbf1-regulated genes, or to the few genes that are not regulated by Rbf1. Thus, different from WOR1 and Gat201, Rbf1 regulates not the genes of a specific, defined pathway, but is required for the regulation of all b-dependent processes. Similarly, the C. neoformans Cir1 regulator is involved in the regulation of all major virulence traits [24], [25].
The cell cycle block of the b-induced filaments is only released upon plant penetration. Our data reveal a complex contribution of different key players to control the cell cycle. At least four different transcription factors, namely bE/bW, Rbf1, and the two Rbf1-dependent factors Biz1 and Hdp1 are involved in cell cycle regulation. The ectopic expression of any of these factors leads to the formation of G2 arrested hyphae [6], [8, C. Pothiratana and J. Kämper, unpublished], which argues for a complex transcriptional network with different levels of relays that allow the integration of various stimuli, as for example, the unknown signal that leads to the release of the cell cycle after penetration of the host plant. The regulatory control achieved via b, Rbf1, Biz1 and Hdp1 may funnel into the transcriptional regulation of different key factors for cell cycle control, as we observe the transcriptional regulation of different cyclins (cln1, clb1, clb2 and pcl12) and of the potential Wee1 kinase Um03928. The Um03928 homologue in S. pombe, Cdr2, is required for the proper formation of septae, and functions as mitotic inducer via the negative regulation of the central cell cycle regulator Wee1 [15], [26]; the U. maydis Wee1 was shown to trigger filamentous growth and a G2 arrest [9], [16]. Obviously, the b-induced G2 cell cycle arrest is controlled by several independent regulatory pathways.
The induction of b leads to the formation of polar growing hyphae, and several of the b-dependently regulated genes reflect this morphological change and the altered requirements of the cell for e.g. long distance transport or cell wall remodeling. However, the most interesting trait by which the b-induced filament differs from other filaments like the pheromone-induced conjugation tube is its ability to infect the host. The exploitation of the b-dependently regulated genes provides for the first time comprehensive insights into the complex developmental processes during morphogenic switching and pathogenic development of U. maydis. The pathogenic potential of the hyphae may for once be marked by an altered cell wall composition, as we observe the differential regulation of several genes involved in cell wall synthesis, including chitin synthases and chitin deacetylases. Rebuilding or masking of the cell wall is a strategy of pathogens to evade perception or to protect themselves from defense responses of the host [27]. However, deletion of either of the two b-regulated chitin deacetyases does not affect virulence in U. maydis (B. Günther, J. Kämper, B. Moerschbacher, unpublished), and neither are the two chitin synthases chs1 and chs4 required for pathogenicity [28], most likely due to overlapping and/or redundant functions.
The other intriguing characteristic of the b-filament is the secretion of various potential effector proteins. Such effectors are thought to be involved in suppression of host defense responses and redirection of nutrient flow during biotrophic growth. The expression of putative effectors prior to the contact with the plant indicates a priming mechanism of the fungal hypha to facilitate rapid suppression of plant defense responses and the fast establishment of the biotrophic interface subsequent to plant penetration. The observation that the temperature-induced inactivation of the bE1/bW2-heterodimer in planta abolishes expression of various additional candidate effector genes [29] that are not identified as b-regulated in this study, implies that the temporal expression of these genes is subject to combinatorial gene regulation involving the bE/bW-heterodimer and other plant-induced factors.
The competence of the b-filaments to penetrate the plant cuticula is reflected by the induction of Biz1 and the MAP kinase Kpp6. Both factors have been shown to be required for efficient formation of appressoria and subsequent penetration. Rbf1 is required for the b-dependent induction of both genes, which explains the absence of appressoria in rbf1 mutant strains.
Rbf1 is required for the induction of most, but not all b-regulated genes. All genes that are exclusively regulated by bE/bW harbor b-binding sites within their promoters, and it is conceivable that these genes are regulated via direct binding of the bE/bW-heterodimer. The majority of the b-regulated genes, however, lack putative b-binding sites, indicating that the b-dependent regulatory circuit involves additional transcription factors.
Similarly, those genes for which rbf1 is required and sufficient for regulation may be directly regulated by Rbf1. Our data indicate that Rbf1 binding to the promoter of the dik6 gene is required for Rbf1-mediated dik6 expression, which emphasizes the function of Rbf1 as a transcription factor. However, the actual Rbf1 binding site has not been determined yet. The in silico analysis of rbf1-regulated genes may be constrained by the fact that Rbf1 triggers the induction of at least three transcription factors, leading to a superimposition of direct and indirect effects. For a small fraction of genes, both b and rbf1 are required for regulation, which can be explained by a combinatorial action of two transcription factors [30]. A substantial number of genes is down-regulated upon b-induction. Intriguingly, two of the very few known transcription factors that can act both as transcriptional activators and repressors, the S. cerevisiae Rme1 protein and the human YY1 protein, are both C2H2 zinc-finger proteins [31], [32]. Thus, it is well possible that the repression of genes is also directly mediated via Rbf1.
The dimorphic switch and the onset of pathogenic development trigger a multilayered regulatory cascade that involves several transcription factors (Fig. 7). Is there a specific reason that the bE/bW-heterodimer regulates only a small number of genes directly and more than 90% in dependence on a second master regulator? For once, additional regulators allow more signals to be integrated into the regulatory circuits, which may help to quickly adapt to changing environmental conditions during biotrophic development, thereby avoiding nutrient stress or plant defense responses. In particular, Rbf1 interconnects the a- and b-dependent regulatory pathways, as both pheromone-response [18] as well as b-induction leads to rbf1 expression. Since we could not determine a specific function for Rbf1 in the pheromone-dependent signalling pathway and deletion of rbf1 is not required for conjugation tube formation and for pheromone-induced G2 cell cycle arrest, we consider it unlikely that rbf1 plays a central role in a-dependent signalling or gene regulation. One possibility is that the pheromone-induced expression of b and rbf1 primes the cells for post-fusion events. The a- and b-dependently induced cell cycle arrest is independently coordinated; thus, the pheromone-induced rbf1 expression facilitates rapid switching of developmental programs thereby minimizing the time preceding plant infection (Fig. 7).
In essence, our study provides fundamental new insights into the complex regulatory traits of sexual, as well as pathogenic development of U. maydis. The identification of key factors points towards an emerging picture that explains how multilayered regulatory pathways can dynamically interact to control complex developmental decisions. We believe that this work is not only relevant for U. maydis but can also serve as a model for other fungi and higher organisms.
Escherichia coli strain TOP10 (Invitrogen) was used for cloning purposes. Growth conditions and media for E. coli [33] and U. maydis [8], [34], [35] and the quantification of appressoria formation [6] have been described previously. U. maydis strains with relevance for this study are listed in Suppl. Table S5. U. maydis strains carrying crg1 expression constructs were induced in array medium [8] or CM medium [34] supplemented with 1% arabinose instead of 1% glucose as described in [8]; equivalently, for nar1-induction, array medium supplemented with 3.8g/l KNO3 instead of glutamine (4.38 g/l) as nitrogen source was used. Mating assays and plant infections are described in reference [35]. For pheromone stimulation of U. maydis cells we followed the protocol of [36].
Molecular methods followed described protocols [33]. DNA isolation and transformation procedures for U. maydis were carried out as described [37]. For all gene deletions, we used the PCR based approach described in [38]. For the Rbf1-3xeGFP fusion, 1 kb of the 3′end of the rbf1 ORF and 1 kb of the 3′ UTR were PCR-amplified, introducing two SfiI sites and removing the stop-codon of rbf1; both fragments were ligated to an SfiI 3xeGFP-HygR fragment of pUMA647 (K. Zarnack and M. Feldbrügge, unpublished) in pCR2.1 TOPO (Invitrogen) as backbone, yielding pMS85.
To replace the b-mating type locus with the arabinose inducible rbf1 allele, the rbf1-ORF was PCR amplified, creating an NdeI site at the start and a NotI site following the stop codon, and cloned into pCR2.1 TOPO. The NdeI-NotI rbf1-ORF-fragment, a 1.3 kb BstXI(blunt)-NdeI crg1-promotor fragment and a 0.3 kb NotI-EcoRI(blunt) nos-terminator fragment from pRU12 [11] were integrated into the StuI site of pCRΔb [38] to generate pCRΔb-crg:rbf1. For the generation of HA-tagged bE1- and Rbf1-fusion proteins 1 kb of the 3′end of the ORF and 1 kb of the 3′ UTR were PCR-amplified, introducing two SfiI sites and removing the stop-codon; respective fragments were ligated to an SfiI 3xHA-HygR fragment of pUMA792 (M. Feldbrügge, unpublished) and cloned into pCRII TOPO, yielding plasmids pDS1 and pDS3. After linearization plasmids were integrated into the bE1 and rbf1 loci, respectively, of strain AB31 by homologous recombination.
All PCR amplified fragments were verified by sequencing. For transformation, either linearized plasmid DNA or PCR generated linear DNA was used; homologous integration was verified by Southern blot. For complementation of the rbf1 deletion, a 3kb region upstream of the rbf1 ORF was PCR amplified introducing a 5′-FseI and a 3′-NdeI restriction site and inserted with the NdeI-NotI rbf1-ORF fragment of pCRΔb-crg:rbf1 into pRU11-NotI6474 (a pRU11 [11] derivative in which the NotI site at position 6474 has been removed by a fill up reaction) by three-fragment ligation to generate pRbf1. Generation of pRbf1Δbbs−1377 was performed as described for pRbf1, with the exception that the bbs-motif at position −1377 within the 3kb rbf1- upstream region was removed by standard PCR techniques [39].
For generation of dik6 promoter-GFP fusion constructs the 2448 bp dik6 promoter fragment was PCR-amplified and integrated into pRU4 [11] digested with HpaI and NdeI. From the resulting plasmid dik6 promoter fragments of 816 bp, 638 bp and 298 bp were recovered as BclI(blunt)/NdeI, MscI(blunt)/NdeI and HindIII(blunt)/NdeI fragments and integrated into pRU4 [11] digested with HpaI and NdeI. Internal deletions in the dik6 promoter were introduced by standard PCR techniques [39]. PCR amplified fragments were integrated into pRU11 via FseI/NdeI restriction sites [11].
RNA extraction and qRT-PCR analysis for rbf1, bW, bE and ppi was performed as described [8]. Full-length rbf1 cDNA was isolated using the GeneRacer Kit (Invitrogen), cloned in pCR2.1 TOPO (Invitrogen) and sequenced. For overview of primers used see Suppl. Table S6.
50 ml cultures of U. maydis were grown until OD600 = 0,6–1,0 and cross-linked by addition of formaldehyde (1% final concentration) for 15 min at RT; glycine was added to a final concentration of 0.125 M, cells were harvested by centrifugation and washed three times in TBS (50 mM Tris-HCl, 150 mM NaCl, pH 7.6). The pellet was resupended in 1.5 ml FA lysis buffer (50 mM HEPES-KOH [pH 7.5], 150 mM NaCl, 1 mM EDTA, 1% [v/v] Triton-X-100, 0.1% [w/v] sodium deoxycholate, 0.1% [w/v] sodium dodecyl sulfate [SDS]) supplemented with 2 mM PMSF, 5 mM benzamidine and 1× Complete EDTA-free (Roche). Cells were lysed with a cell mill (Retsch MM200, 25Hz, 5min) in liquid nitrogen pre-cooled grinding jars and the powdery cell extract thawed on ice. 1 ml aliquots of the resulting suspension were sonicated on ice; sonication was set to yield a DNA average size of 400–500 bp. After centrifugation (17000g, 15 min, 4°C) the supernatant was used as the input sample in immunoprecipitation experiments. For each experiment, 400 µl aliquots of the input sample were incubated with 25 µl monoclonal anti-HA-agarose beads (Sigma-Aldrich) over night at 4°C on a rotating wheel.
Washing of beads and recovery of the immunoprecipitated DNA was done according to the ChIP protocol from the Haber Lab (http://www.bio.brandeis.edu/haberlab/jehsite/protocol.html) with the following modifications. All washing steps were carried out at 4°C and repeated one more time, with exception of the TE wash. Proteinase K treatment was done with 50 µl TE containing 3.5 mg/ml Proteinase K without glycogen, and phenol/chlorophorm extraction was done without LiCl. Samples were analysed by qPCR on a Bio-Rad iCycler using the Mesa Green qPCR MasterMix Plus (Eurogentec) with 400 nM Primer (each) and 1 µl immunoprecipitated DNA or 1/100 diluted input DNA, respectively. Amplicons were normalized to input DNA using the Bio-Rad IQ5 software.
Custom-designed Affymetrix chips were used for DNA-array analysis. Probe sets for the individual genes are visualized at http://mips.helmholtz-muenchen.de/genre/proj/ustilago/Target preparation, hybridization and data analysis was performed essentially as described before [40], with the following alterations: 5 µg RNA were used for first strand cDNA synthesis at 50°C with Superscript II (Invitrogen); for all experiments, an adjusted P-value for the false discovery rate [41] of ≤0.01 and a change in expression of ≥2 was used for filtering. For analysis of b-dependent gene expression strain AB31 (a2 Pcrg1:bE1/bW2) was compared to strain AB32 (a2 Pcrg1:bE2/bW2) and strain AB33 (a2 Pnar1:bE1/bW2) was compared to strain AB34 (a2 Pnar1:bE2/bW2) at the given time points. For analysis of rbf1-dependent gene expression strain AB31 (a2 Pcrg1:bE1/bW2) was compared to AB31Δrbf1 (a2 Pcrg1:bE1/bW2 Δrbf1) and strain CP27 (a2 Δb::Pcrg1:rbf1) was compared to strain JB2 (a2 Δb) at the given time points. Expression values were calculated as mean of two biological replicates. All array data have been submitted to GEO/NCBI (GSE18750, GSE18754 and GSE18756).
De novo promoter motif search was performed using the TAMO framework [42] extended to include AlignAce, Bioprospector, Cismodul, Improbizer, Meme, MDScan and Weeder. Output of each algorithm was collected, converted into a position weight matrix and scored with a hypergeometric test reflecting a random selection null hypothesis [43].
Flow cytometry measurements were performed as described before [20]. Cell counting was performed with a Neubauer counting chamber.
Microscopic analysis was performed using an Axioimager equipped with an Axiocam MRm camera or a Lumar V12 equipped with an Axiocam HRc (Zeiss, Jena, Germany). Nuclei were stained with DAPI Vectashield H-1200 (Vector Laboratories), fungal cell walls with 2 µg/ml Calcofluor white (Sigma, St. Louis, MO) in PBS. All images were processed with Axiovision (Zeiss, Jena, Germany).
clp1 (um02438) XP_758585, rbf1 (um03172) XP_759319, hdp1 (um12024) XP_761909.1, hdp2 (um04928) XP_761075, biz1 (um02549) XP_758696, cln1 (um04791) XP_760938, clb1 (um03758) XP_759905, clb2 (um10279) XP_758735, cdr2-like protein (um03928) XP_760075, pcl12 (um10529.2) XP_760585, DNA polymerase epsilon (um01008) XP_757155, DNA replication licensing factor (um06402) XP_762549.
|
10.1371/journal.pcbi.1005785 | The influence of astrocytes on the width of orientation hypercolumns in visual cortex: A computational perspective | Orientation preference maps (OPMs) are present in carnivores (such as cats and ferrets) and primates but are absent in rodents. In this study we investigate the possible link between astrocyte arbors and presence of OPMs. We simulate the development of orientation maps with varying hypercolumn widths using a variant of the Laterally Interconnected Synergetically Self-Organizing Map (LISSOM) model, the Gain Control Adaptive Laterally connected (GCAL) model, with an additional layer simulating astrocytic activation. The synaptic activity of V1 neurons is given as input to the astrocyte layer. The activity of this astrocyte layer is now used to modulate bidirectional plasticity of lateral excitatory connections in the V1 layer. By simply varying the radius of the astrocytes, the extent of lateral excitatory neuronal connections can be manipulated. An increase in the radius of lateral excitatory connections subsequently increases the size of a single hypercolumn in the OPM. When these lateral excitatory connections become small enough the OPM disappears and a salt-and-pepper organization emerges.
| Columns of neurons in the primary visual cortex (V1) are known to be tuned to visual stimuli containing edges of a particular orientation. The arrangement of these cortical columns varies across species. In many species such as in ferrets, cats, and monkeys a topology preserving map is observed, wherein similarly tuned columns are observed in close proximity to each other, resulting in the formation of Orientation Preference Maps (OPMs). The width of the hypercolumns, the fundamental unit of an OPM, also varies across species. However, such an arrangement is not observed in rodents, wherein a more random arrangement of these cortical columns is reported. We explore the role of astrocytes in the arrangement of these cortical columns using a self-organizing computational model. Invoking evidence that astrocytes could influence bidirectional plasticity among effective lateral excitatory connections in V1, we introduce a mechanism by which astrocytic inputs can influence map formation in the neuronal network. In the resulting model-generated OPMs the radius of the hypercolumns is found to be correlated with that of astrocytic arbors, a feature that finds support in experimental studies.
| The cortex is the outermost layer of cerebral tissue, composed of neuronal cell bodies and protoplasmic astroytes. The neurons in the cortex are arranged in columns, and the neurons in each column usually respond to similar features. In the macaque these columns, known as microcolumns or minincolumns have a density of 1270 minicolumns per mm2, with each minicolumn having around 142 pyramidal cell bodies [1]. Now the 3-d volume of cortical tissue could be locally approximated as a 2-d sheet of nodes, with a single node representative of all the neurons within a particular column. With this approximation it becomes possible to describe a 2-d map in the neuronal space with each node responding to a particular feature in the stimulus space.
A number of such stimulus modality-specific feature maps are topographic in nature, meaning that features that are similar in the stimulus space are mapped onto neighboring locations in the cortical space. A few examples include the tactile map in the primary somatosensory cortex [2], the whisker map in the barrel cortex [3], and the orientation, direction and retinotopic maps in the primary visual cortex [4]. Understanding the mapping function allows prediction of what features a particular neuron will respond to. A model which simulates the development of such maps, would aid in understanding which factors contribute to the development of such features maps. These factors could include internal factors such as the connectivity between the nodes, or the available area of the cortex onto which the features are to be mapped. Similarly features of the stimuli used for training the model themselves act as external factors.
Self-organizing maps (SOMs) have been used extensively to simulate the development of cortical maps [5–12]. A SOM has two constraints: coverage and continuity. Optimal coverage implies all input stimuli are mapped evenly on to the output space. Continuity implies that neighboring neurons in the output space respond to similar stimuli. The SOM uses local learning rules in order to optimize coverage and continuity. A biologically realistic variant of SOM, namely the Gain Controlled Adaptive Lateral (GCAL), has been used to investigate the factors involved in the development of a number of feature maps in the primary visual cortex (V1) [13]. The GCAL model consists of sheets of neurons. Each neuron in each layer could have 3 kinds of connections, each of which is trained using a normalized Hebbian learning rule:
A common feature in most SOMs is the presence of a mechanism by which neighboring neurons respond to similar features whereas those further away respond to dissimilar ones. The GCAL model achieves this by having short range excitatory connections, and longer range inhibitory connections. However in V1 inhibitory connections are local (short range) and may dominate responses [14], whereas the long range connections are excitatory. The effective long range inhibition is achieved by excitatory neurons synapsing onto inhibitory neurons which in turn synapse onto other neurons in its vicinity. For high contrast stimuli, it is known that the long range connections are in effect (multi-synapse) inhibitory in nature [15]. From a computational perspective, the radius of the effectively short range excitatory connections is important in determining the size of the orientation hypercolumn [6, 16]. In the absence of any excitatory connections, with Hebbian trained afferent connections and anti-Hebbian trained lateral connections, a sparse representation yielding independent components of the training set is realized [17]. This implies that an OPM will give way to a salt-and-pepper organization, without a smooth shift in orientation preference among neighboring neurons, in the absence of lateral excitatory connections. OPMs are present in carnivores (such as cats and ferrets) and primates but absent in rodents [18]. The term ‘salt-and-pepper’ was originally used to describe the maps seen in rodents, since the orientation preference of neighboring neuronal columns appeared to be uncorrelated and resembled a random pattern. However, recent experimental evidence suggests that the map is pseudo-random and exhibits some local similarities in orientation preference [19].
We hypothesize a possible link between astrocytic arbors and presence of OPMs and try to show that larger astrocytic arbors are more conducive to the generation of OPMs. We investigate the above hypothesis using computational modeling. We propose a GCAL model having 2 V1 layers: one representative of neurons, whereas the other of astrocytes. The synaptic activity of V1 neurons is given as input to an astrocyte layer. The activity of the astrocyte layer is now used to modulate bidirectional plasticity of lateral excitatory connections in the V1 layer. By simply varying the radius of astrocytes, the effective extent of lateral excitatory neuronal connections can be manipulated. An increase in the effective radius of lateral excitatory connections subsequently increases the size of a single hypercolumn in the OPM. When these effective lateral excitatory connections become small enough the OPM disappears and a salt-and-pepper organization emerges.
Hubel and Wiesel proposed that the emergence of orientation preference in principal (layer 4) neurons in the primary visual cortex is primarily due to the spatial arrangement of LGN afferent connections [20], though the effect of recurrent connections is now also clear. This contribution of afferent feed-forward connections is also emphasized by Paik and Ringach, who attribute the development of orientation preference maps across species to the Moire interference patterns created due to the spatial arrangement of Retinal Ganglion Cells (RGCs) [21]. While the contribution of feed forward connections to map formation is undeniable, as verified by a number of experiments, the contribution of recurrent lateral connections between cortical columns is also prominent. At the level of columns, rather than at the level of single neuron, it is known that for high contrast inputs, due to the recruitment of local inhibitory inter-neurons, long range lateral connections are predominantly inhibitory in nature [15, 22, 23]. This configuration of lateral connections is essential for map formation [7]. What shapes the lateral circuitry in cortical networks? Are there mechanisms which could ensure that short range connections are excitatory, whereas as the long range connections are in effect (considering the contribution of interneurons) predominantly inhibitory?
We hypothesize that protoplasmic astrocytes could play a key role in this regard. Although there are a number of mechanisms by which astrocytes and neurons communicate with each other [24–27], not all these mechanisms contribute to long term plasticity, crucial for the development of cortical maps. It must however be noted that there are a number of ways in which astrocytes could influence long term plasticity. These mechanisms could be summarized as follows:
In each of these mechanisms the effective synaptic strength is influenced by astrocytic activity. NMDA-dependent LTP/LTD is known to be a function of the postsynaptic calcium influx [34]. The postsynaptic calcium influx is likely dependent on astrocytic activity as well. The astrocytic influence could be abstracted using a plasticity or learning rule (such as a BCM curve), where the threshold controlling LTP vs. LTD is dependent on the astrocytic activity. The lateral excitatory connections in the modified GCAL model are modeled in such a manner.
The Gain Control, Adaptation, Laterally Connected (GCAL) model, has been used to develop stable and robust orientation maps [35]. This model builds on the LISSOM model and has, as the name suggests, a mechanism which ensures gain control of input activations and homeostatic adaptation of weights. The model has 3 layers: a photo-receptive input layer, an ON/OFF LGN layer and a V1 layer.
The activity of the ON/OFF LGN layer is given as L for a node i, j in the layer.
L i , j ( t + 1 ) = f ( γ o ∑ a , b x a , b ( t ) C i j , a b k + γ s ∑ a , b L i , j ( t ) C i j , a b s ) (1)
where (a, b) denotes a neuron in the receptive field of the (i, j)th neuron in the output layer, with input given as xab; Cij, ab represents the weight from the (a, b)th neuron to the (i, j)th neuron. A constant multiplier to the overall strength is given by γo; γs represents the gain-control. The weights Cij, ab are defined as a difference of Gaussians.
C i j , a b = 1 Z c e x p ( - ( a - i ) 2 + ( b - j ) 2 2 σ c 2 ) - 1 Z s e x p ( - ( a - i ) 2 + ( b - j ) 2 2 σ s 2 ) (2)
where Zc, and Zs denote the normalization factors, σc, and σs regulate the width of the gaussians.
The term C i j , a b s denotes the lateral inhibition received from other ON/OFF units.
C i j , a b s = 1 Z s e x p ( - ( a - i ) 2 + ( b - j ) 2 2 σ c 2 ) (3)
The firing rate of a V1 neuron is dependent on only 3 kinds of inputs, namely: afferent inputs from the LGN (Lab(t − 1)), lateral effectively excitatory inputs, and lateral effectively inhibitory inputs. Thus the firing rate (yij(t)) is given as:
y i j ( t ) = f ( p ∑ a , b A i j , a b L a b ( t - 1 ) + q ∑ k , l E i j , k l y k l ( t - 1 ) - r ∑ k , l I i j , k l y k l ( t - 1 ) ) (4)
where p, q, r are scaling factors; Aij, ab is the afferent weight from neuron (a, b) to neuron (i, j); Eij, kl is the lateral excitatory weight from neuron (k, l) to neuron (i, j) and similarly Iij, kl is the lateral inhibitory weight from neuron (k, l) to neuron (i, j). The function f is a half wave rectifier in order to ensure that the activations are positive with a variable threshold point given as ρ. The activations yij(t) are allowed to adapt in a homeostatic fashion. The output activity yij and the threshold ρ are adapted as follows:
y ¯ i j ( t ) = ( 1 - β ) y i j ( t ) + β y ¯ i j ( t - 1 ) (5) ρ ( t ) = ρ ( t - 1 ) + λ y ¯ i j ( t ) - μ (6)
where β is the smoothing parameter and λ is the homeostatic learning rate; y ¯ i j ( t ) is initialized to the average V1 activity (μ).
In order to model astrocytic activation we simulate an additional layer whose input is the synaptic activity (gs) present at each node of the V1 layer. Thus the activation of a single node in this astrocyte layer is given by Sij.
g s i j ( t ) = p ∑ a , b A i j , a b L a b ( t - 1 ) + q ∑ k , l E i j , k l y k l ( t - 1 ) - r ∑ k , l I i j , k l y k l ( t - 1 ) (7) S i j ( t ) = ∑ i , j ∈ R a s t r o g s i j ( t - 1 ) (8)
where the radius of the astrocyte is given as RAstro. There is some debate regarding the precise nature of GABA induced calcium oscillations in the astrocyte and the subsequent gliotranmission [36]. Hence we run an additional simulation which does not consider the effect of GABA induced gliotransmitters. Now the activation of a single node in the astrocyte layer is given by Sij.
g s i j ( t ) = p ∑ a , b A i j , a b L a b ( t - 1 ) + q ∑ k , l E i j , k l y k l ( t - 1 ) (9) S i j ( t ) = ∑ i , j ∈ R a s t r o g s i j ( t - 1 ) (10)
The lateral inhibitory and afferent weights are trained using the same normalized Hebbian rule given by:
w i j , m n ( t + 1 ) = w i j , m n ( t ) + η y i j ( t ) P m n ( t ) ∑ m n ( w i j , m n ( t ) + η y i j ( t ) P m n ( t ) ) (11)
where Pmn is the generalized notation for the pre-synaptic activity originating from the neuron (m, n); η is the learning rate. These learning rates can be different for each of the connections: ηA, ηE and ηI are the learning rates for the afferent, excitatory and inhibitory connections respectively.
However the lateral excitatory connections adapt using a variant of the BCM rule with a threshold function θ being a function of the astrocytic activation at the corresponding node. It has been previously proposed that astrocytes introduce metaplasticity by shifting the BCM curve [31].
E i j , k l ( t + 1 ) = E i j , k l ( t ) + η E y i j ( t ) ( y i j ( t ) - θ i j ) y k l ( t ) (12) θ i j = ( 1 - S i j ) (13)
Astrocytes communicate with each other via gap junctions; however only distal branches are connected, resulting in astrocytic microdomains with less than 10% overlap [37]. The gap junctions could be modeled using Gaussian random lateral excitatory connections to the 8 nearest neighboring nodes. A schematic of the model is shown in Fig 1
The parameters used for the GCAL model are a superset of those used in the standard LISSOM model. The complete list of parameters are given in Table 1. The simulations are performed using the Topographica simulator [38].
We vary the astrocytic radius and observe the changes in the orientation map developed. The experimentally reported astrocytic radii are estimated using the Glial fibrillary acidic protein (GFAP) as the astrocytic marker. However, the GFAP marked region accounts for only 15% of the actual astrocytic volume. Hence we scale the astrocytic radii by a factor of 2 in the simulations. The model is trained for 10000 iterations. The training regime consists of elongated 2-dimensional Gaussians with centers and orientations drawn from a uniform random distribution.
The astrocytic radius is varied and the corresponding orientation maps developed are studied (Figs 2 and 3). It is observed that on reducing the astrocytic radius, the periodicity of the map increases and the width of a single hypercolumn decreases. Thus in a given area of cortical tissue 3 x 3 mm, the number of orientation hypercolumns would increase as we reduce the astrocytic radius. The neuronal and astrocyte maps developed have similar orientation preferences which could be quantified by their stability index. The stability index between the astrocytic and orientation maps is shown in Fig 4. These results demonstrate that the astrocyte radius has a profound effect on OPM formation.
The development of a few of these maps and their stability indices across iterations are shown in Figs 5, 6, and 7. The V1 orientation preference map is probed at 250, 500, 750, 1000, 2500, 5000, 7500 and 10000 iterations. It is observed that the map developed becomes stable after a few initial iterations, as quantified by the corresponding stability indices. These results demonstrate the model develops stable orientation maps.
We also simulate 2 additional conditions which could effect the development of the orientation maps: (1) Considering there is no GABA induced gliotransmission: Since the effect of GABA induced calcium oscillations is not well understood in literature, we also simulate the map development ignoring the corresponding term as described in Eq 9. (2) Considering the effect of gap junctions in the astrocyte layer: The basic simulation does not consider the effect of gap junctions among astrocytes. As described in the methods section, we introduce gap junction by considering excitatory connections among the nearest neighbors in the astrocyte layer. The maps formed for these 2 conditions are shown in Figs 8 and 9 respectively. The maps developed using all 3 conditions (basic, no GABA, Gap junctions) appear visually similar and their features, which are further quantified (See Fig 10), show a similar trend. These results indicate that the correlation between astrocyte radius and hypercolumn widths is robust for all the conditions considered.
The number of pinwheels observed in the neuronal (V1) orientation map in the simulated region (3 x 3 mm) is shown in Fig 10(A). As expected, the number of pinwheels falls with increasing astrocytic radius. The number of pinwheels per hypercolumn remain approximately constant, centered around π for the maps in which a clear orientation preference map (OPM) structure is present (Fig 10(B)). However for smaller astrocytic radii the map begins to disintegrate. These results strengthen the hypothesis that the astrocytic radii influence the formation of orientation maps. For higher astrocytic radii the number of pinwheels per hypercolumn stabilizes to values around π. This result is in keeping with experimental findings which show that the number of pinwheels per hypercolumn is a constant π across species [39]. The trend observed in the simulated widths of the hypercolumn and the corresponding astrocytic radii are comparable with the scant experimental evidence available as shown in Fig 10(C).
The transition from a salt and pepper kind of map to a smooth orientation map could be quantified using 2 methods: (1) Change in the number of pinwheels/ hypercolumn: Experiment results indicate that the number of pinwheels/ hypercolumn remains a constant across species, even with differing hypercolumn widths [39]. Thus, if such a ratio is no longer maintained, the map developed no longer resembles a smooth orientation preference map. However, the map developed is also not truly random since there might be local patches with similar orientation preference. A recent study has shown that in rodents the map only appears to be random, and has significant local orientation similarity [19]. (2) Local similarity in orientation preference: This method quantifies the local smoothness of the map developed. The mean angle of separation between the orientation preference of a node and all others within a predefined radius of interest is computed and compared for different astrocytic radii.
We then compare the results using the 2 methods and observe that a sharp transition between salt and pepper and a smooth orientation maps is absent (Fig 11). Rather, an intermediate state which exhibits local patches of orientation similarity, but lacks the features of a true orientation map is seen.
A fascinating feature of orientation mapping is that not all species display a smooth transition in orientation preferences as we probe along the cortical surface. Rodents, in particular have neuronal columns which are orientation specific but arranged in a seemingly randomized fashion across V1. This kind of organization is referred to as a salt and pepper configuration. The presence or absence of OPMs and their potential consequences for information processing is a topic of current interest.
Another interesting fact in those species which do have OPMs is that the size of the hypercolumn varies from species to species. However the number of pinwheels per hypercolumn appears to remain constant across species. Self organizing mechanisms have been extensively utilized to model OPMs. These models rely on a mechanism that ensures that neighbouring neuronal columns respond to similar features, whereas distant ones to different features. This is invariably implemented by invoking local excitatory and larger inhibitory connections. However cortical inhibitory connections are known to be short range, whereas excitatory ones are longer ranged laterally. These long ranged inhibitory connections have been explained away as long ranged excitatory neurons recruiting local inhibitory neurons, such that the net effect is inhibitory. However, a mechanism that ensures that short range connections are effectively more excitatory than inhibitory has proven elusive. The arguments summarized above have been discussed in detail by Swindale [7]. He postulates the possibility of extracellular diffusion of chemical messengers mediating this short range excitatory connectivity.
However there are a number of issues with the diffusion hypothesis. Firstly diffusion, a passive process, would ensure roughly similar excitatory radii across species and would thus imply by extension similar widths of orientation hypercolumns across species. In reality the widths of hypercolumns vary widely across species. Rodents do not have a smooth topographical variation in orientation preference and hence do not have defined hypercolumns [18]. Thus diffusion alone would not explain the variation in hypercolumn widths.
Secondly the pyramidal apical dendrites, which are used to define the width of a minicolumn (also called microcolumn) are roughly the same (≈ 30μm) for the rhesus macaque and the rat [40]. Thus the chemical messenger which diffuses should have similar effects at the level of cortical columns. However as stated earlier this again does not hold true.
Astrocytes are known to regulate both excitatory and inhibitory cortical circuits, via a combination of glutamate and GABA re-uptake by transporters, gliotransmitter release, and regulation of neuronal excitability [27–29, 31, 41, 42]. Indeed, optogenetic astrocyte calcium activation modulates the excitatory-inhibitory balance and increases response selectivity of excitatory neurons within local cortical microcircuits [28]. Thus the extent of astrocyte influence may directly influence the range of local influence in the cortex.
As mentioned earlier, protoplasmic astrocytes are thought to contribute to metaplasticity [30]. As shown in Fig 12, astrocytes release gliotransmitters which are known to shift a BCM like curve to the left, implying greater LTP for lower postsynaptic firing rates in excitatory neurons [31]. This constitutes a greater increase in synaptic strength for those synapses in the vicinity of the gliotransmitter releasing astrocyte.
Now, astrocytes are understood to release gliotransmitters in correlation with their internal calcium levels [30]. Glutamate in the synaptic cleft, either via receptors or transporters, mediates the calcium levels in the engulfing astrocyte. Astrocytes associated with synapses corresponding to those layer 4 pyramidal neurons, which receive direct thalamic input, would have greater calcium levels as compared to other astrocytes. Hence the metaplasticity induced in the neighboring excitatory synapses in the domain of these astrocytes would also be more pronounced. Over time, this would lead to a greater excitatory drive for those neurons in the vicinity of the neuron receiving the direct thalamic input, as shown in Fig 13.
We hypothesize that astrocytes influence local synapses and specify the radius of lateral excitatory connections. This in turn influences the size of hypercolumns across species. A comparative table specifying hypercolumn widths and astrocyte radii is specified in Table 2. In the standard GCAL model there is a constraint placed on the maximum radius of the lateral excitatory connections [35]. Indeed, this constraint is necessary to ensure that the lateral excitatory connections are shorter in range than the lateral inhibitory ones. Such a configuration is essential for the development of the self organized orientation map. In our proposed model the lateral excitatory connections have no defined maximum radius. The limit on the lateral excitatory connections is implicitly imposed due to the fact that the astrocytes control the BCM threshold of the excitatory neuronal synapses within their (astrocyctic) radius of influence and ensures LTP. For synapses outside the astrocytic radius, the threshold is such that LTD occurs and these connections are pruned off automatically.
As a computational principle, any mechanism that can control the lateral excitatory radius with respect to the lateral inhibitory radius, could produce similar maps as those shown in this manuscript. However, several lines of evidence indicate that astrocytes are strongly involved in this mechanism. First, astrocytes have been shown to regulate the excitatory to inhibitory balance in local neuronal circuits [28]. Importantly, astrocytes express transporters for both glutamate and GABA, and can thus regulate the strength of both excitatory and inhibitory synaptic transmission. Second, they have a significant role in regulating local synaptic plasticity, in particular local neuronal excitation, via a range of mechanisms that include modulation of NMDA receptors as well as integrating other plasticity-mediating neuromodulators such as acetylcholine, noradrenaline [43, 44]. Together, these effects are well placed to implement the BCM rule. Third, astrocytes are known to form microdomains with less than 10% overlap [37]. Thus, the astrocytic organization automatically results in the formation of local domains, which are influenced by these transmitters/modulators. Fourth, the radii of these local domains are known to be less than the effective lateral inhibitory radius, thus resulting in the required short range excitation and longer range inhibition.
We simulate the development of orientation maps with varying hypercolumn widths, by simply varying the radius of astrocytic connections using the LISSOM model with an additional layer simulating the astrocytic activation. We observe that increasing the astrocytic radius, and thereby the effective radius of lateral excitatory connections in the V1 neuronal layer, the width of the hypercolumn developed shows a proportionate increase. When the effective lateral excitatory radius is reduced so as to almost prevent any similarity in orientation preference of neighboring neurons, the OPM disappears and a salt and pepper configuration of neuronal arrangement of orientation preference is seen.
|
10.1371/journal.pcbi.1002511 | Google Goes Cancer: Improving Outcome Prediction for Cancer Patients by Network-Based Ranking of Marker Genes | Predicting the clinical outcome of cancer patients based on the expression of marker genes in their tumors has received increasing interest in the past decade. Accurate predictors of outcome and response to therapy could be used to personalize and thereby improve therapy. However, state of the art methods used so far often found marker genes with limited prediction accuracy, limited reproducibility, and unclear biological relevance. To address this problem, we developed a novel computational approach to identify genes prognostic for outcome that couples gene expression measurements from primary tumor samples with a network of known relationships between the genes. Our approach ranks genes according to their prognostic relevance using both expression and network information in a manner similar to Google's PageRank. We applied this method to gene expression profiles which we obtained from 30 patients with pancreatic cancer, and identified seven candidate marker genes prognostic for outcome. Compared to genes found with state of the art methods, such as Pearson correlation of gene expression with survival time, we improve the prediction accuracy by up to 7%. Accuracies were assessed using support vector machine classifiers and Monte Carlo cross-validation. We then validated the prognostic value of our seven candidate markers using immunohistochemistry on an independent set of 412 pancreatic cancer samples. Notably, signatures derived from our candidate markers were independently predictive of outcome and superior to established clinical prognostic factors such as grade, tumor size, and nodal status. As the amount of genomic data of individual tumors grows rapidly, our algorithm meets the need for powerful computational approaches that are key to exploit these data for personalized cancer therapies in clinical practice.
| Why do some people with the same type of cancer die early and some live long? Apart from influences from the environment and personal lifestyle, we believe that differences in the individual tumor genome account for different survival times. Recently, powerful methods have become available to systematically read genomic information of patient samples. The major remaining challenge is how to spot, among the thousands of changes, those few that are relevant for tumor aggressiveness and thereby affecting patient survival. Here, we make use of the fact that genes and proteins in a cell never act alone, but form a network of interactions. Finding the relevant information in big networks of web documents and hyperlinks has been mastered by Google with their PageRank algorithm. Similar to PageRank, we have developed an algorithm that can identify genes that are better indicators for survival than genes found by traditional algorithms. Our method can aid the clinician in deciding if a patient should receive chemotherapy or not. Reliable prediction of survival and response to therapy based on molecular markers bears a great potential to improve and personalize patient therapies in the future.
| In the past decade, several studies have used microarray gene expression data from tumors to predict the clinical outcome of patients (see Table S1). The tumors included breast cancer [1]–[5], lung cancer [6]–[8], lymphomas [9]–[13], leukemia [14], [15], and others [16]–[18]. Outcome is usually measured by categorical, often binary variables such as survival up to a certain time, recurrence of tumor or metastasis before a certain time, or success of treatment. Predicting such variables from gene expression levels can be viewed as a classification problem, and the set of genes used for prediction is commonly referred to as a signature. Accurate outcome prediction can be used clinically to select the best of several available therapies for a cancer patient. For instance, a low risk patient can be advised to select a less radical therapy.
Whereas differences in gene expression between tumor and healthy tissue or between different tumor tissues are often strong, gene expression differences between patients with the same type of tumor but different outcome are more subtle. For example, distinguishing acute myeloid from acute lymphoblastic leukemia has been demonstrated to be up to 100% accurate using only a few genes [19]–[21]. In contrast, outcome prediction is a much harder problem, with classification accuracies commonly in the range of 50–70%. It is therefore not surprising that many studies suffer from one or several of the following three problems: (i) limited or overoptimistic prediction accuracy, (ii) limited reproducibility, and (iii) unclear biological relevance of the genes used for prediction. For example, an early study predicting breast cancer metastasis using 70 genes [3] was subsequently found (i) to have lower than initially published predictive accuracy on the same or independent data sets [22], [23], (ii) to be difficult to reproduce [24], and (iii) to have used 70 genes that can be easily replaced by 70 different but equally predictive genes derived from the same data, questioning the biological relevance of the particular 70 genes of the original study [25]. Notably, predictive gene sets derived from different studies for the same disease show almost zero overlap, questioning their biological relevance.
To address and overcome these problems we have developed a computational network-based strategy for outcome prediction. Our algorithm, NetRank, couples gene expression measurements with a network of known relationships between the genes' products. NetRank is based on Google's PageRank algorithm [26]. PageRank uses the hyperlink information between web documents to better decide which documents are the most relevant ones. Similarly, NetRank uses biological interaction information between genes' products to better decide which genes are the most relevant for outcome prediction. Such interaction information is available in protein–protein, transcription factor–target, or gene co-expression networks. The inclusion of network information serves two purposes. First, gene products with many interactions should have a higher biological relevance since they can exert a bigger influence on a biological system. Second, considering network neighbors can help the algorithm to ignore correlations between expression and outcome that have no underlying biological causality. Such correlations can arise simply by chance, often due to the fact that microarray measurements are noisy and that the number of samples is typically several orders of magnitudes smaller than the number of genes investigated.
We wanted to test the NetRank idea on outcome prediction for pancreatic cancer, for which no microarray-derived signature was yet published. Pancreatic ductal adenocarcinoma accounts for approximately 130,000 deaths each year in Europe and the United States [27], [28]. It has an extremely poor prognosis with a 5-year survival rate below 2% [29], [30]. Currently, only a few prognostic factors for pancreatic cancer survival are used in the clinical setting, among them CA 19-9, alkaline phosphatase, LDH, levels of white blood cells, aspartate transaminase, and blood urea nitrogen [31]. A considerable number of protein markers for pancreatic cancer prognosis have been investigated using immunohistochemistry [32]. However, the clinical value of most of these markers remains to be determined, and also most of these markers were found by chance or educated guesses rather than a systematic, genome-wide approach.
The aim of our study was therefore (i) to carry out a genome-wide screen for genes whose expression in pancreatic cancer tissue samples reliably correlates with the patient survival time, and (ii) to use these genes as a molecular signature for reliable survival prediction. To this end, we collected and analyzed tissue samples from patients with pancreatic ductal adenocarcinoma from Germany in a multi-center study. Applying NetRank to gene expression profiles of these samples identified seven candidate marker genes prognostic for outcome. To assess the clinical value of our identified marker genes, we validated them on an independent patient cohort. We found that signatures based on these markers were more accurate than traditional clinical parameters and more accurate than signatures identified with other computational approaches.
We obtained gene expression profiles of sufficient quality from 30 pancreatic ductal adenocarcinoma samples of patients that underwent surgery in German university hospitals between 1996 and 2007, hereafter referred to as the screening dataset (see Materials and Methods for quality criteria and experimental details). For each patient, the clinical parameters age, sex, cancer staging according to the tumor-node-metastasis (TNM) classification, and survival time after operation were recorded. Table 1 shows an overview of the patient characteristics. For predicting the prognosis of a patient, we assigned patient samples to either a poor or a good prognosis group depending on patient survival time. Such an assignment is straightforward if the survival time is bimodally distributed. However, such a bimodal distribution is often absent in cancer patient survival times (see Figure S1). In this case, a common choice is to split by the median survival time independently of the distribution (see for example [13], [33]), which has the advantage of yielding two classes of equal size. Thus, we defined prognosis as poor if the patient survival time was less than the median survival time of 17.5 months, and as good otherwise. This resulted in two prognosis groups with 15 patient samples each. The goal was to identify a signature of genes whose expression levels allowed to correctly predict the prognosis group of a patient. Since we wanted the resulting signature to be applicable in a clinical setting using immunohistochemistry staining of the signature proteins, we opted for a signature size in the order of five to ten genes.
To identify signature genes, various state of the art methods exist. To evaluate these methods in comparison to our own NetRank method on the screening dataset, the following workflow was employed (see Figure 1). After filtering out low expression and low variance genes, 8,000 genes remained as potential signature genes. Five different methods for ranking genes according to their power to discriminate between the two prognosis groups were tested: (i) fold change, as defined by the ratio of a gene's mean expression in one group over the other group, (ii) the t-statistic, (iii) Pearson and Spearman rank correlation coefficients of a gene's expression with the survival time of the patient, (iv) the SAM (Significance Analysis of Microarrays) method [34], and (v) our NetRank algorithm (see Materials and Methods for details). In addition, selecting genes randomly was included as a control method. For each method, a support vector machine classifier was trained using the 5–10 top ranked genes as features. Prediction accuracy as defined as the percentage of correctly classified samples was evaluated with different training and test set sizes. All feature selection and machine learning steps were subjected to Monte Carlo cross-validation, which is a recommended and relatively un-biased evaluation strategy [22], [35] (see Figure 1 and Materials and Methods for details).
We introduce NetRank, a modified version of the PageRank algorithm [36]. As employed by the Google Internet search engine, the PageRank algorithm uses network information (hyperlinks) between documents in the world wide web to assess the relevance of a document. A document is important if it is highly cited by other documents. Moreover, citations from important documents have more weight than citations from unimportant documents. Thus, in order to measure the relative importance of a document within the set of all web documents, PageRank ranks a document according to the number of highly ranked documents that point to it. Similarly, NetRank assigns a score to a gene which is influenced by the scores of genes linked to it. This linkage can be defined in several ways. Morrison et al. [37] described an adaptation of PageRank which uses networks where genes are connected if they share a Gene Ontology annotation. Here, we employ known transcription factor–target relationships (from TRANSFAC [38]), protein–protein interaction (from HPRD [39]), and gene co-expression (from COXPRESdb [40]) to define three different gene–gene networks, which were used with NetRank. NetRank first assigns as a score for each gene the absolute correlation of its mRNA expression level with the patient survival time in the dataset. The network is then used to spread this correlation to its neighbors and beyond. The genes with the highest NetRank score are then selected as signature genes (see Materials and Methods for details).
We first compared the three above mentioned networks for NetRank and found that signatures obtained using the TRANSFAC transcription factor network consistently had higher predictive accuracies than those using the protein interaction or the co-expression network. We therefore decided to only use the TRANSFAC network for NetRank in the following experiments.
For all training set sizes, signatures selected with NetRank using the TRANSFAC transcription factor–target network showed higher predictive accuracies than those selected by any of the other four methods (Figure 2A). NetRank showed a maximum accuracy of 72% ( standard error of the mean, s.e.m.) with a training set of 28 samples and a signature size of 7 genes. This compares favorably to studies in other cancers, which show accuracies in the range of 50–70%. We found that NetRank is especially beneficial for small training set sizes, where there is a 7% ( s.e.m.) improvement in accuracy compared to the Pearson correlation method.
Since many single prognostic markers for pancreatic cancer have been described in the literature, we next asked whether markers found with NetRank were superior to these literature markers. To this end, 51 markers identified via a literature search were used to train a support vector machine with different training set sizes (see Materials and Methods and Table S2). Surprisingly, we found that the NetRank markers showed on average a 12% higher accuracy than the literature markers (Figure 2B).
Using NetRank, we identified seven genes (STAT3, FOS, JUN, SP1, CDX2, CEBPA, and BRCA1) as most relevant for predicting survival in patients with pancreatic ductal adenocarcinoma. These seven marker candidates were validated in two ways: first, by quantitative RT-PCR of the screening dataset to confirm the microarray gene expression measurements, and second, by immunohistochemical analysis of protein levels in an independent dataset of 412 patients (the validation dataset, see Table 1).
Of our seven markers, we found high expression to be associated with shorter survival for STAT3, FOS, and JUN, and high expression associated with longer survival for SP1, CDX2, CEBPA, and BRCA1. This is in line with most previous studies. STAT3 is a well-known oncogene and persistently activated in many human cancers, including all major carcinomas [41]. FOS and JUN, which constitute the AP-1 transcription factor, have been linked to both tumor progression and suppression [42]. SP1 was reported to be associated with poor prognosis in gastric cancer and recently also in pancreatic ductal adenocarcinoma [43], [44]. BRCA1 is a DNA damage repair protein where loss-of-function mutations typically lead to early onset of breast cancer and ovarian cancer [45].
Figure 3A shows the direct network neighbors of the seven candidates. The network is shown in power graph representation, which reduces the number of edges drawn without information loss [46]. The underlying network of regulatory relationships was obtained from the TRANSFAC database [38]. The network shows that the seven markers (yellow) regulate a total of 323 targets. The correlation of the expression of a gene with the survival of the patient in the screening dataset is shown in red. Genes with larger circles were previously described in the literature as being associated with survival in pancreatic cancer.
The marker protein with the most regulatory interactions is the transcription factor SP1. Some of its targets are additionally regulated by other markers such as CEBPA, STAT3, FOS, and JUN. One interesting module defined by the genes that are regulated by SP1 and FOS is shown in Figure 3B. It contains many genes already known to be associated with survival in pancreatic cancer as well as some genes highly correlated with survival in our data, such as HBA1, F3, CCL2, IL2, and GJA1. A subset of this module is defined by genes that are also regulated by JUN. This subset contains the genes IL2, TGFB1, MT2A, and GJA1, which correlate well with survival.
Among the interaction partners of the markers, more than one-third has been previously reported as associated with survival or prognosis in pancreatic cancer, including PPARG, MUC4, and SMAD3 [47], [48]. A pathway analysis using KEGG [49] showed that 91 of the interacting genes are involved in signaling pathways, most prominently the MAPK and JAK-STAT signaling pathways. Furthermore, 53 of the interacting genes are involved in known KEGG cancer pathways (see Table S3).
To validate our findings, we analyzed protein levels of our markers in an independent set of 412 patients (the validation dataset, see Table 1). We wanted to test how well the proteins encoded by the marker genes are indicative for the survival of a patient when assessed by immunohistochemical staining of the patient's tumor. Using tissue microarrays, immunohistochemistry stainings were obtained for each of the seven marker proteins STAT3, FOS, JUN, SP1, CDX2, CEBPA, and BRCA1 for each patient in the validation dataset (see Figure S2). The predictive accuracies of the marker staining intensities (encoded in two levels, low or high, see Materials and Methods) were evaluated after training a support vector machine classifier in a leave-one-out cross-validation procedure. The classifier predicted patients to belong either to a low risk (good prognosis) group, or to a high risk (poor prognosis) group. Using backward elimination, starting from the full set of markers, markers were removed one at a time until the accuracy of the trained classifier failed to improve. The clinical parameters tumor size (T), regional lymph nodes (N), distant metastasis (M), histological grade (G) and residual tumor (R) were tested in the same manner for comparison.
Since some patients in the validation dataset received adjuvant therapy (mostly chemotherapy with gemcitabine), and the adjuvant therapy had an influence on survival time (although not quite as expected, see Figure S3), we split the validation dataset into a group of patients with and without adjuvant therapy. Note that the decision to treat a patient with adjuvant chemotherapy is so far not based on any molecular markers (which was one motivation for our study). Chemotherapy is part of the standard treatment for pancreatic cancer in Germany since many years and is recommended for every patient. However, patients in a reduced state of health and patients who refuse it will not receive chemotherapy.
The accuracies of our signatures are comparable to those found in other cancer studies. Stratford et al. [18] found six genes differentially expressed in tumors from pancreatic cancer patients with localized disease compared to metastatic disease using the significance analysis of microarrays (SAM) method [34]. Based on these six genes, they classified patients into high- and low-risk groups with 1-year survival rates of 55% and 91%, respectively. Our signatures classify patients into high- and low-risk groups with 1-year survival rates of 54% and 76%, respectively (adjuvant six-gene signature) and 55% and 69%, respectively (non-adjuvant five-gene signature). Unfortunately, Stratford et al. [18] did not report a classification accuracy percentage. Most surprisingly, although patients and methods were different, the six genes identified in their study and our seven genes share one gene of the Fos family. As mentioned before, there has hardly been any overlap among the signatures published so far for one tumor type. The discovery of FOS in both methods thus highlights its importance for tumor progression and outcome in pancreatic cancer, and further underlines the ability of our method to find reproducible and biologically significant markers.
NetRank depends on a number of parameters (see Materials and Methods for a full description): the choice of the genes' initial values that spread through the network, the damping factor which influences the amount of spread, the choice of the network, and the role of noisy and uninformative genes, which are filtered out. Next, we investigate NetRank's dependence on these parameters.
Here, we present a novel method for identifying prognostic markers from genome-wide gene expression data. A key feature of the method is that it judges the relevance of a gene as marker not only by its expression (or rather the correlation of its expression with survival), but also by the expression of its neighbors. Thus, it can detect and therefore avoid markers that correlate with survival simply by chance or noisy measurements, but not due to an underlying biological causality. We applied this method to microarray data from 30 freshly frozen samples of pancreatic ductal adenocarcinoma and obtained a prognostic marker set of seven genes. This set showed an accuracy of 72% in predicting the prognosis of a patient. To ensure validity of this result, we employed a rigorous Monte Carlo cross-validation procedure. We then validated these genes using high-throughput immunohistochemistry of samples from surgically resected tumors from an independent cohort of 412 patients; roughly half of these received adjuvant therapy. From the marker set we derived a six-gene signature for patients with adjuvant therapy and a five-gene signature for patients without adjuvant therapy. Both signatures improve prediction of patient prognosis compared to the use of clinical parameters when used for immunohistochemical staining of the tumor tissue. The additional predictive value of the signature markers compared to clinical parameters was 9% for patients with and 6% for patients without adjuvant therapy (as the best combination of clinical parameters only showed a predictive accuracy of 61% and 59%, respectively). Whereas the use of microarrays in clinical practice is limited by the large number of genes, complicated analytical methods, and the need for fresh-frozen tissue, RT-PCR or immunohistochemistry of a small number of proteins can be done routinely in a clinical setting.
Note that the samples were obtained during initial surgery, before any of the patients received adjuvant therapy. The expression signatures we identified predicted clinical outcomes specific for patients with and without adjuvant therapy. These signatures could be used to stratify patients for adjuvant treatment of the disease: A patient that is classified as low risk (good prognosis) by the adjuvant therapy signature should receive adjuvant chemotherapy treatment, whereas a patient that is classified as low risk by the no adjuvant therapy signature might have a longer survival without chemotherapy.
Our signature genes can also help to stratify pancreatic cancer patients for new therapies. STAT3 was found to be the best single prognostic marker, with a high expression of STAT3 indicating a high risk. STAT3 inhibitors might therefore be promising therapeutic agents. It is known that a large percentage of pancreatic cancers feature aberrantly activated STAT3 [59]. Very recently, novel STAT3 phosphorylation inhibitors were demonstrated to suppress growth in pancreatic cancer cell lines [60].
For breast cancer, an FDA approved microarray-based test that uses the 70-gene signature by van't Veer et al. [3] to assess the metastatic risk in patients with node negative breast cancer is commercially available and can be utilized clinically. In a validation study on an independent data set of 307 node-negative breast cancer patients [23], the 70-gene test was shown to have a sensitivity of 90% and a specificity of 42%. The accuracy, however, resulted in 50%, which is equivalent to guessing. The reported ROC curve for predicting time to distant metastases shows the same area under curve of 68% as our signature for predicting prognosis in patients with adjuvant therapy (Figure 4A). Using an appropriate cut-off, our signature also shows a high sensitivity of 83% with a specificity of 45% (upper right corner of the ROC curve, Figure S4A). It therefore can be used reliably to identify a group of patients who seem to benefit from the adjuvant therapy.
Our study emphasizes the benefit of systematic network-based approaches that incorporate background knowledge for identifying biologically relevant marker genes. Correlations between gene expression levels and a clinical variable of interest can arise simply by chance, without any underlying biological cause, especially with few patient samples. One example for such a spurious correlation in our screening dataset is the HBA1 gene, which encodes for hemoglobin alpha, and which showed a strong negative correlation with survival. Although HBA1 would have been a candidate marker when ranked merely by correlation, it was not ranked among the top ten markers by NetRank. Since we found the idea of a cancer tissue expressing hemoglobin interesting and worth exploring further, we decided to include HBA1 in the immunohistochemistry validation. However, immunohistochemical staining for hemoglobin in the validation dataset was incapable of defining significantly different risk groups. In addition, adding HBA1 to our signatures did not improve, but impaired their predictive accuracies. We conclude that the strong negative correlation of HBA1 expression levels with survival time in the screening dataset might have been caused by chance and not by any underlying biologically relevant causality. Network-based methods such as NetRank can add such causality for example in the form of known gene regulatory networks, resulting in the identification of markers that are more likely to be truly relevant.
Recent work by the Ideker lab also emphasized the benefit of network-based approaches [61]. Their study demonstrated that markers based on protein interaction subnetworks are first more reproducible than individual marker genes and second can improve classification accuracy for the van't Veer breast cancer dataset by 8% compared to the original 70 genes [3]. A further approach documenting the usefulness of networks employed PageRank to identify genes cross-talking between already published cancer genes [62].
During the progress of our study, another study [63] was published which used PageRank on protein interaction networks to improve recursive feature elimination for support vector machine learning. They found that this improved the prediction of ERBB2 status and relapse in breast cancer. However, neither of these two studies validated their genes on an external patient cohort to demonstrate the validity of markers found with PageRanking based on biological networks. Moreover, we found that the use of transcription factor–target networks yields more accurate signatures than the use of protein interactions networks in cancer outcome studies.
The use of background knowledge in order to get more robust and more biologically meaningful signatures comes at a price. It is in the nature of the NetRank algorithm to favor genes with many connections, since they can increase their ranking, whereas uncharacterized genes with no connections cannot. Hence, marker genes found with NetRank are more likely to be well known and well-described in the literature and less likely to be previously uncharacterized. We also found that the predictive accuracy of the immunohistochemistry -based markers was lower than that of the microarray-based markers. One potential bias stems from the different design of tissue microarrays, which vary in the number of cores per case, core size, and density. In addition, the semi-quantitative evaluation of the immunohistochemical staining tends to be less accurate and less objective than microarray-based gene expression profiling.
In conclusion, the expression signatures we identified predicted clinical outcomes in patients with surgically resected pancreatic ductal adenocarcinoma specific for patients with and without adjuvant therapy. Since these signatures could be used to stratify patients for adjuvant treatment of the disease, they are a potential additional piece of information in clinical decision making and can help to reduce costs, improve patient survival, and quality of life.
Two hundred forty-four freshly frozen tissue samples of pancreatic adenocarcinoma were obtained from surgical specimens from patients who underwent operations between 1996 and 2007 at German university hospitals in Berlin, Dresden, Heidelberg, Mannheim, Munich, and Regensburg. Informed consent was obtained from all patients included in this study. From each of the frozen tissue samples, 4 m slides were obtained, stained with hematoxylin and eosin, and re-evaluated by a pathologist (G. K.) experienced in pancreato-biliary pathology. Of these, 56 tissue samples contained tumorous tissue without any contamination from normal acini or islets and had suitable RNA quality. Of these, 30 were obtained from patients without any adjuvant therapy, and were used as the screening dataset. The clinical characteristics of this dataset are given in Table 1.
The validation dataset consisted of surgically resected PDAC samples from 517 patients who underwent operation between 1991 and 2008 at university hospitals in Berlin, Dresden, Jena, and Regensburg, Germany. Informed consent was obtained from all patients included in this study. Patients were followed up to 15 years by telephone inquiries, registry at cancer centers, and residents' registration offices. Out of the 517 patients, 105 were excluded because of missing data. The clinical characteristics of this dataset are given in Table 1. After the completion of this study we became aware of the fact that two patients (without adjuvant treatment) were present in both our screening and validation data set. To ensure that this caused no bias in our results, these two patients were excluded from all test sets in the validation analysis presented here.
Support vector machines are powerful supervised machine learning algorithms for classification problems [65]–[67]. We used a support vector machine to classify pancreatic tumors samples into poor or good prognosis groups based on the expression levels of selected genes. Here, we used the LIBSVM implementation as provided in the R package e1071 (version 1.5-18, obtained July 2008 from http://cran.r-project.org/web/packages/e1071/). The expression level of each gene was used as an independent feature to train the classifier. No kind of aggregation was used. All feature selection and machine learning steps were subjected to Monte Carlo cross-validation, which is a recommended and relatively un-biased evaluation strategy [22], [35] described in the following.
For ranking of genes, NetRank combines the correlation of a gene's expression level with the survival time of the patient with a network of known gene–gene relationships. The ranking can be computed iteratively. Here, we follow the notation and implementation in [37]:(1)Here denotes the ranking of page after iterations, is a symmetric adjacency matrix for the gene network, so if genes and are connected, and otherwise. is a vector of absolute Pearson correlation coefficients of gene expression values with the patient survival time, and is a fixed parameter describing the influence of the network on the rank of a page. Setting corresponds to no influence of the network and full influence of the gene expression data, whereas setting corresponds to full influence of the network and no influence of the gene expression data. The value appears to be used by Google [26]. The rank of a gene depends on the rank of all genes that link to it. Scaling by in the summation ensures that each gene has equal influence in the voting procedure. Each gene gets a rank of automatically and also gets times the votes given by other genes.
The iteration to convergence in (1) corresponds to solving the equation(2)where is the identity matrix, is the transpose of , and . With the choice of (no influence of the network, full influence of the gene expression data), equation (2) has the solution . That is, the rank of a gene solely depends on the correlation of its expression with survival time. For (full influence of the network, no influence of the gene expression data), equation (2) becomes(3)
To identify genes mentioned in the literature as prognostic immunohistochemistry markers for pancreatic cancer, we used GoGene [68] and combined the results of queries “pancrea* prognos* immunohisto* paraffin” and “pancrea* survival immunohisto* paraffin”. GoGene performs a PubMed query with the search term and then identifies gene names in the abstracts reported by PubMed. Table S2 shows the literature genes with the PubMed IDs of the abstracts in which they were found.
|
10.1371/journal.ppat.1004037 | A Role for LHC1 in Higher Order Structure and Complement Binding of the Cryptococcus neoformans Capsule | Polysaccharide capsules are important virulence factors for many microbial pathogens including the opportunistic fungus Cryptococcus neoformans. In the present study, we demonstrate an unusual role for a secreted lactonohydrolase of C. neoformans, LHC1 in capsular higher order structure. Analysis of extracted capsular polysaccharide from wild-type and lhc1Δ strains by dynamic and static light scattering suggested a role for the LHC1 locus in altering the capsular polysaccharide, both reducing dimensions and altering its branching, density and solvation. These changes in the capsular structure resulted in LHC1-dependent alterations of antibody binding patterns, reductions in human and mouse complement binding and phagocytosis by the macrophage-like cell line J774, as well as increased virulence in mice. These findings identify a unique molecular mechanism for tertiary structural changes in a microbial capsule, facilitating immune evasion and virulence of a fungal pathogen.
| Polysaccharide capsules are important virulence factors in pathogenic microbes that provide a protective coat against host immunity. Cryptococcus neoformans is a pathogenic encapsulated yeast that is a major opportunistic infection, causing approximately 600,000 cases of meningitis per year in AIDS patients globally, and whose polysaccharide capsule is a major virulence factor. While extensive work has detailed the chemical components forming the cryptococcal capsule, the molecular events leading to the higher order assembly of the capsule, and its consequences on immune subterfuge remain unknown. In the present studies we used a proteomics method to identify a novel hydrolytic enzyme, lactonohydrolase (Lhc1) and used a variety of biophysical methods including dynamic and static light scattering as well as motility studies to show that extracted capsular polysaccharide undergoes remodeling in a LHC1-dependent fashion. This results in a more tightly compacted capsular structure that alters binding of anti-capsular antibodies and reduces binding by both human as well as mouse serum complement. Furthermore, LHC1-dependent capsular alterations serve to increase the virulence of the fungus in a mouse model, suggesting a novel role for this class of enzyme in capsular remodeling and immune evasion in microbial pathogenesis.
| Polysaccharide capsules (PC) are highly diverse hydrated structures that provide microbes with a key defense against the host immune system [1]. For example, bacterial capsules confer resistance to complement-mediated opsonophagocytosis [2] and are an important property of highly virulent bacteria such as Neisseria meningitidis [3]. Among fungal pathogens, a prominent virulence factor of the opportunistic pathogen Cryptococcus neoformans is a large polysaccharide capsule with potent anti-phagocytic properties [4]. C. neoformans is a common cause of meningitis in parts of Africa [5], accounting for approximately 600,000 deaths annually [6].
The cryptococcal capsule is a hydrated polysaccharide gel, constituted by high-molecular weight polysaccharide polymers such as glucuronoxylomannan (GXM) which represents almost 90% of the total capsule with the remainder being glucuronoxylomannanogalactan (GXMGal) [7]. GXM is composed of a large backbone of 6-O-acetylated α-1,3-mannose residues with β-D-xylopyranosyl, β-D-glucuronosyl monosubstituted side chains [8]. Extensive work by numerous investigators has provided key insights into synthesis and virulence role of the capsular primary structure [7]. However, genes controlling or regulating higher order structures of the capsular polysaccharide have not been identified. This has been in part due to difficulties in assessing the tertiary structure of the cryptococcal polysaccharide.
Thus, to identify genes that may control the higher order organization of the capsular structure, we used a focused proteomic approach to identify capsular-associated proteins that may participate in remodeling of the cryptococcal capsule. Since current models suggest that the primary structure is synthesized within the cell cytoplasm [9], the hypothesis was that secreted proteins might be more likely to be involved in capsular tertiary structure. This approach identified a capsular lactonohydrolase of C. neoformans and a targeted mutant strain demonstrated a larger capsule size that was more permeable to dextran particles in a mutant strain defective in this hydrolytic activity. Recently applied biophysical methods [10] were then used to demonstrate that the mutant polysaccharide (PS) was larger, more hydrated and branched, evidenced by altered capsule nuclear magnetic spectra, zeta potential and polysaccharide hydrodynamic dimensions. The mutant also displayed an increase in antibody and serum-dependent phagocytosis by the macrophage cell line J774.16 cells, an increase in serum complement binding and reduced virulence in mice that could be reversed by depletion of complement using cobra-venom. These data thus identify LHC1 as a unique example of a gene locus involved in modification of higher order capsular structure of a microbial pathogen and its role in immune evasion.
After extensive washing of cells, dimethyl sulfoxide (DMSO) was used to solubilize and remove the outer layers of the cryptococcal capsule without breakage of the cell wall as described previously [11]. Strain B-3501 was used because its smaller capsule produced relatively less capsular polysaccharide that could complicate protein purification. Interestingly, after recovery of crude protein from dialyzed DMSO-solubilized material by adsorption on diethylaminoethanol-agarose, only two prominent bands were identified on Coomassie-blue stained PAGE gels (Fig. 1A). Protein sequencing identified three cryptococcal proteins (see supplemental Table S1 in Text S1), each matching protein sequence within the serotype D (www.ncbi.nih.gov) as well as the H99 serotype A database (www.broad.mit.edu), indicating their presence in two strains representative of two important serotypes capable of causing human disease. The small number of protein bands was remarkable, considering the large number of secreted proteins of C. neoformans [12] and may be due to the presence of only a small population of capsular-associated proteins or to incomplete adsorption of proteins from solubilized capsule material by the DEAE-agarose matrix. Analysis of the CNAG_04753 amino acid sequence from the higher mobility band showed strong homology to a number of fungal lactonohydrolases including that from Fusarium oxysporum (E = e-119; Fig. 1B) and contained three conserved domains for this class of hydrolytic enzymes [13]. Interestingly, using the PROCARB carbohydrate binding prediction tool based on a database of known and modeled carbohydrate-binding protein structures [14], three putative amino acids were identified that could represent amino acids involved in such binding,W28, N454, and R456—all aromatic amino acids that have the capacity to form Pi(π) bond complexes with hexose sugars, a common mechanism of lectin binding to carbohydrates [15]. Sequence analysis of the lower mobility band (68 kDa) identified a mixture of a conserved hypothetical protein and a protein showing closest homology to Kex1 of yeast. Since these latter two proteins were less likely to be involved in capsular modifications, they were not analyzed further.
A deletion strain was created in serotype A strain H99 to help identify a role for the putative lactonohydrolase from C. neoformans, Lhc1 using a strain of the serotype that is most predominant in human infections, serotype A [16]. As shown in Fig. 1C, a large increase in the size of the capsule was observed in the lhc1Δ mutant strain by India Ink microscopy grown in the presence of CO2, which was restored to approximately that of wild-type (wt) after complementation by a 3.6-kb fragment of the LHC1 gene. Larger capsule was also evident in YPD after a 1 day incubation that showed poor capsule induction in the wt strain or after capsule induction in ASN minimal media, 1∶10 Sabouraud or RPMI media (Fig. S1 in Text S1). In contrast, deletion of LHC1 had only a minor effect on other virulence factors such as laccase, measured by melanin formation (Fig. 1D) and no effect on urease activity or growth in YPD at 37°C (data not shown). Analysis of capsular radius of lhcΔ mutant cells using India ink microscopy induced by growth in the presence of 5% CO2 (Fig. 1E; p<0.01), ASN minimal media, 1∶10 SAB or RPMI demonstrated a significantly increased capsular radius compared to either the wt or complemented strains (Fig. S1 in Text S1; p<0.05). Interestingly, large capsules were also expressed by the lhc1Δ mutant in mouse brains (Fig. 1F, right panel) compared to that of wt (left panel) or the complemented strain (data not shown). These data establish a role for LHC1 in the wt capsular phenotype both in vitro and in vivo.
To confirm the identity of LHC1, we assayed for hydrolysis of the aliphatic lactone, D-pantolactone (Fig. 2A) using a previously-described high performance liquid chromatographic method [17] after growth of wt and lhc1Δ mutant cells in minimal media for 3 days 30°C. As shown in Fig. 2B, wt fungal cells converted approximately 11.3+/−2.6% (SEM, N = 3) of the D-pantolactone to the corresponding acid in 30 min, whereas no significant hydrolysis was evident in the lhc1Δ mutant. A small shoulder on the substrate peak of the mutant reaction could represent an unknown breakdown product. However, we were not able to detect hydrolysis of the aromatic substrate 3,4-dihydrocoumarin by changes in UV absorption using a previously described method [18], suggesting a restriction to aliphatic lactones that might be expected within the polysaccharide matrix (data not shown). Recombinant Lhc1 was inactive in both assays which may be due to cryptococcal specific conformational modifications or a requirement for a specific carbohydrate binding cofactor for activity.
Further studies sought to confirm a capsular localization of Lhc1 suggested by DMSO-solubilization from intact cells. Western blots using mouse antiserum developed against a recombinant maltose-binding protein (MBP)-tagged Lhc1 fusion protein demonstrated an immunoreactive band from wt or LHC1 complemented, but not lhc1Δ mutant strains of the appropriate molecular mass from SDS extracts of pelleted fractions enriched in cell wall/capsule but not cell lysis supernatants enriched in cytosolic proteins after homogenization (Fig. 2C). Lhc1-immunoreactivity was also not detected from culture supernatants or 20× concentrated culture supernatants (data not shown), further suggesting that Lhc1 was a capsular-associated protein. Control antibody raised against MBP alone showed no cross-reactivity against C. neoformans cellular materials by western blot as described (Fig. S5 in Text S1) [19].
In addition, Lhc1 was expressed under its native promoter as a C. neoformans codon-optimized mCherry fusion protein (Fig. 2D), which suggested that 1) Lhc1 expression is repressed under nutrient-rich conditions where capsule is repressed and 2) is successfully expressed and localized to capsule under conditions where capsule is induced, including 1∶10 SAB (Fig. 2D), ASN minimal media or RPMI (Fig. S2 in Text S1). Regulation appeared to be at the transcriptional level as quantitative RT-PCR studies demonstrated induction under capsule inducing conditions in either ASN minimal media, 1∶10 SAB or RPMI media after 24 h incubation, which was present for both the serotype A H99 strain as well as the serotype D strain (B-3501) used to identify the Lhc1 protein (Fig. S3 in Text S1). Additional studies utilized fluorescence immune-microscopy of sectioned C. neoformans cells to demonstrate capsular reactivity in a human autopsy specimen of a 30 year old female who died of overwhelming C. neoformans meningoencephalitis (Fig. 2E). These data suggest a role for Lhc1 expression in human infections. In summary, lactonohydrolase expressed from C. neoformans was localized to the capsular matrix of the fungus, although removal with detergent and DMSO as well as migration within an SDS-PAGE gel matrix suggested a non-covalent interaction.
Because of the larger capsule size of the lhc1Δ mutant evident on India ink staining, additional studies were conducted to assess alterations in the chemical structure of the capsule. Neutral sugar analysis did not identify large changes in substituent sugars although some increases in xylose and glucuronic acid as well as reductions in glucose were noted (Table S2 in Text S1). NMR spectroscopy of isolated soluble GXM from wt and the lhc1Δ strain also showed subtle differences in the GXM spectra that suggested alterations in higher order structure of the polysaccharide (Fig. 3A). The acetylation ratio was almost identical to the wt; however, mannosyl residues substituted with glucuronic acid residues were more frequently also 6-O-acetylated in the mutant (M6G-ac in Fig. 3A) in contrast to the deacetylated residues in the wt (M6G-deac). The ratio of mannose: xylose: glucuronic acid residues was 3∶2.1∶1, similar to wt. However, the exact chemistry of the cross-linkage of GXM chains was difficult to determine for the lhc1Δ strain. Samples are usually prepared for NMR spectroscopy by sonication to break-up large polysaccharides into smaller repeating units to reduce relaxation times and hereby allow for 2D NMR correlation spectroscopy [20]. However, conventional preparation in this case did not result in sufficiently small polysaccharide fragments to yield usable 2D NMR data for the determination of the smallest repeating carbohydrate structure in the lhc1Δ strain. High-power sonication resulted in complete break-up into monosaccharide units, suggesting that the larger polysaccharide fragments of the mutant exhibited higher levels of structural complexity in the form of branching or intermolecular cross-links that required higher sonication energies.
Because NMR spectroscopy suggested alterations in the higher order structure of the capsule, permeability was assessed using a fluorescent-labeled dextran dye (2,000 kDa) used previously in this organism [21]. India ink was added to the dextran suspension to allow demarcation of the exterior surface of the capsule. Full capsule thickness was determined as the distance between the fungal cell wall and the outer edge of the India ink exclusion zone (Fig. 3C). Comparison of the three strains demonstrated increased permeability of the lhc1Δ mutant versus that of either the wt or the complemented strain, as determined by the width of the zones of red fluorescence (wt: 1.38±0.07; lhc1Δ: 2.41±0.06; lhc1Δ+LHC1: 1.53±0.12; p<0.0001 for lhc1Δ versus either wt or complemented strain-Fig. 3B, C). Ratios of dextran penetration versus full capsule thickness were also calculated, which showed increased fractional penetration of dextran in the lhc1Δ versus wt or complemented strains. (wt: 0.61±0.03; lhc1Δ: 0.72±0.03; lhc1Δ+LHC1: 0.46±0.03; p<0.05 for lhc1Δ versus either wt or complemented strain; Fig. S4 in Text S1). These data show that LHC1 plays a role in reducing capsule permeability of C. neoformans to large molecules.
Since altered cross-linking or branching of the capsular polymer may affect the dimensions of the C. neoformans capsular PS, hydrodynamic sizes of DMSO-extracted capsular PS were determined by dynamic light scattering as described [22]. These data (Fig. 3D) demonstrated two sets of particle distributions as described previously for the DMSO-extracted polysaccharide [23]. Interestingly, the particle distributions from the lhc1Δ mutant strain were much larger and more heterogeneous than those from the wt strain, suggesting a role for LHC1 in reducing capsular PS dimensions.
Zeta potential (ζ) of polysaccharide samples were also determined for the capsular material from the mutant strain. Zeta potential is a measurement of charge and is defined as the electric potential gradient between a boundary liquid in contact with a solid and the mobile diffuse layer in the body of the liquid. Colloidal suspensions having ζ that deviate from zero (>×30 mVolts) have greater solvent hydration and tend to remain in stable suspension, whereas those with values closer to zero tend to aggregate [24]. Using this approach ζ was found to be −28.28±0.30 mV for the wt strain and −34.46±0.61 mV for the lhc1Δ mutant (Fig. 3E; p<0.001). This suggests that the larger particles of the lhc1Δ mutant seen in the polydispersity profile (Fig. 3E) were also more highly hydrated either by increased cross-links/branching and/or by differences in glucuronic acid availability, the latter suggested by the neutral sugar analysis.
We next utilized light scattering to assess higher order capsular structure. These biophysical methods recently demonstrated evidence for branching/cross-linking within the polysaccharide matrix that is difficult to assess by chemical methods alone [10]. For these studies, DMSO-solubilized capsule was analyzed without size fractionation to reduce bias that could be introduced by excluding important capsular constituents. As shown in Table 1, deletion of LHC1 was associated with changes in a number of macromolecular parameters including average molecular mass (Mw), radius of gyration (Rg), hydrodynamic radius (Rh), mass density and the 2nd virial coefficient (A2). Interestingly, while shape factor and A2 were restored by complementation, the other parameters were not, suggesting a sensitivity to gene dosing of some of these parameters relative to wt cells by the heterologous insertion of the LHC1 gene as previously described [25] and examined more recently [26]. The significant increase in Mw and Rh are consistent with the capsular dimension results (Fig. 3E), demonstrating that the increase in capsule size observed in the mutants is due to the presence of larger polysaccharide molecules. Interestingly, the ratio of Rg/Rh, referred to as the shape factor, ρ, was much lower in the mutant than wt and the complemented strain. This was a key parameter demonstrating higher structural complexity of the cryptococcal polysaccharide with low values suggesting higher levels of branching [10]. In addition, the A2 coefficient was altered in the mutant. The A2 coefficient is a property which describes the interaction strength between the molecule and a solvent, giving insights into the tendency of polysaccharide-polysaccharide interactions in that solvent [27]. Solubilized material from the mutant strain manifested a negative A2 value (−2.6±1.1×10−4 cm3 mol/g2) compared to a positive value of both the wt (1.82±0.6×10−4 cm3 mol/g2) and the complemented strain (1.74±0.5×10−4 cm3 mol/g2), suggesting that the strength of molecular-solvent interactions in the mutant strain is lower than the molecular-molecular interactions, relative to the wt strain, and that intra-branch interactions in the mutant are stronger than in the wt strain. Thus, the lhc1Δ strain appeared from the biophysical data to produce a population of higher molecular weight capsular PS that showed more solvent hydration and exhibited a greater degree of branching/cross-linkage. To obtain additional structural data, wt and lhc1Δ cells were induced for capsule on 1∶10 SAB media and subjected to cryo EM. As shown in Fig. 3F, representative micrographs demonstrated a condensed PS structure in the wt with reduced radius and a larger, more highly branched PS structure in the mutant strain, consistent with the biophysical data. This again suggests a model whereby the capsular adherent lactonohydrolase either directly or indirectly results in the remodeling of secreted polysaccharide particles to reduce particle size and branching/cross-linkages.
To determine the functional significance of the LHC1-dependent altered higher order structure in the capsule, we compared binding of mAb to the capsule. Antibodies showed a punctate pattern of binding with subtle but significant differences in the number of puncta observed between the LHC1 strains (Fig. 4A, B). Antibody negative (Fig. 4A) or isotype controls (data not shown) showed no significant binding. Differences in puncta observed after antibody deposition on C. neoformans capsule have been associated with differences in antibody-mediated protection but the mechanism has not been elucidated [28]. Opsonization with the mAb 18B7 yielded increased phagocytosis of the lhc1Δ strain by the J774.16 macrophage-like cell line with an almost doubling of the phagocytic index defined as number of fungal cells/macrophage (wt: 1.4±0.3; lhc1Δ: 3.85±0.02; lhc1Δ+LHC1: 0.8±0.3; p<0.02– lhc1Δ versus wt or complemented strain; Fig. 4C, left panel), with little phagocytosis evident, using a mouse IgG1 isotype control (Fig. 4C, right panel). Successful phagocytosis by macrophages plays a major role in killing of this facultative intracellular pathogen [29]. Opsonization was unsuccessful with mAbs 12A1 and 2D10 as IgM antibodies are not opsonizing, although these latter two IgM antibodies are capable of opsonizing serotype D C. neoformans strains through facilitation of a unusual conformational change in the capsule [30]. Increased antibody opsonization was also associated with decreased survival of the mutant strain after prolonged incubation in J774.16 cells (Fig. 4D). In summary, differences in capsular PS higher order structure as determined by biophysical methods translated into demonstrable changes in antibody-mediated rates of phagocytosis.
Similar to that found after antibody opsonization, the lhc1Δ mutant strain opsonized with human serum was more readily ingested than the wt strain (Fig. 5A-left panel), an effect that was abolished after heat inactivation (Fig. 5A-right panel), implying complement-dependent opsonization. The capsule serves as a site for deposition of C3 fragments of the alternative pathway of the complement cascade, which promotes C. neoformans phagocytosis [31], while the polysaccharide blocks activation of the classical pathway that can occur at the cell wall of avirulent non-encapsulated strains [32]. Incubation of fungal cells with human PBMC's after opsonization with fresh human serum resulted in increased fungal killing (Fig. 5B). Quantitation by flow cytometry demonstrated that C3 deposition was increased in the mutant strain, even after the addition of EGTA to inhibit the classical pathway, but was abolished (data not shown) after heat treatment (Fig. 5C). Complementation with the LHC1 locus led to a partial yet significant reduction in complement binding. Fluorescence microscopy using antibody to human C3 further demonstrated C3 binding within the enlarged capsule of the lhc1Δ strain (Fig. 5D) that was reduced after LHC1 complementation. Complement binding was heterogeneously deposited on the cells, most likely due to the presence of focal initiation sites, as previously described [32], [33]. Most of the C3 was identified in the form of iC3b (Fig. S6 in Text S1), as previously described [34]. To evaluate the consequences of the reduced LHC1-dependent complement binding in C. neoformans, we modeled the studies in mice, a species that would also allow testing for virulence [35]. As shown in Fig. 5E, opsonization with mouse serum reproduced the results using human serum and flow cytometry using antibody to C3 again demonstrated increased C3 binding. Inoculation of mice using an intravenous model showed reduced virulence of the lhc1Δ strain that was restored after complementation with the wt gene (Fig. 5G-left panel). Interestingly, after depletion of complement using cobra venom factor [33], the differences in virulence between wt and mutant strains disappeared (Fig. 5G-right panel), with a small increase in overall virulence of the wt strain, as previously described after complement depletion [36]. These data suggest a role for LHC1 in reducing mouse as well as human complement binding to C. neoformans and support a role for complement in mediating LHC1-dependent mammalian virulence.
A number of formidable pathogenic microbes express PS capsules that are potent virulence factors. Despite their importance in pathogenesis, many aspects of capsular architecture remain poorly understood. The cryptococcal capsule is particularly large and complex, resulting in cells with diameters up to 50 µm in diameter that cannot be ingested by phagocytic cells [37]. The primary structure of the cryptococcal capsule has been well characterized [7]. Xylose and glucuronic acid substituents are attached to a mannose backbone that are readily detectable by NMR spectroscopy. However, given that GXM polymers have masses in excess of 1 mDa, current methods of analytical chemistry cannot reliably detect the rare sugar modifications that may be responsible for tertiary structural complexity [38]. Thus, to identify proteins possibly involved in the modification of the capsular structure, DMSO extraction of intact cells was followed by extensive dialysis and capture on a charged agarose matrix which identified a lactonohydrolase that co-localized with the capsule. Further studies utilizing an mCherry-tagged recombinant protein, biochemical as well immunolocalization confirmed this capsular localization. While the putative Lhc1 sequence does not contain a putative N-terminal leader sequence, several unconventional protein secretion mechanisms have been described including secretion of protein-filled exosomes [39]. Lactonohydrolases are hydrolytic enzymes that cleaves the lactone group within carbohydrates to produce the corresponding organic acid and are expressed by a wide variety of plants and plant-associated fungi but have not been implicated in capsule modifications [40].
Deletion of the LHC1 locus resulted in a larger capsule that blocked penetration by larger particles comprising India Ink, but allowed increased diffusion of a fluorescent dextran polymer. This phenotype was associated with a reduction in virulence in a mouse model and the increased capsule size persisted during infection in brain. Classical analytical approaches [8] yielded only subtle differences between the wt and mutant strain, including a slight increase in backbone sugars such as mannose and xylose [38] as well as branching glucuronic acids by neutral sugar analysis. NMR spectroscopy also suggested only subtle differences that reflected retention of much of the mannose backbone and primary branching, with a similar mannosyl∶xylosyl∶glucuronic acid ratio, but potentially stronger cross-linkage between GXM chains, given the resistance of the mutant PS to sonication. These results differed somewhat from the elemental analysis and may be due to the preparation of the material for NMR spectroscopy which used soluble GXM, rather than DMSO-extractable polysaccharide material used for the elemental analysis.
Such subtle changes in primary structure as well as increased diffusion of dextran dye in the mutant suggested a role for LHC1 in capsular tertiary structure that could result in a larger, more open structure in the mutant strain. Biophysical methods were thus used to study DMSO-extractable PS of the cryptococcal capsule that comprise the outer interface with phagocytes during infection [21] and also contained the adherent Lhc1 protein. Notable was the larger molecular mass and increased particle size of the PS from the lhc1Δ mutant by polydispersity measurements, which correlated with the overall larger capsule size visualized by India ink microscopy. Previous work had noted an association between extracted particle size and overall capsule size between cryptococcal strains [23]. In addition, the larger negative zeta potential of PS from the mutant strain suggested a more solvated surface for these larger molecules and a tendency towards less aggregation that could provide a more open structure for antibody and complement deposition. The mutant strain PS was also found to have a lower shape factor ρthan wt, determined by the ratio of Rg, the radius of gyration and Rh, the hydrodynamic radius, which provides an important measure of whether a molecule is loosely linear or branched [41], [42]. C. neoformans wt strains tend to produce PS which have low ρ, more similar to branched polysaccharides such as glycogen and amylopectin [10]. This suggests that the surface structure of the larger lhc1Δ mutant PS contains an even more highly-branched/cross-linked surface component that is also more efficiently solvated. Small changes in the number of glucuronic acid residues in the mutant strain, detected in the composition assay, could have contributed to this increased surface hydration, either alone or in combination with a more highly branched surface structure. A lower ρ value could also reflect aggregation of components of the PS capsule [43], but would be inconsistent with the more highly negative zeta potential value that suggests a lower tendency towards aggregation. In addition, examining the two values contributing to ρ, most of the change was due to that of the radius of hydration, which is a measure of the hydrated zone surrounding a given polysaccharide particle, with very little change in the radius of gyration, again suggesting a more branched, hydrated structure.
Combining the results of this study with prior contributions from several laboratories and more recent studies of the capsular architecture suggest a tentative model that may help to illustrate the modifications of C. neoformans capsular PS by Lhc1 and exclusion of anti-microbial products (Fig. 6). The location of a hydrolytic lactonohydrolase within the PS capsular structure and the smaller size of the PS particles in the wt strain suggest that the enzyme either directly or indirectly plays a role in remodeling secreted PS fibrils (Fig. 6A). Recent data has suggested that the capsule is assembled by non-covalent binding of PS fibrils that have a branched structure [10]. Hydrolysis of outer branching units within PS surface structure by Lhc1 in wt cells (Fig. 6B, right panel) could reduce the size of the capsular PS (compared to that of the lhc1Δ mutant; Fig. 6B, middle panel), and its overall branching if the outer hydrated segments were also highly branched. This would result in a reduced hydrated surface with a smaller radius of hydration and a less negative zeta potential in the wt cells, resulting in a smaller, more compact capsule. This structure was supported by cryo-scanning electron microscopy which demonstrated a compact structure in the wt cells composed of truncated fibrils, whereas the mutant strain demonstrated a much more open lattice composed of larger fibrils, likely representing the larger PS demonstrated by the polydispersity measurements. The poor fitting and better solvation of the unprocessed mutant PS within the capsular gel lattice in the lhc1Δ mutant thus resulted in increased diffusion of dextran particles and greater penetration by anti-microbial products or exposure of binding epitopes. PS processing by Lhc1 could be due to the presence of trace lactone units forming some of the PS crosslinks or other carbohydrate linkages specifically tailored to the hydrolytic activity of Lhc1. However, identifying trace lactone groups within a large polysaccharide background is particularly difficult. To put this problem in perspective, the molecular mass of GXM is >1 mDa [44]; thus, detection of rare lactone groups is probably beyond the current analytical horizon. If true, the occurrence of lactone-dependent crosslinking would imply that this step of capsular assembly occurs in the extracellular space and that random concentration gradients of reagents involved in such processes within the capsule could contribute to the remarkable antigenic variability reported within the capsule structure [45]. Localization of the protein to the capsule, suggested by the method of isolation, western blots of cell fractions and immune- and fluorescence-microscopy, also supports a remodeling role in capsular synthesis. Indeed, the mCherry-Lhc1 localization studies appear to suggest production of the enzyme during capsule induction at the outer region of the capsule which is the region that is thought to represent an inducible outer layer that forms on top of an older inner capsule layer [46]. Lhc1 protein production during capsular induction also suggests such an extracellular mechanism. However, an indirect or regulatory role cannot be ruled out. Interestingly, the gene has previously been found to be induced in response to hypoxia, which may facilitate virulence in the relatively hypoxic environment of infected tissue [47].
Using either mAb 18B7 or serum as opsonin, the lhc1Δ mutant exhibited higher rates of phagocytosis than the wt or complemented strain. Patterns of antibody binding by the IgM mAb 12A1 and 2D10 were altered in the lhc1Δ mutant, whereas only subtle differences were noted using the IgG mAb, 18B7, which may have been due to the smaller size of the IgG antibody or to differences in epitopes to which they bind. As suggested by the serum-dependent phagocytosis data, more complement was deposited on the capsule of the lhc1Δ mutant than the wt strain both in the presence or absence of EGTA, which blocks the classical pathway. The increased deposition of C3 on the lhc1Δ capsule was accompanied by reduced virulence in a mouse model which was reversed by depletion of complement by cobra venom factor in the mice, suggesting a role for complement in the differential virulence between mutant and wt strains in complement-sufficient mice. There is strong evidence supporting a role for the alternative pathway in innate protection against Cryptococcus, since animals deficient in C3 but not C4 have reduced survival [48], [49]. Deposition of complement is important for efficient opsonization of the fungus, rather than direct killing by production of a membrane attack complex [50]. Complement fragments iC3b were also increased on the lhc1Δ mutant, which are important opsonic ligands and forms rapidly on the cryptococcal capsule after complement deposition [34]. Reducing alternative pathway activation by LHC1 thus facilitates a role for C. neoformans as an effective pathogen. In summary, the current studies demonstrate a role for LHC1 in the remodeling of the polysaccharide capsule of C. neoformans that represents a unique mechanism of virulence optimization among pathogenic microbes.
Research involving human participants was approved by the NIAID intramural institutional review board and written, informed consent was obtained from all study participants before participation and was conducted according to the principals in the Declaration of Helsinki. All experimental procedures involving animals were conducted under guidelines of the National Institutes of Health and protocols approved by the Institutional Animal Care Committees (IACUC) of the Intramural NIH/NIAID and the University of Illinois at Chicago.
Cryptococcus neoformans ATCC 208821 (H99) was a generous gift of J. Perfect. Strain H99 ura5, [51] was employed as a recipient strain for deletion and expression studies and were maintained on media described in Supplemental Material and Methods. Plasmid pCIP containing the URA5 gene was a kind gift of K.J. Kwon-Chung.
C. neoformans strain B-3501 was grown at 30°C in RPMI supplemented with 2% glucose to stationary phase. Fifty grams of cells were harvested, washed extensively in 10 mM sodium phosphate, pH 7.0 and exchanged into dimethylsulfoxide as described [11] and incubated overnight with shaking at 37°C. Supernatant was harvested and dialyzed extensively against 10 mM sodium phosphate buffer, pH 7.0, followed by recovery of protein by passage of the dialysate on a 1 ml (DEAE)-Sepharose column. The column was extensively washed in 10 mM sodium phosphate, pH 7.0 and eluted with the same buffer containing 0.5 M NaCl. Eluate was again dialyzed and subjected to PAGE. Proteins were then transferred to nitrocellulose membranes and subjected to automated protein sequencing after protease Lys-C digestion as described [52].
Standard methods were used for disruption and complementation of the LHC1 gene in strain H99 as described previously using two PCR-amplified fragments and a 1.3-kb PCR fragment of the URA5 gene previously described to effect a 1.4-kb deletion within the LHC1 coding region (see Supplemental Materials and Methods in Text S1) and was complemented using a 3.6-kb genomic fragment of the LHC1 gene.
Fungal cells or recombinant enzyme were assayed for hydrolysis of the aliphatic lactone D-pantonylactone using a previously-described method [17] after induction in minimal media (0.1% glucose, 1 g/L asparagine, 20 mM sodium phosphate, 1 g/L YNB without amino acids and ammonium sulfate) for 3 days 30°C (see Supplemental Materials and Methods in Text S1) and assayed for hydrolysis of 100 mM D-pantonylactone by high performance liquid chromatography by reference to standard D-pantoic acid.
Full-length and an N-terminal fragment of lactonohydrolase was expressed in E. coli as a fusion protein with maltose-binding protein by using the pIH902 expression system (New England Biolabs, Beverly, Mass.) and the recombinant maltose-binding protein–lactonohydrolase fusion protein (MBP-Lhc1) purified on amylose-Sepharose according to the manufacturer's directions as described in Supplemental Materials and Methods. Mice were immunized by a standardized protocol with either MBP-Lhc1 or MBP alone as control and MBP antibodies removed as described in Supplemental Materials and Methods.
Histopathological material was prepared, embedded and fixed and incubated with either anti-Lhc1 or anti-MBP antibody and observed using fluorescence microscopy as described in Supplemental Materials and Methods.
To induce capsule, yeast cells were grown in 3 mL of RPMI in a 12-well plate incubated in a CO2 enriched environment (GasPak EZ CO2, Becton Dickinson) in a 37°C water jacketed incubator for 4 days. Alternatively, capsule was induced by growth on 1∶10 dilutions of Sabaraud's media at 30°C or RPMI agar for the indicated times and Lhc1 transcript was measured by quantitative RT-PCR and is described in Supplemental Material and Methods. TMR-Dextran 2,000 kDa (TMRD, Invitrogen) staining of C. neoformans capsule on cells grown under capsule-inducing conditions was performed as described previously [21]. The distance between the outside of the cell wall and the staining front was measured using Slidebook software. One-way ANOVA was used to assess statistical significance among the strains; Tukey's t test was used to perform pairwise analyses post-hoc.
Glycosyl composition analysis was performed on DMSO-extracted capsule by combined gas chromatography/mass spectrometry (GC/MS) of the per-O-trimethylsilyl (TMS) derivatives of the monosaccharide methyl glycosides produced from the sample by acidic methanolysis as described [53].
Soluble glucoroxylomannan (GXM) was prepared and analyzed by NMR as previously reported [54]. Briefly, Isolated soluble GXMs were sonicated and lyophilized three times before being dissolved in 0.5 ml of 99.96% D2O (Sigma, St. Louis, Mo.) for nuclear magnetic resonance (NMR) analysis by standard methods [55]. NMR spectra were acquired with a Bruker Avance 600 NMR spectrometer, using a 5-mm (1H, 13C) inverse-detection dual-frequency probe, operating at 600.13 and 150.913 MHz, respectively as described.
Capsular PS samples from H99 wt, lhc1Δ and lhc1Δ LHC1 strains were isolated by DMSO extraction, prepared and subjected to light scattering analysis as described [10] and is described further in Supplemental Materials and Methods. Cryo-electron microscopy was performed on indicated cells and is described further in Supplemental Materials and Methods.
Capsule was measured by microscopy after the fungal cells were suspended in India ink [56], urease production by incubation on Christensen's agar [57], and laccase by melanin production on nor-epinephrine agar [58]. Virulence studies were conducted according to a previously described intravenous mouse meningoencephalitis model [59] using 10 CBA/J mice for each C. neoformans strain. In a second experiment, animals were treated with cobra venom factor (CVF) according to Shapiro et al. [36], described in Supplemental Materials and Methods.
Phagocytosis and fungal killing assays were conducted using J774.16 cells by the method of Shapiro et al [36] or the method of Miller and Mitchel [60] using human PBMCs and is described in Supplementary Materials and Methods. The phagocytosis index was determined by microscopic examination of the number of fungal cells ingested or adherent divided by the number of total macrophages.
Fungal binding of C3 from mouse and human serum was determined by flow cytometry using goat anti-mouse C3-FITC (ICN), anti-human C3-FITC (Invitrogen) or anti-human iC3b (Quidel) in the presence or absence of EGTA by the method of [33] using a Becton Dickenson LSR Fortessa flow cytometer. C3 binding was determined by fluorescence microscopy using a FITC-labeled anti human C3 (Invitrogen) and visualized by fluorescence microscopy.
The capsule radius was measured in India ink experiments using 10 cells for each strain and means compared using ANOVA with Tukey's test post hoc. Errors were expressed as standard error of the mean (SEM). Statistical significance of mouse survival times was assessed by Kruskall-Wallis analysis (ANOVA on Ranks). Statistical analyses for capsular biophysical measurements were carried out using Bi-ZPMwA Zimm Plot Software (Brookhaven Instruments).
90Plus/BI-MAS Software was used for effective diameter and polydispersity parameters (Brookhaven Instruments). Comparison of phagocytic index was performed by a non-parametric t-test with Welch's correction. Plots, curve fits, Pearson or Spearman correlations (r), and statistical analysis were performed using GraphPad Prism version 5.0a, GraphPad Software, San Diego, California, USA.
|
10.1371/journal.pgen.0030067 | Rapid Birth–Death Evolution Specific to Xenobiotic Cytochrome P450 Genes in Vertebrates | Genes vary greatly in their long-term phylogenetic stability and there exists no general explanation for these differences. The cytochrome P450 (CYP450) gene superfamily is well suited to investigating this problem because it is large and well studied, and it includes both stable and unstable genes. CYP450 genes encode oxidase enzymes that function in metabolism of endogenous small molecules and in detoxification of xenobiotic compounds. Both types of enzymes have been intensively studied. My analysis of ten nearly complete vertebrate genomes indicates that each genome contains 50–80 CYP450 genes, which are about evenly divided between phylogenetically stable and unstable genes. The stable genes are characterized by few or no gene duplications or losses in species ranging from bony fish to mammals, whereas unstable genes are characterized by frequent gene duplications and losses (birth–death evolution) even among closely related species. All of the CYP450 genes that encode enzymes with known endogenous substrates are phylogenetically stable. In contrast, most of the unstable genes encode enzymes that function as xenobiotic detoxifiers. Nearly all unstable CYP450 genes in the mouse and human genomes reside in a few dense gene clusters, forming unstable gene islands that arose by recurrent local gene duplication. Evidence for positive selection in amino acid sequence is restricted to these unstable CYP450 genes, and sites of selection are associated with substrate-binding regions in the protein structure. These results can be explained by a general model in which phylogenetically stable genes have core functions in development and physiology, whereas unstable genes have accessory functions associated with unstable environmental interactions such as toxin and pathogen exposure. Unstable gene islands in vertebrates share some functional properties with bacterial genomic islands, though they arise by local gene duplication rather than horizontal gene transfer.
| Genes vary greatly in their long-term phylogenetic stability, and there exists no general explanation for these differences. Stable genes persist as a single copy over a wide range of distantly related species, whereas unstable genes undergo frequent duplication and loss in a process called birth-death evolution. The vertebrate cytochrome P450 (CYP450) gene superfamily includes many genes that are present in a single copy in species ranging from teleost fish to mammals and other groups of genes that undergo active birth-death evolution across the same species. The author found that nearly all stable CYP450 genes encode enzymes known to function in the synthesis and degradation of steroid and retinoid hormones (and related molecules). These hormones function in core developmental pathways in vertebrates. In contrast, most unstable CYP450 genes encode enzymes that detoxify foreign small molecules (called xenobiotics—foreign biochemicals). In addition, many of the unstable CYP450 genes are subject to natural selection to change their amino acid sequence over time (positive selection), probably in response to changes in xenobiotic exposure. These findings suggest that stable and unstable genes differ in their rates of birth-death evolution, because stable genes have core endogenous functions whereas unstable genes respond to changing environmental conditions.
| Genes in animal and plant gene families vary greatly in their phylogenetic stability. Some genes persist as a single copy over a wide phylogenetic range of species, with few or no gene duplications or losses on different evolving lineages. Other genes undergo frequent duplication and loss in a process called birth-death evolution [1,2]. Gene duplication and subsequent divergence of the duplicate copies underlie the formation of gene families and are thought to be an important source of genetic complexity and evolutionary change [3–5]. Though many specific examples of stable and unstable genes have been documented, there exists no general explanation for their different patterns of evolution. To explore the evolutionary basis for these patterns, I sought a gene family with characteristics that make it amenable to detailed analysis and interpretation. The vertebrate cytochrome P450 (CYP450) gene superfamily is well suited for the following reasons. First, it is a large family with both phylogenetically stable members and unstable members. Second, the biochemical and organismal functions of most genes in the family have been extensively studied in humans and rodents. Third, their protein products are large and have a well-conserved domain content, providing abundant information for studies of molecular evolution. Finally, strong interest in the family has resulted in relatively high quality annotations of gene structure in a wide range of vertebrate species.
The human genome contains approximately 60 CYP450 genes, which encode membrane-bound oxidase enzymes that act on a wide variety of substrates (recent reviews include [6–8]). The CYP450 enzymes can be divided into those that act on endogenous small molecules and those that act on xenobiotic compounds. The endogenous-substrate enzymes function in biosynthesis or catabolism of steroids, sterols, retinoids, prostaglandins, and fatty acids. The xenobiotic-substrate enzymes are expressed primarily in the liver and epithelial tissues and function to defend against environmental toxins and carcinogens. These xenobiotic CYP450 genes have been studied extensively in humans because they comprise one of the major (and most polymorphic) activities that determine drug half-life and are thus important in pharmaceutical development [6,7]. Xenobiotic CYP450 enzymes have also been studied extensively because in some cases they convert exogenous compounds into active carcinogens, presumably as a negative side-effect of broad substrate specificity [9]. Approximately 22 of the human CYP450 enzymes are thought to act primarily on endogenous substrates and approximately 15 other CYP450 enzymes are thought to act primarily on xenobiotic substrates (Figure S1). A few of the remaining enzymes are reported to have activity toward both endogenous and xenobiotic substrates; it is unclear whether one or the other activity (or both) is their primary function. Finally, there are about ten CYP450 enzymes with no known activities, most of which have been identified only recently from human genome sequence.
To study the molecular evolution of the CYP450 superfamily, I considered for analysis about 20 vertebrate genomes in various stages of draft sequencing, assembly, and annotation. Of these, sequence coverage, assembly, and annotation of 11 genomes were sufficiently complete for my needs. Of these 11, Pan troglodytes was discarded because of its extreme similarity to human, leaving ten nearly complete genomes: six placental mammals, one bird, one amphibian, and two teleost fish. Analysis of gene stability among these genomes reveals that the rate of birth-death evolution among genes acting on xenobiotic substrates is much higher than among those acting on endogenous substrates.
As starting material, I gathered complete sets of predicted CYP450 genes from the genomes of human (Homo sapiens), rhesus macaque (Macaca mulatta), dog (Canis familiaris), cow (Bos taurus), mouse (Mus musculus), rat (Rattus norvegicus), chicken (Gallus gallus), clawed frog (Xenopus tropicalis), zebrafish (Danio rerio), and pufferfish (Takifugu rubripes). All known and predicted splice forms were screened to identify a single canonical coding sequence from each gene for molecular evolutionary analysis (see Materials and Methods). Figure S1 shows a maximum-likelihood protein tree derived from all 628 of these sequences, marked to indicate known or probable functions based on the human proteins. A representative segment of the tree is shown in Figure 1. In the tree, CYP450 enzymes that act on endogenous substrates are generally represented by a single gene in each organism, and the phylogeny of that gene approximates the known species phylogeny. In some cases there are two related proteins in one or both fish species, presumably resulting from the known whole-genome duplication on the fish lineage [10,11]. In all but one case, each orthologous group of genes is strongly separated on the tree from all other CYP450 genes, indicating that they became functionally specialized well before the divergence of teleost fish from mammals (the single exception is described below). In sharp contrast, nearly all CYP450 enzymes that act on xenobiotic substrates are encoded by genes that have undergone active duplication and loss throughout vertebrate evolution, reflected in multiple species- and lineage-specific gene expansions. The tree segment in Figure 1 includes two endogenous-substrate enzymes and one class of xenobiotic-substrate enzymes; they are typical of the complete tree (Figure S1). Table 1 summarizes gene numbers and variance across species for three groups of CYP450 genes.
Among the endogenous-substrate ortholog groups, an expected protein from a specific species was occasionally absent, usually from cow, chicken, or frog. One example is found in Figure 1 (a CYP46A1 cholesterol 24-hydroxylase ortholog is missing from cow). Most of these apparent gene losses probably result from incomplete sequence assemblies or gene annotations. Such apparent gene losses were found most often in the cow, chicken, and frog genomes presumably as a result of relatively immature assemblies and annotations of these genomes. In several cases, translated BLAST nucleotide (TBLASTN) searches of the cognate genome clearly indicated the likely presence of part or all of an unpredicted gene ortholog (see Materials and Methods; Tables S1 and S2). There were also a few cases of apparent species-specific gene duplications outside of fish. For example, there were two predicted rat genes encoding CYP51A1-like proteins (Figure S1). Although these cases may include undiagnosed pseudogenes, it is reasonable to suppose that occasional gene duplications may be tolerated in these groups. Finally, there was one case in which enzymes that act on endogenous substrates (corticosteroid precursors) appear to have become specialized recently and specifically in terrestrial vertebrates (CYP11B1 steroid 11-β-hydroxylase and CYP11B2 aldosterone synthase), probably in connection with resistance to dehydration [12]. Nevertheless, the general picture among endogenous-substrate CYP450 genes is one of striking phylogenetic stability, with very low rates of gene duplication and loss across jawed vertebrates.
Among CYP450 enzymes with xenobiotic substrates the picture was dramatically different. Almost without exception, one or more species or lineages had multiple genes that grouped in the tree with one or no genes from other species. The sample shown in Figure 1 is typical of the complete tree. These differences in gene content are far too extensive to be accounted for by incomplete genome sequence and annotation. In particular, the human and mouse genomes are especially mature in sequence and annotation, and the pattern of gene duplication and loss is robustly apparent in comparing these two genomes (e.g., Figure 1; Table 2). Extensive bootstrap testing on several subtrees supported the occurrence of independent gene duplications and deletions on both the mouse and human lineages (Figure S2).
In principle, genetic exchange among closely related genes within a species (concerted evolution) could explain some aspects of these patterns by causing lineage-specific clades of genes from sequence homogenization rather than recent duplication. In support of this possibility, a recent paper describes evidence that concerted evolution played a role in the divergence of the CYP1A1 and CYP1A2 genes in mammals from their bird orthologs [13]. However there are several reasons to think that this case is unusual. First, all of the syntenic CYP450 gene clusters between human and mouse contained different numbers of genes, indicating that gene duplication or loss must occur (two examples are shown in Figures S3 and S4). Second, except for a few of the most closely-related gene pairs, nucleotide divergence among duplicate genes was substantial, nucleotide changes were distributed widely across the coding sequence, and regions of higher nucleotide similarity usually coincided with regions of stronger purifying selection in the family in general (unpublished data). Finally, in many cases the tree relationships were inconsistent with concerted evolution being a dominant influence on sequence relatedness (for example, Figure S5). These results suggest that concerted evolution plays, at most, a minor role in shaping evolution in the family as a whole.
Most human and mouse genes for CYP450 enzymes are scattered widely across their genomes, but there are some exceptions. Clusters of genes are found at human chromosome bands 1p33, 7q22, 10p13, 10q26, and 19q13. In each of these five human clusters, the CYP450 genes are adjacent, with no other confirmed genes interspersed among them. Within each cluster, all the genes encode closely related proteins with many cases of apparent gene duplication since the split from the rodent lineage. Strikingly, nearly all of the CYP450 enzymes with known xenobiotic substrates were found in these gene clusters (e.g., see Figure 1 chromosome positions). The remaining genes in these clusters were also phylogenetically unstable but are not yet implicated in xenobiotic detoxification. Each of the human clusters had a syntenic counterpart cluster in mouse (Figures S1, S3, and S4). Very few of the genes in these clusters could be assigned as one-to-one orthologs because of repeated duplication and deletion events on both lineages. These syntenic gene clusters must have originated from a shared ancestral gene or genes, with repeated local gene duplications and losses giving rise to lineage-specific groups of related genes. The patterns of divergence are consistent with single gene duplications that occurred at various times, though multigene duplications followed by loss of most of the duplicate genes cannot be ruled out. Genomic clustering of related genes was also apparent in the other eight genomes, but assembly quality was more problematic, and I did not investigate the patterns in detail. These results are consistent with increased rates of birth-death evolution specific to xenobiotic-substrate genes, with local gene duplications giving rise to clusters of related genes.
If exposure to xenobiotic compounds changed over evolutionary time, the CYP450 detoxifier genes might be subject to positive selection driven by changing substrates. To test this possibility, I used maximum-likelihood codon analysis on paralogous groups of genes from many parts of the overall CYP450 tree [14,15]. Most attention was given to tree regions where there was evidence of frequent recent gene duplication, both because this might indicate selection to diversify enzyme function and because this provided abundant sequence data for analysis. Specifically, 203 genes in clades with five or more closely-related genes were tested, including six groups from primates, nine from rodents, and six from a mixture of primates and rodents. Details of the method are given in Materials and Methods, and summary statistics for the results are given in Table S3. A total of five of the 21 tested sets showed highly significant evidence of positive selection. The specific amino acid sites under positive selection were analyzed for a set of 12 rodent Cyp2d genes, which showed the strongest signal of positive selection. In this set, a Bayes-Empirical-Bayes method identified nine sites with strong evidence of positive selection (p > 0.9 for each site). The structural positions of these nine sites were compared by alignment with two closely related CYP2 proteins for which crystal structures have been determined [16,17]. Both in the primary sequence alignment and in the 3-D protein structure, the nine sites under positive selection were clearly associated with substrate-binding or substrate-channel regions (Figures S6 and S7). In the protein structure, the distance from positive selection sites to a bound warfarin ligand was significantly smaller than for random sets of sites (p ≈ 0.0002, see Materials and Methods). In addition to these nine sites, all three small indels among the 12 rodent genes coincided with regions of substrate binding (Figure S6). No method exists to analyze such indels for positive selection, but it is plausible that they affect substrate interactions and might be subject to such selection. These results are consistent with changing xenobiotic substrates driving the observed positive selection.
For comparison with the unstable genes, nine sets of genes from stable CYP450 groups were also tested for positive selection (marked on Figure S1). In each case, the orthologous genes from the six mammals were tested as a group. Because there are almost no gene duplications in these groups, it was not possible to include paralogs in these sets, but the degree of divergence of the genes (total tree length) was in the same range as for the paralog groups, indicating that the maximum-likelihood analysis should have similar power [18]. No evidence of positive selection was found for any of the nine sets, as expected based on their function as enzymes acting on endogenous substrates.
Phylogenetic instability among genes in families has been observed repeatedly, including among immunoglobulin genes [19], major histocompatibility complex genes [20], T-cell receptor genes [20,21], mammalian olfactory receptor genes [22], nematode chemosensory genes [15,23], globin genes [24,25], cichlid opsin genes [26,27], zinc-finger genes [28], F-box and MATH-BTB genes [29], and others. Selective pressure from environmental change or pathogens has been suggested as a driver of duplication and diversification for many of these examples, but alternatives are difficult to rule out. The vertebrate CYP450 gene superfamily affords special insight into gene instability because probable biochemical functions of most of the genes are known and because it includes both stable and unstable genes. I found a strong correspondence between genes implicated in xenobiotic detoxification and evolutionary instability in the form of gene duplication and deletion. These findings strongly support the involvement of changing external selective pressure in high rates of birth-death gene evolution. Positive selection for amino acid change has been also been reported in many of the phylogenetically unstable gene families, suggesting that diversification is achieved by a combination of gene duplication and selection-driven divergence in sequence.
For other unstable gene families, lack of stable family members for comparison or lack of extensive functional information prevent the highly informed analysis possible for the CYP450 genes. Nevertheless, a similar explanation for birth-death evolution is plausible for most or all such families. Immunoglobulin genes, major histocompatibility complex class I and class II genes, and T cell receptor genes function in pathogen defense, and an extensive literature documents gene duplications and deletions in these families during vertebrate evolution. Mammalian olfactory receptor genes and nematode chemosensory genes function in response to environmental chemical stimuli and might be subject to changes in desired ligand specificity or sensitivity. Lake Malawi cichlid fish have undergone duplications of opsin genes that have diverged to different spectral sensitivities, and expression of specific sets of these genes results in diverse adaptive visual properties, possibly driven by mate choice and foraging preferences [26,27]. In other cases the role of external selective forces in driving birth-death evolution is less obvious but plausible; for example the F-box and MATH-BTB genes in nematodes and plants have been hypothesized to function in pathogen defense, though the evidence is indirect [29].
Much less is known about specific functions of CYP450 genes in nematodes and plants. Based on maximum-likelihood protein trees with three Caenorhabditis species (briggsae, elegans, and remanei) and with the plants Arabidopsis thaliana and Oryza sativa (rice), it is clear that birth-death evolution affects most members of the CYP450 family in these groups as well. For example, only 7% of Arabidopsis CYP450 genes and 38% of C. elegans genes had strict one-to-one orthologs within their groups (unpublished data). As in the human and mouse genomes, closely related unstable genes tended to be in genomic clusters. The exact degrees of stability and instability cannot be directly compared because the phylogenetic distances differ, and genome annotation qualities may vary.
A few CYP450 genes in humans lack clear functional information. The relative evolutionary stabilities of these genes may be predictive of general aspects of their functions. Among stable genes of unknown function, human CYP20A1 and orthologs are well separated from all other CYP450 genes on the protein tree and are encoded by a single gene in each of the ten species analyzed (top left of Figure S1). This pattern suggests an ancient and critical function that is shared among all the species. A less clear-cut but similar pattern is seen for an unnamed human CYP27 (Chromosome 2) and the CYP2U1 and CYP2R1 genes. I speculate that these four enzymes will prove to be involved in biosynthesis or catabolism of endogenous substrates of importance throughout vertebrates. Consistent with this interpretation, CYP20A1 expression is enriched in immune system cell types, CYP2U1 is expressed broadly, and CYP2R1 expression is enriched in testis and immune system [30]. Unstable genes for which I could find no described function include CYP4A43, CYP4Z1, CYP4Z2 (possible pseudogene), CYP4X1, CYP4A11, CYP4A22, CYP4F2, CYP4F3, CYP4F8, CYP4F11, and CYP4F12. Among these, in addition to birth-death evolution, the CYP4F group is subject to significant positive selection, consistent with substrate-driven selection for amino acid change. I speculate that each of these genes functions in xenobiotic detoxification or some other process that acts on environmental substrates. Consistent with this interpretation, expression of most of these genes is enriched in liver (CYP4A11, CYP4A22, CYP4F2, CYP4F3, CYP4F11, and CYP4F12) or in tracheal tissue (CYP4X1) [30], which are tissues where other xenobiotic CYP450 genes are known to function [6,7]. Finally, the evolution of the CYP2W1 and CYP2J2 genes is less clear, with duplications in some species but only a single gene in humans. CYP2W1 has a broad expression pattern [30], and evolution of the group appears most consistent with a relatively stable function in mammals but extensive duplication and diversification in amphibians and teleost fish. CYP2J2 expression is enriched in liver [30], and the group includes multiple genes in most species but only a single gene in humans and macaques.
Groups of closely related unstable CYP450 genes are strongly clustered in the human and mouse genomes. A similar pattern of genome clustering is found in a number of other human and mouse gene families [31–33]. These patterns indicate that DNA duplications that persist in the genome occur predominantly locally. In contrast, about half of very recent large DNA duplications (segmental duplications) on the human lineage are interchromosomal, and some of the remainder occur nonlocally on the same chromosome [34–37]. It is possible that these segmental duplications represent a gene duplication mechanism that has dramatically and recently increased on the primate lineage [34]. Alternatively, nonlocal segmental duplications may be more unstable than local duplications and fail to persist over longer evolutionary periods, so that they rarely give rise to persistent functional duplicate genes.
The combination of evolutionary instability and gene clustering has parallels with “genomic islands” found in bacteria. Genomic islands consist of blocks of DNA that are present in some bacterial isolates but absent from other closely related isolates (reviewed in [38]). Though the full picture is still developing, it is likely that most genes in genomic islands have auxiliary organismal functions such as pathogenesis, antibiotic resistance, symbiosis, and specialized metabolism [38]. Despite some parallels, the mechanisms by which bacterial genomic islands and unstable gene islands in animals arise are different: clustering in bacteria is a consequence of horizontal transfer of contiguous blocks of DNA, whereas clustering in animals is restricted to closely-related genes and arises from local gene duplication events without horizontal gene transfer.
Sequence data were downloaded from ENSEMBL [39] release 40, which contained the following genome builds: H. sapiens NCBI36, M. mulatta MMUL1, B. taurus Btau2.0, C. familiaris BROADD1, M. musculus NCBI36, R. norvegicus RGSC3.4, G. gallus WASHUC1, X. tropicalis JGI4.1, D. rerio ZFISH6, and T. rubripes FUGU4.40. Data for P. troglodytes were not included because the genes are too similar to human to be informative. Data for Monodelphis domestica and Tetraodon nigroviridis and several low coverage genomes were considered but rejected because current gene prediction sets appeared to be too incomplete or inaccurate (unpublished data). Starting with a hand-collected set of human CYP450 sequences, all known isoforms of human CYP450 proteins were collected from a protein BLAST (BLASTP) search of all human proteins. Many genes had multiple transcript and protein forms. Each such case was analyzed to identify a single protein isoform that appeared full length and most family-typical, based on alignment with related CYP450 proteins. This analysis resulted in a set of 60 reference human proteins that were full length or nearly full length, including three possible pseudogenes. Proteins from other species were collected from BLASTP searches with these 60 human proteins as query. All isoforms were kept initially, and for each gene the isoform with the highest BLASTP score against a human protein was kept for further analysis (for the most part, this approach discarded obviously incomplete proteins). Subsequent tree analysis of these proteins indicated the surprising absence of a few CYP450 proteins with endogenous substrates. Some of these were found to be present (but unpredicted) in the cognate genomes by TBLASTN searches (Table S1). Other genes are probably present in unassembled sequence; the number of missing genes and genome sequence coverage in the mammalian genomes is summarized in Table S2.
After discarding a few proteins that were missing more than 30% of the expected CYP450 protein or clearly aligned badly, the remaining set of 628 proteins were aligned with ClustalW (default settings except gap clustering turned off) [40]. A complete list of these protein sequences is found in Dataset S1 with ENSEMBL gene identifiers [39]. After removal of aligned columns with more than 30% gap residues, the resulting multiple alignment had 417 sites, at least 70% of which were present in any given sequence. A maximum-likelihood tree was constructed from this alignment with PhyML (JTT matrix, four rate categories, gamma shape parameter 1.0) [41]. Tree nodes were rotated, colored, and displayed with Bonsai [42]. Additional annotations were added to the tree image for Figure S1. The complete tree was subjected to 200 bootstrap repeats with PhyML with the same settings, which required ten days of central processing unit time. Subtrees of specific interest were tested with 1,000 bootstrap repeats (Figure S2).
Multiple sets of five to 15 closely related primate and rodent genes were selected from the tree in Figure S1. For each set, a protein alignment was made using ClustalW (default settings except gap clustering turned off) [40]. This protein alignment was used to generate the corresponding codon alignment and to construct a maximum-likelihood protein tree with PhyML [41]. The tree and codon alignment were analyzed with CodeML from PAML 3.14 [14], using models 7 and 8, with three starting dN/dS (ω) values for model 8 to avoid local optima. The neutral model 7 assumes a β-distribution of 10 dN/dS ratio classes constrained to lie between 0 and 1.0, whereas the selection model 8 permits one additional dN/dS ratio class without constraint. In order to minimize the effects of gene prediction and alignment errors, aligned columns with a gap in any sequence were excluded from analysis (“cleandata” option in CodeML). The transition/transversion ratio (κ) was estimated by CodeML. Statistical significance was assessed using a χ-square test on twice the difference in log-likelihood values (ΔML) for models 7 and 8 with two degrees of freedom, a statistic shown to be conservative in simulations [43]. Assignment of specific sites under likely positive selection was based on the Bayes-Empirical-Bayes test implemented in CodeML. In Figures S6 and S7, marked sites are subject to positive selection with p > 0.9. Extensive gene conversion can exaggerate signals of positive selection, but direct analysis of gene conversions among the tested genes suggests that they are unlikely to confound the analysis. GENECONV [44] was used to test for possible gene conversion events in the complete genomic regions of all the mouse and human genes analyzed for positive selection (unpublished data). In all cases except one, detected regions of conversion affected only a few genes in <10% of their coding region, suggesting that they will not significantly affect the signal of positive selection. The exception was among mouse genes in the CYP2D family, but even in this case a minority of genes had evidence of gene conversion and none were affected in more than 20% of their coding region.
The significance of the association of positive selection sites to bound warfarin ligand in the CYP2C9 structure was assessed as follows. The mean physical distance of the alpha backbone carbon atom to a reference warfarin atom was measured for the nine amino acid residues under positive selection. Warfarin atom C10 (carbon 10, atom 7439 of Protein Data Base structure 10G5) was used as the reference; it was selected as centrally located in the binding pocket/access channel based on visual inspection of the three-dimensional protein structure. The expected mean distance for randomly located amino acid residues was determined by repeatedly choosing nine random amino acid residues and computing the mean distance from their alpha backbone carbon atoms to the reference warfarin atom. Mean distances for 10,000 simulation repeats fitted a normal distribution (Figure S8), and all but two had mean distances larger than the mean distance for positively selected sites.
The Protein Data Base (http://www.pdb.org) number for the structure discussed in this paper is 1OG5. |
10.1371/journal.pcbi.1004423 | The Beta Cell in Its Cluster: Stochastic Graphs of Beta Cell Connectivity in the Islets of Langerhans | Pancreatic islets of Langerhans consist of endocrine cells, primarily α, β and δ cells, which secrete glucagon, insulin, and somatostatin, respectively, to regulate plasma glucose. β cells form irregular locally connected clusters within islets that act in concert to secrete insulin upon glucose stimulation. Due to the central functional significance of this local connectivity in the placement of β cells in an islet, it is important to characterize it quantitatively. However, quantification of the seemingly stochastic cytoarchitecture of β cells in an islet requires mathematical methods that can capture topological connectivity in the entire β-cell population in an islet. Graph theory provides such a framework. Using large-scale imaging data for thousands of islets containing hundreds of thousands of cells in human organ donor pancreata, we show that quantitative graph characteristics differ between control and type 2 diabetic islets. Further insight into the processes that shape and maintain this architecture is obtained by formulating a stochastic theory of β-cell rearrangement in whole islets, just as the normal equilibrium distribution of the Ornstein-Uhlenbeck process can be viewed as the result of the interplay between a random walk and a linear restoring force. Requiring that rearrangements maintain the observed quantitative topological graph characteristics strongly constrained possible processes. Our results suggest that β-cell rearrangement is dependent on its connectivity in order to maintain an optimal cluster size in both normal and T2D islets.
| High or low blood glucose levels are detrimental to human health. The hormone-secreting cells primarily responsible for maintaining glucose at physiologically appropriate levels are embedded in small clusters within the pancreas, the so-called islets of Langerhans. These islets have an irregular arrangement of cells, β cells that secrete insulin, α cells that secrete glucagon, and other cells with less well-understood functions. While the arrangement of β cells is irregular, these cells need to be touching for the islet to respond to glucose with insulin secretion. We first use a mathematical formalism called graph theory to show that cell arrangements in islets from diabetic and control donors are significantly different. The question we then address is: Is there some set of moves of islet cells that will preserve the observed arrangement? The aim is to gain insight into the biological processes by which islets are formed and maintained. We find moves on β-cell graphs that leave the same significant aspects of cell arrangements unchanged. These moves turn out to be severely restricted, and suggest that β cells may prefer to move from larger clusters but can move to a cluster of any size, possibly to maximize their exposure to blood vessels.
| Pancreatic islets of Langerhans make up 2% of the average pancreatic mass (or 0.000028% of human body mass), yet contribute significantly to the regulation of blood glucose levels. These micro-organs consist primarily of α, β, and δ cells that produce the hormones glucagon, insulin, and somatostatin, respectively.
β-β cell contacts are necessary for proper islet function [1–3]. Their electrical coupling allows for the synchronization of intercellular [Ca2+] oscillations which results in pulsatile insulin release upon glucose stimulation [4–8] and increases insulin production two-fold as compared to isolated β cells [9] (which show partial recovery in insulin production after reaggregation [10]). This coupling is dependent on Connexin36 (Cx36) gap junction channels [11, 12] since Cx36-deficient mice show altered insulin pulse dynamics and glucose intolerance [13]. Prediabetic mice display impaired Cx36 coupling [14] suggesting a possible role in the progression to T2D. In humans, β cells contain Cx36 gap junctions and levels of Cx36 mRNA correlate with insulin expression [15]. However, Cx36 knockdown reduces incretin-stimulated, but not glucose-stimulated, insulin secretion [16] suggesting the importance of Cx36 may be not through glucose response but through the response to incretins which itself is disrupted by lipotoxicity. Interestingly, the upregulation of Cx36 occurs in unison with the main wave of β-cell differentiation [17], further illustrating the possible dependence of β-cell function on gap junction coupling. It has also been shown that Cx36 protects β cells from apoptosis under cell injury [18] in mice.
Human islets have a unique cytoarchitecture with direct consequences on islet function [19]. The islet’s cytoarchitecture creates the anatomical basis for functional coupling between β cells [20]. However, what is the correct architecture for optimal function of pancreatic islet cells? Qualitatively, this arrangement is non-random [21] and species-dependent [19, 22], with rodent islets displaying a β-cell core surrounded by an α-δ mantle and human islets varying in arrangement in a size-dependent manner [21, 23, 24]. Smaller human islets (effective diameter < 60 μm) consist of an inner core of β cells surrounded by an α- and δ-cell mantle (similar to rodents), whereas in larger islets, α and δ cells are also found within the core possibly due to fusion of subunits [25] or lobules [26] consisting of the mantle-core architecture arranged around penetrating blood vessels [26, 27].
Rare cellular replication and apoptotic events balance β-cell mass over a lifespan [28]. However, with sustained hyperglycemia, changes in β-cell mass have been observed. As metabolic load increases, a complementary increase in β-cell mass is detected in mice [29], rats [30], and humans [31–33]. In mice, β-cell mass increase is achieved through β-cell replication [34], yet in humans β-cell replication is not detectably increased [35, 36], leading some to suggest β cells form under a different mechanism [36]. Nestin-positive islet-derived progenitor cells, which can differentiate into insulin-producing cells, are found in rat and human islets [37]. However, Street et al. [38] found that nestin-positive cells in adult human islets were not colocalized with insulin positive cells suggesting nestin is not expressed in the mature human β cell. With T2D, this initial increase is followed by β-cell mass loss [36] and functional failure believed to be from glucolipotoxicity [14]. Inflammation also affects β-cell mass whereby low concentrations of interleukin-1β (IL-1β) promote β-cell function and prolonged high glucose exposure increases IL-1β which, in turn, increases Fas-mediated apoptosis [39].
β-cell migration is also suggested to occur, most notably during islet development but has also been shown in the adult pancreas. Cole et al. [40] demonstrated that individual human β cells, previously suggested by [41], and large aggregates of cells can form within the ductal epithelium and migrate during gestation. They also found Vimentin (a mesenchymal protein) positive adult human β cells suggesting that adult β cells are capable of remodeling. Recently it was shown that loss of Fbw7 induces pancreatic ductal stem cells to proliferate into endocrine cells in the adult human pancreas [42], with evidence for subsequent migration into islets. Note that the coordinated activation of genes essential for endocrine cell proliferation, migration, vasculogenesis and hormone secretion has been demonstrated [43]. Thus, overall, the processes that govern human β-cell mass maintenance are not well understood. Supporting evidence for any specific process gleaned from rodent models must be considered in the context of possible structural differences between rodent and human islets.
The governing principle determining the optimal arrangement of cells in the healthy and T2D individual is a mystery. Further, there are observed architectural differences in T2D that may contribute to its progression [44], such as β-cell mass loss preferentially in large islets [21], hypertrophy in β cells [45], and amyloid plaque formation in islets [46, 47]. How these architectural changes affect the organization of β cells and their function remains unknown.
Given observations of the results of stochastic development and maintenance processes, there are two basic ways in which one can analyze these processes quantitatively. To set the stage for our islet β-cell analysis, let us consider a hypothetical dataset consisting of a sample of real numbers drawn from a one-dimensional normal distribution. The first standard approach is to make a histogram of the sample, and this histogram recapitulates an approximate normal distribution. A second approach is to consider the numbers as resulting from the equilibrium distribution of a stochastic Langevin process. In this approach, the same sample of normally-distributed random numbers leads us to the Ornstein-Uhlenbeck process with a random walk balancing a linear restoring force in the Langevin dynamics of each number, interpreted as the position of a particle on a line. These two approaches are equally valid views of the data, a static view and a stochastic view, and we shall apply them in turn to the study of β-cell arrangements in islets.
We propose that graph theory is an appropriate framework for quantifying β-cell arrangement because it captures quantitative aspects of connectivity flexibly. Graph theory has been used in many contexts in biology, in obvious contexts such as metabolic or regulatory networks, and in more esoteric contexts such as functional MRI analysis [48] and tongue carcinoma prognostication [49]. Geometric graphs have been used to study protein-protein interaction networks [50]. With respect to islets, functional graphs representing β-cell activity in individual mouse islets were created to analyze their small-world characteristics [10]. Images of islet sections show regularities in structure on casual inspection, but the significance of such regularities is difficult to ascertain without a mathematically sound framework for quantifying islet architecture. Here, we apply graph theory to the architecture of islets, and let the vertices of a graph consist of β cells in each islet while the graph's edges represent intercellular connectivity (Fig 1A–1C). This geometric framework captures key architectural characteristics quantitatively in terms of graph-theoretic constructs, as we shall show.
A precondition for applying either approach is the availability of a large-enough dataset. We use the dataset published in [21], consisting of ~9,000 control and ~8,000 T2D two-dimensional sections of islets containing ~200,000 control and ~110,000 T2D cells. All graphs that we consider refer to this two-dimensional dataset, and we are cognizant of the limitations of working with two-dimensional sections of actual three-dimensional structures, the islets. We do address the differences between two-dimensional and three-dimensional graphs of islet cell arrangements by explicitly computing graphs in a smaller islet dataset that does have three-dimensional coordinates for every cell, showing that quantitative measures of these graphs are qualitatively similar. We show first, in the static view, that graph theory quantitative measures, such as mean connectivity of β cells, size of β-cell clusters and number of clusters per islet, can distinguish between control and T2D islets. We then test, in the stochastic view, various models of β-cell degree- and β-cell cluster size-dependent stochastic models of hypothetical graph rearrangements to find possible topological constraints on β-cell rearrangement processes. It is these theoretical relocation processes that are the analogs of the linear restoring force and the random walk defining the normal distribution in the Ornstein-Uhlenbeck example above. To avoid confusion, we emphasize that in the absence of clear experimental evidence on the actual processes that maintain human β cells, our use of verbs such as `move’ is only a theoretical construction. The aim here is to use imaging data to pinpoint the precise mathematical equilibrium processes so that these mathematical constructs can be framed in terms of their contribution to the actual biological realizations, when the experimental data becomes unambiguous.
Twelve T2D and 14 non-diabetic (control) human organ donor specimens were collected and prepared as described in [21]. Briefly, a virtual slice was taken across each pancreatic specimen containing the head, body, and tail regions. This slice was then stained for insulin, glucagon, somatostatin, and DAPI. Tissue sections were imaged and two-dimensional (x,y) coordinates for each endocrine cell nucleus were recorded using a nucleus marker DAPI with identification of each cell type. The T2D group consisted of 4 males and 8 females ranging in age from 38–81. These subjects died of causes other than diabetes and it is assumed they were normoglycemic with ongoing treatment. The control group had 9 males and 5 females with ages ranging from 15–81. Further information about this dataset can be found in [21]. This dataset, after applying the described imaging method, produced location and cell-type information for ~200,000 cells found in ~9,000 control islets and ~110,000 cells found in ~8,000 T2D islets (Table 1) which was used for further analysis. The control and T2D datasets can be found in S1 Dataset.
Here our focus is on β-cell clusters so henceforth we consider only β cells as the vertices of the graph (Fig 1C). Edges between β cells, as determined by the nearest-neighbors of cells, describe the connectivity of the β-cell cluster.
A cell has a non-zero radius. If we fix a neighborhood size (discussed in the next subsection), for a given cell, not every cell with a center in a neighborhood around this cell is a nearest-neighbor of the cell as not all nearby cells share a contact. The assignment of graphs to the position dataset must take this into account, so that only edges representing nearest-neighbor cell-cell contacts are added to the graph. A typical case is shown in Fig 1D. To determine the neighbors of cell 1, we utilized a shadow algorithm, depicted in Fig 1E, as follows: Neighboring cells were first sorted by distance, then as the sorted list of cells was traversed, edges were added to the graph only if the cell was not in a previous cell’s ‘shadow’. The previous cell’s shadow angle, θ, was calculated by θ=tan−1(4d), where d is the distance to the previously added cell and the radius of a cell is 4 μm. In the example in Fig 1D, cell 8 is not a neighbor since the distance between it and cell 1 is greater than the neighborhood radius. After sorting the remaining cells by their distance to cell 1, it is determined that cells 3 and 5 are in the shadow of cells 2 and 4, respectively. Therefore the neighbors of cell 1 are cells 2, 4, 6, 7, and 9 (Fig 1E). Between 2 and 6 percent of total edges are removed from a β-cell graph of neighborhood radius of 10 microns by the shadow algorithm. This percentage increases with respect to an increase in neighborhood radius (Fig 1F). Graphs representing each control and T2D islet were computed. Differences between large and small islets, where small islets are defined by an effective diameter of < 60 microns, have been reported [23]. We will, therefore, analyze the large and small islet groups (defined by the 60 micron diameter) separately as well.
While graphs can be deduced with the shadow algorithm for any choice of neighborhood size, it is critical to use the data to determine a range of possible neighborhood sizes. To this end we calculated the pair distribution function, g(r)dr=IsletArea2πrN2∑i∑j≠iδ(r−rij) where N represents the number of vertices, rij is the distance between vi and vj, and δ(r−rij)={1 if r=rij0 otherwise. This function describes how the density of cells varies as a function of distance from a given reference cell.
We calculated the pair distribution function for each β-cell graph in the Control and T2D groups. The distributions for large and small islets were calculated separately, since architectural differences exist between them as previously shown in [21, 23, 24]. Bounding boxes of each islet were found and used for the islet area in the calculation. The pair distribution function was then averaged for each radius over all islets in the control and T2D datasets.
We also examined the pair distribution functions for other cell types found in islets, namely α-α, α-β, α-δ, β-δ, and δ-δ. For distributions of the same cell type, the formula above was used. For edges between vertices of mixed cell types (say, cell types a and b), a different version of g(r) was used to capture the distribution of b with respect to a, gab(r)dr=IsletArea2πrNaNb∑i=1Na∑j=1Nbδ(r−rij) where Na and Nb are the number of a and b vertices, respectively.
Human islets have been described with a variety of qualitative forms: α-β cell core subunits [25], lobules [26], cloverleaf patterns [24], ribbon-like structures [27], and folded trilaminar plates [23]. Our aim was to characterize the structure of islets in an alternate form, amenable to quantitative analysis in a static or a stochastic Langevin approach. The structure of graphs can be characterized quantitatively in a variety of ways. Here we define the most basic graph measures: the mean degree of vertices, the mean number of components per graph, and the mean number of vertices per component.
The measures we consider (degree, component size and cells per component) are amongst the simplest numbers associated with a graph. We found (see Results) that these were already biologically interesting as they could distinguish between T2D and normal islets. One measure we decided against using was the number of cliques, defined as a subgraph such that every vertex in the subgraph shares an edge with every other vertex in the subgraph. We found preserving the number of cliques to be too strong a constraint to maintain with random rearrangements.
The degree, dv, of vertex v is the number of edges containing v as a vertex (Fig 1G). For islet graphs, this measures the number of β cells in direct contact with a given β cell. To define dv, let E(a,b) = 1 if eab, the edge between vertices a and b, exists and E(a,b) = 0 otherwise. Then dva=∑i=1nE(a,i) where n is the total number of vertices in the graph. Thus, the mean degree can be defined as 1N∑j=1Ndvj=1N∑j=1N∑i=1nE(j,i) where N is the number of vertices in the dataset. For a given neighborhood radius, the degree of each cell of each islet was calculated and averaged together for the control and T2D groups.
The second characteristic, the connected components, describes discrete clusters of β cells. A component of a graph is a subset of vertices contained in the graph such that each vertex in the set shares an edge with at least one other vertex in the subset (Fig 1H). That is, if a closed curve can be drawn around a subset of a particular graph without crossing an edge, and this subset contains no other smaller subsets with this property, then the subset is a component. A component can also be singular, i.e. contain only one vertex, and thus describe lone β cells. In light of this, both the singular + nonsingular components and the nonsingular-only components, which represent β-cell clusters, were examined. For a given neighborhood radius, the components of each islet were found. The number of components per islet and number of cells per component for singular + nonsingular and nonsingular-only components were then averaged over the control and T2D groups for small, large, and all (small+large) islets.
While islet structure images obtained from selected two-dimensional (2d) sections of pancreata have been used to support qualitative conclusions regarding islet function for decades, we wondered if graph theoretical measures deduced from 2d sections might have artifacts that would impact the validity of our graph theory approach. A systematic study of three-dimensional (3d) volume estimates from 2d is given in [51]. In other words, could it be that actual 3d islet graphs leads to graph measures that are very different from the graph measures deduced from 2d sections, without invalidating all the conclusions in the literature that are based on 2d sections alone? There are not many published human 3d islet structures to test this, but to examine the effects of studying 3d islets in 2d slices, we took a different dataset consisting of 28 islets (17,539 α cells, 32,947 β cells] that were sliced every 15 microns resulting in (x,y,z) coordinates where x and y represent the placement of the cell in the slice and z represents the height of the slice in the islet. We then perturbed the z value of each cell by a random number between -7.5 and 7.5 such that the cells maintained a minimal distance of 4 microns. We randomized this dataset 10 times resulting in a three-dimensional dataset that was then used for a direct comparison between resulting graph measures in 3d and in 2d after simulated sectioning.
For the 3d case, graphs were created with specified edge type and neighborhood-sphere size. For 2d case, the dataset was sliced every 15 micron throughout the volume with respect to the z coordinate, i.e. the volume was sliced at z = 0, 15, 30, …. To examine the dependency of the placement of the slice, the starting point of the slicing was shifted by a certain number of microns i.e. the volume was sliced at z = 1, 16, 31, … for slice-start = 1, the volume was sliced at z = 2, 17, 32, … for slice-start = 2 and so on for slice-starts ranging between 0 and 14. For each slice, two-dimensional graphs were created where edges were created with specified edge type and neighborhood size (based solely on the (x,y) coordinates of the cells).
To simulate stochastic graph-altering processes akin to the random walk and linear restoring force that define the Ornstein-Uhlenbeck process, we simulated vertex deletion and addition moves in each islet. We emphasize that these processes are a mathematical representation of all the developmental and homeostatic processes that resulted in the observed islet graphs. The question we focused on is: What possible graph stochastic processes would preserve the observed simple quantitative graph measures of such a set of islet graphs? Thus the deletion and addition moves do not represent actual β-cell death and birth processes, which are well-known to be rare [28], but just as random steps and a linear restoring force with appropriate parameters will preserve the histogram of a sample of normal random numbers, we aimed to find graph dependencies for vertex addition and deletion moves that would preserve the observed distributions of graph measures.
We considered two different classes of models depending on the dependencies of the vertex addition and deletion processes. Each model was characterized by relative likelihood (RL) functions chosen such that cells whose graph characteristics are either above (RL+) or below (RL-) a given parameter value have a higher likelihood of an event such as the theoretical addition of a cell in its vicinity or the theoretical deletion of that cell. The sigmoidal functions are
RL+=0.5+0.5*tanh(x−rlp*)
(1)
and
RL−=0.5−0.5*tanh(x−rlp*)
(2)
where x is the graph measure defining that class of models and rlp* is the addition/deletion parameter value (Fig 2A and 2B). In general, stochastic simulations with hard transitions, as opposed to the sigmoidal functions defined here, will lead to artifacts. We tried to avoid such artifacts with the use of the RL+/- functions. In a deterministic process, such sharp transitions for where a cell can or cannot be placed would be required, as such sharp transitions would be part of the definition of the rules for such a process. There is no evidence for such determinism in any published experimental data.
We describe each class of model in turn. Each simulation has a given RL function for the modeled addition (RLa) step with a given relative likelihood parameter (rlpa) and modeled deletion (RLd) step with a given relative likelihood parameter (rlpd). We will introduce the shorthand, M for RL-, P for RL+, a given (RLa, RLd) combination as [RLa] [RLd], and a given simulation as [RLa][RLd][rlpa][rlpd] where [RLa] and [RLd] are either M or P and [rlpa] and [rlpd] represent the given theoretical addition and deletion parameters. For example, MP01 describes the simulation where M = RLa, P = RLd, rlpa = 0 and rlpd = 1.
β cells receive cues from their environment and neighboring cells [1–3, 6, 10]. We hypothesized that the theoretical addition or deletion of a cell could be influenced by the number of β-cell contacts. Here we used the degree-dependent RL functions as described in Eqs 1 and 2 where x represents the degree of the cell. Notice that for RL+, cells with larger degrees have a higher relative likelihood than cells with smaller degrees of a given event (addition or deletion), whereas for RL-, the opposite holds.
(Fig 2A and 2B, where x is deg and rlp* is rlpa or rlpd). For combinations PP, PM, MP, and MM, 500 simulations consisting of 500 iterations of cell deletion and addition per graph for each combination of rlpa = 0,…,7 (representing the observed degrees in the original architecture) crossed with rlpd = 0,…,7 for each dataset (control large, control small, T2D large, and T2D small islet graphs) were run. This totaled 512,000 simulations.
In vitro isolated β cells are known to spontaneously re-aggregate into cluster-like structures [52]. It has also been shown that β-cell clusters are necessary for pulsatile glucose stimulated insulin response in vivo [6–8],[12],[53]. Finally, the proximity of β cells to islet capillaries is essential for survival [54]. Greater β-cell distance from associated capillaries has been shown to result in β-cell disappearance during islet transplantation [55]. Here we test to see if theoretical cell addition and/or cell deletion are influenced by component size. The addition (rlpa) and deletion (rlpd) relative likelihood functions used were as described in Eqs 1 and 2 where x now represents the size of a cell’s component. Notice that for RL+, cells existing within smaller components have a higher relative likelihood than cells within larger components of a given event (addition or deletion), whereas for RL-, the opposite holds.
For each combination of PP, PM, MP, and MM, 500 simulations consisting of 500 iterations of cell deletion and addition per graph for each combination of rlpa = 1,…,5 (representing the overwhelming majority of component sizes in the original architecture) crossed with rlpd = 1,…,5 for each dataset (control large, control small, T2D large, and T2D small islet graphs) were run. This totaled 200,000 simulations.
We implemented methods for stochastic changes to graphs consistent with the dataset. For each iteration, one vertex was deleted from and added to each graph based on a given probability function for the theoretical deletion and addition processes. We emphasize again that these are theoretical abstractions of cell rearrangements, not to be conflated with β-cell birth and death. In particular, note that these theoretical processes are balanced in that they preserve cell number in each islet at each step in the simulation. Obviously no biologically valid birth and death simulation could coordinate the preservation of cell number. Such a rearrangement is not a trivial process because the islet graphs represent finite-sized cells with steric constraints, not idealized mathematical points. Particular care must be taken when a β-cell is moved from one component in an islet to another, because there are geometric constraints that need to be preserved for the cell-cell contact interpretation of edges in the resulting graph. The process of removing a vertex from a graph is straightforward. However, adding a vertex to a graph can cause geometric complications, especially since there is a minimal distance between two vertices that needs to be maintained. The addition process used is: (i) for the chosen vertex, vparent, a new vertex, vnew, is added at an angle chosen randomly from angles not occupied by a neighbor of vparent and a random distance between 8 and 13 microns; (ii) a ‘frozen’ list is created consisting of vparent and vnew; (iii) a ‘problem’ list is created consisting of vertices within dmin distance from vnew where dmin is the minimal distance observed in the islet; (iv) for the first vertex on the ‘problem’ list, v1, the closest vertex on the ‘frozen’ list, vc, is found and v1 is moved radially from vparent a given distance such that the new distance between v1 and vc is between dmin and 10 μm; (v) the vertices within dmin from v1 are added to the bottom of the ‘problem’ list; and (vi) v1 is removed from the ‘problem’ list and added to the ‘frozen’ list (vii) steps (iv) through (vi) are repeated with v1 representing the first vertex in the ‘problem’ list until it is empty. This algorithm intuitively envisions a propagating wave from vparent in the direction of vnew where vertices too close to vnew are moved outward, and vertices that are now too close to the just-moved vertices are moved outward, and so on until the exterior vertices of the graph are reached (Fig 2C–2E).
Graphs were created using the Boost Graph Library [56] which utilizes a depth-first search algorithm for computing components [57]. Additional code was written in C++. For simulations, this study utilized the high-performance computational capabilities of the Biowulf Linux cluster at the National Institutes of Health. The Mann-Whitney test, with a Bonferroni multiple comparison correction, was used for statistical tests.
To verify simulation convergence, the maximal standard deviation and maximal change in mean of each graph measure were calculated with the addition of each simulation. For simulation i, the standard deviations of simulations 1, …,i were calculated for each (rlpa,rlpd). The maximal standard deviation is the maximum of these values and was found and plotted for i = 1…n, where n is the number of simulations. The change in mean was found with the addition of each simulation for each (rlpa,rlpd). The maximal change in mean is the maximum over all (rlpa,rlpd). For convergence, as i approaches n the maximal standard deviation approaches a constant and the maximal change in mean approaches zero.
The number of islets obtained from each donor is given in Table 1, with the number of cells in large and small islets. We counted the number of endocrine cells of each type in large and small islets (Table 2). Notice the decrease in all cell types in large T2D islets as compared to control islets.
The pair distribution function showed a peak between 8 and 13 microns for all types of edges in large islets. Due to this, all measures were calculated for graphs with neighborhood radii between 8 and 13 (with radii 5–7 and 14–16 added for comparison). β-β edges are shown in Fig 3 with other edge types in S1–S3 Figs. Small islet pair distribution functions do not show the expected asymptotic approach to 1, suggesting that there are not enough small islets for averaging over islets to produce stable results.
We found that graph measures determined from 2d sections and 3d islet graphs are qualitatively the same for any choice of neighborhood size (Fig 4). Conclusions drawn from 2d sections should therefore be reflective of 3d differences. Of course, the significance of differences between T2D and normal sections has nothing to do with the 3d vs. 2d issue addressed here as any geometric difference would affect both types of sections alike.
Both large and small T2D islets have a greater mean degree than control for all neighborhood radii (Fig 5). The number of vertices per non-singular component is greater in the T2D large-islet group than control for graphs of neighborhood radii between 7 and 12 microns (Fig 6). The difference between T2D and control was not statistically significant in small islets or the combination of small and large islets across the range of neighborhood radii (Fig 6A and 6C).
The number of components per graph is significantly lower in large islet graphs (Fig 7B) and when all islets are combined (Fig 7A) with T2D than the control group. There is no statistically significant difference between the groups for small islets (Fig 7C). The size dependence and cumulative distributions of the three graph measures for all, large, and small islets, respectively, are shown in S4–S6 Figs. The number of components per islet is significantly different between control and T2D groups even when comparing by subject (S4 Table, all measures were compared by subject in S1–S15 Tables).
The differences in mean degree and number of components seen in T2D islets cannot be explained by the loss of β cells alone (S7 Fig). The differences observed in T2D cannot be explained by age (S8 Fig). We also compared the number of components per islet with the mean degree of the islet controlling for the number of cells per islet. There was a statistically significant difference (P< 0.0001) between the control and T2D group. Therefore, the decrease in components observed in the T2D group as compared to the control group is not solely from the increase in mean degree. There is a difference in the mean degree and number of cells per component between large and small islets for the control and T2D datasets (S16 Table). However, this difference is dependent on the neighborhood radius.
To see the effects of different parameter values, we ran the component-dependent MP model on an individual islet with the original graph measures given in Fig 8. The resulting architectures are a mix of the islet containing varying numbers of large clusters (Fig 8, rlpd = 1) and varying numbers of clusters consisting of only 2 or 3 cells (Fig 8, rlpd = 5). See S1–S4 Videos for simulations of the component-dependent model using (rlpa,rlpd) = (3,1) for the PP, PM, MP, and MM models, respectively.
Table 3 shows all the stochastic graph processes with parameter ranges in which they are consistent with the observed graph measures. Random theoretical addition and deletion processes do not preserve the graph measures found for the experimental data (S17 Table). There are two solutions possible for both degree-based and component-based models, for both normal and T2D islets. The solutions have different characteristics. The solutions that have precise parameter values are those in which cell deletion is more likely to occur from larger components (degrees) while cell addition is unconstrained by component size (degree). These solutions are termed MP solutions (Fig 9). The other solutions are PP solutions in which both cell deletion and cell addition have a range of correlated values (Fig 10). Notice that the cell deletion parameter decreases as the cell addition parameter decreases, an intuitively natural compensation. These PP models (degree- or component-based) are therefore, mathematically-speaking, over-parametrized, but that is not a biologically valid reason for rejecting them.
We computed the correlation between mean degree and mean component size for each islet. We found that these are correlated (Fig 11F). Thus the degree-based and component-based solutions we found are qualitatively the same process.
All results for the component-based and degree-based models for small and large islets can be found in S9–S16 Figs. The maximal mean difference and maximal standard deviation (calculated as described in Materials and Methods-Computation and Statistics subsection) for each model is given in S17–S24 Figs, respectively.
Not only is it important for the resulting simulations to exhibit the original architecture’s measures, but their edge distributions should be similar as well. For example, the pair distribution functions averaged over all degree-based simulations were calculated for measure-equilibrium models MP41 and PP42 and for non-measure-equilibrium model PM20 (Fig 11A–11C). The measure-equilibrium models show two joined peaks of varying heights between edge lengths 8 and 13, which correlates with the peak found in the pair distribution of control large islets. However, the PP model distribution has values greater than 1 for radii greater than 13, whereas the non-measure-equilibrium model distribution shows multiple peaks ranging from a distance of 4 to 20, and then decays after 25 microns. To quantify the differences found when comparing the pair distribution functions for MP and PP simulations with the experimental pair distribution functions, we subtracted the difference for each radius for several MP and PP measure-equilibrium solutions (averaged over 10 simulations each) (Fig 11D and 11E) and integrated the absolute value of the difference between radius 0 and radius 50 (Fig 11D and 11E, values in parentheses). There is a noticeable increase in the total difference for the PP models. There is also an increase in total difference for non-measure-equilibrium solutions, such as MPX2 and MPX3 (S25A and S25B Fig) where X = 0–7. This is most evident within the radial distance interval of 8 and 13 where the total difference for MPX2 averaged over X = 0–7 is 3.42 and for MPX3 is 3.96 as compared to the measure-equilibrium solution set, MPX1, which is 1.88. The model with the closest match to the pair distribution function found in the control and T2D (S25C Fig) data is MP01.
The distributions of clusters (components) and the distributions for mean degree and number of cells per component for MP01 (MP02 and PP01, two other measure-equilibrium solutions, are given for comparison) simulations are given in S26 Fig. The experimentally observed distribution (S5 Fig) is qualitatively similar to that of the equilibrium solution MP01.
Our main descriptive result is that T2D islets have (statistically significant) quantitative differences compared to control with respect to their connectivity structure as quantified with basic graph theory measures. These subjects were being treated for T2D at the time of death, and diabetes was not the cause of death in any case. It is likely that the usual natural history of T2D progression (initial β-cell mass increase [31, 32] followed by β-cell mass loss [36] and functional loss from glucolipotoxicity [58] and inflammation [39]) has resulted in the differences we quantified. However, unlike animal models, it is not possible to attribute these changes solely to prolonged hyperglycemia and dyslipidemia.
It is striking that large T2D islets exhibit higher mean degrees, larger components but fewer components compared to control. Did the process of developing diabetes lead to the loss of cells that were not sufficiently connected, and the loss of components that were not large enough for the synergistic functional couplings between β cells needed to sustain them? Indeed, prolonged hyperglycemia, with concomitant higher demands for insulin secretion, leads to a decrease in Gjd2 expression (which codes for Connexin36 gap junction channels) in mice [59]. Further, Cx36 plays a protective role against apoptosis in the presence of toxins [18] suggesting β cells with fewer connections have less apoptotic protection. A greater number of cells in contact could be a compensation for fewer gap junction channels between each pair of cells in contact, resulting in increased degree and cells per component.
Our central deductive result from the simulations of stochastic processes on islet architectures is that it is possible to find a stochastic process that maintains the graphical measures observed with a slight difference between parameters in T2D subjects compared to control. Surprisingly, the data could rule out entire classes of stochastic rearrangement processes as implausible because of a failure to reproduce the observed architectures of islets for control and T2D. Our results indicate that β cells are preferentially moved from larger β-cell clusters while their destination clusters are picked randomly.
The idea here was not to examine changes in topology due to change in β-cell mass, as that would require having information about the entire natural history of progression to T2D via insulin resistance. Such information would be very valuable but does not exist at this time. In addition, the specimens were obtained from T2D donors with unknown therapeutic interventions, so the T2D dataset reflects possibly stable but inadequate β-cell mass. Therefore, we took each dataset separately and found theoretical cellular rearrangement processes that conserved the topology of each dataset as determined by the three graph measures and the resulting pair distribution functions. The distribution of theoretical clusters was essentially unchanged under the stochastic rearrangement process.
One point of interest is that the structural line between T2D and normal islets appears to be fairly fine, i.e., the parameter difference between the stochastic process for T2D islets and that for control is small. Whether these differences are a reflection of the compensatory processes that took place during the stages of insulin resistance preceding the appearance of T2D, or are reflective of ongoing continuing pathological changes, cannot be ascertained from such donor data. Furthermore, the limited number of 3d islets that we had to test our qualitative similarity between 3d and 2d graph measure trends does not quantitatively guarantee that a future dataset with 3d structures analyzed with 3d graph measures will find the same significant difference between T2D and control 2d graph measures. In other words, we cannot preclude that 3d islet structures analyzed with 3d graph measures could conceivably find that T2D islets are not distinct from control.
Our use of stochastic methods is motivated by two factors: (1) The exact processes of development and homeostasis in the human pancreas are unknown. Imposing a deterministic process would lead to many graph structures that might be artifacts of the simulation, as such a process would have to depend on many more parameters to detail how cells are placed depending on cell vicinity at each step. We do not know of any extant dataset that could possibly be used to constrain such deterministic simulations, nor do we know any literature on experimental results about in vivo endocrine development in the pancreas that could inform the design of an algorithm for deterministically re-arranging cells. (2) When we calculate pair distribution functions from the experimental data, it is striking that there is no evidence of systematic higher order structures, as would be hard to avoid in a deterministic process. Thus, while we cannot conclude that the actual natural history of islet development in the pancreas is stochastic, there is no experimental evidence that the placement of cells within each islet is deterministic.
The invariance of graphical measures turned out to be a constraint strong enough to determine a unique set of stochastic graph rearrangement models. There was really only one true solution since degree- and component-models are related (Fig 11F) and MP01 fits the pair distribution function best. This is a prediction that may be testable with lineage tracing techniques in animal models. Obviously, our mathematical formulation of cell rearrangements is not based on experimental data, but the equilibrium model pair distribution still captures the observed features of the pair distribution function observed in the data, and away from equilibrium, does not. This gives us some confidence that the mathematical formulation may not be too far from reality as it has been shown that endocrine cell proliferation and migration processes are not disjoint [43]. Furthermore, when we repeated the analysis on 10% of the data, we found the same results as when we did the analysis on 100% of the data, with no change in parameters of the stochastic process. This can be interpreted as a 90% check on our initial 10% model.
Among the limitations of our analysis, the most important is the lack of information on vasculature. The importance of vasculature in islets as endocrine organs needs no repetition. While it would be very interesting to carry out a more elaborate version of our simple graph theoretic simulations with data including vascularization, this would require a very large amount of 3d data as 2d sections are unlikely to capture the essential connectivity of capillaries [60].
Another limitation is that the donor characteristics available were not very detailed. With enough donor data at all age groups, it would be of great interest to follow the changes in graph measures with age, and ideally, with insulin resistance. That level of data may make it possible to consider independent addition and deletion processes to take β-cell number changes into account.
While we only addressed the β-cell contacts in this work, the proximity of α cells to β cells may be important for the suppression of glucagon secretion by insulin [61]. It would be interesting to carry out our analysis with graphs including all different endocrine cells as vertices. More data would probably be required to constrain such a model as there are many more possible rearrangements that one would need to consider.
What might be the biological rationale for the MP01 model that is preferred by our theoretical study? As we have no data on enervation and vascularization relative to the cell coordinates in the image dataset, we do not have a concrete basis for suggesting that some unknown functionally optimal amount of contact with neurons or capillaries is the cause. The functional characteristics of the sub-unit structures that Bonner-Weir and co-workers have suggested in islets [24] may provide a clue as to why β-cell migration may proceed from larger clusters to smaller.
A very challenging extension of this work would be to type 1 diabetes, where the auto-immune destruction of islets is an ongoing process. This would require donor data with duration of disease besides more specific donor characteristics. The setup of stochastic simulations to take the duration of disease into account would require considerable thought, and there would likely be dependence on the age of the donor, besides the duration of disease.
|
10.1371/journal.pntd.0006550 | Transrenal DNA-based diagnosis of Strongyloides stercoralis (Grassi, 1879) infection: Bayesian latent class modeling of test accuracy | For epidemiological work with soil transmitted helminths the recommended diagnostic approaches are to examine fecal samples for microscopic evidence of the parasite. In addition to several logistical and processing issues, traditional diagnostic approaches have been shown to lack the sensitivity required to reliably identify patients harboring low-level infections such as those associated with effective mass drug intervention programs. In this context, there is a need to rethink the approaches used for helminth diagnostics. Serological methods are now in use, however these tests are indirect and depend on individual immune responses, exposure patterns and the nature of the antigen. However, it has been demonstrated that cell-free DNA from pathogens and cancers can be readily detected in patient’s urine which can be collected in the field, filtered in situ and processed later for analysis. In the work presented here, we employ three diagnostic procedures—stool examination, serology (NIE-ELISA) and PCR-based amplification of parasite transrenal DNA from urine–to determine their relative utility in the diagnosis of S. stercoralis infections from 359 field samples from an endemic area of Argentina. Bayesian Latent Class analysis was used to assess the relative performance of the three diagnostic procedures. The results underscore the low sensitivity of stool examination and support the idea that the use of serology combined with parasite transrenal DNA detection may be a useful strategy for sensitive and specific detection of low-level strongyloidiasis.
| As international bodies focus efforts on control of the world’s neglected tropical diseases, the critical importance of accurate and sensitive diagnosis becomes a key factor. The problem arises when the infection load in a community is reduced to a level where the standard diagnostic methodologies are insufficiently sensitive to detect the residual infection in the community. There is a need to develop improved diagnostic strategies for many parasitic diseases. One of the more difficult to diagnose helminth parasites is the nematode Strongyloides stercoralis. We have introduced a new approach that detects parasite-specific cell free DNA in urine as a sensitive measure of parasite presence. In the work presented here, we compare the performance of parasitological, serological and urine/DNA-based diagnosis of S. stercoralis infection. Using a Bayesian Latent Class Analysis approach, we provide evidence for the enhanced utility of using both urine and blood for the diagnosis of this parasite.
| The soil-transmitted parasitic nematode Strongyloides stercoralis is increasingly recognized as a significant human pathogen that deserves consideration for inclusion in the public health interventions that are underway to control other medically important soil transmitted helminths (STH) [1] such as Ascaris lumbricoides, Trichuris trichiura and the hookworms Ancylostoma duodenale and Necator americanus [2, 3]. The current STH control strategy does not include S. stercoralis as a target for chemotherapy. One of factors that has negatively influenced the inclusion of S. stercoralis as a target in the STH control efforts is the limited ability to diagnose an infection based on the standard, WHO-recommended, microscopic identification of larval parasites from stool samples [4]. While highly specific when carried out by experienced technical personnel, the sensitivity of this approach is compromised by the unpredictable, intermittent release of small numbers of larvae by adult parasites residing in the intestine [5]. Because this parasite is difficult to diagnose, the prevalence of S. stercoralis infection in many regions is largely unknown. There is a clear need for an improved approach for the diagnosis of S. stercoralis infection to define prevalence and the impact of intervention measures in the field.
In recognition of this need for better diagnostics, serological methods have been devised [6]. While significant advances have been made in terms of sensitivity, detection of specific antibodies is still subject to individual response as well as the antigens used in the tests to measure anti-S. stercoralis antibodies [7]. For increased specificity, nucleic acid-based diagnosis of S. stercoralis from stool samples using qPCR has been introduced. Although specific and amenable to multiplexing for the parallel detection of other pathogens [8, 9], this process has limitations, again, due to the intermittent presence of small numbers of S. stercoralis larvae passed in the feces of most patients. Additionally, collection of stool specimens in the field is labor intensive, costly and cumbersome.
The use of cell-free DNA in blood and other bodily fluids as biomarkers has gained wide acceptance in clinical laboratories. Cell-free DNA is being applied as a diagnostic marker for cancer, prenatal diagnosis and in infectious diseases, including parasitic diseases such as malaria, trypanosomiasis, leishmaniasis, schistosomiasis, strongylodiasis, and filariasis [10–12]. While most methods use blood, cell-free DNA is also readily detected in urine [12–14], saliva [15], stool [16], and sputum [17]. Cell-free DNA that is initially released into the blood can pass through the glomerular barrier and appear as transrenal DNAs in the urine [13] as small fragments of ~150–300 bp [18]. The advantages of transrenal DNA-based diagnosis of infectious disease include: (a) urine collection is non-invasive, (b) urine is easy and cheap to collect and process, and, in theory, (c) transrenal DNA does not depend on the stage of the parasite or the tissue site of infection.
In the current study, we employed a Bayesian Latent Class modeling approach to examine the diagnostic utility of three methodologically distinct diagnostic procedures—traditional comprehensive stool based parasitology, serology that employed a specific recombinant S. stercoralis larval antigen for the detection of anti-parasite antibodies, and a PCR-based analysis of urine for the detection of transrenal parasite DNA [19]. The Bayesian approach was used to address issues of misclassification of data because of different diagnostic targets and, importantly, the lack of a gold standard, to calibrate the diagnostic procedures. The selective modeling approach within a Bayesian framework also allowed us to establish a set of principled, evidence-based expectations about the diagnostic accuracy of the three methods and the overall prevalence, before incorporating the evidence from the observed data with the goal of improving the accuracy in estimates of regional prevalence of S. stercoralis.
Our study was a cross-sectional assessment of diagnostic tests in rural and urban communities in Northwestern Argentina, in the Departments of Oran, San Martin and Rivadavia in Salta province. Eligible communities were those assigned to a sanitary intervention program carried out by the teams from Universidad Nacional de Salta, the Regional Sanitarian Development Association NGO, ADESAR, and the Provincial Ministries of Public Health and First Infancy. The objectives of this collaborative network were to provide medical care and epidemiological surveillance of intestinal parasitic infections in remote villages of the Chaco and Yunga geographic regions. A total of 359 participants provided a stool, a urine, and a serum sample.
The study was carried out and reported in accordance with the Standards for Reporting Diagnostic Accuracy (STARD-BLCM) guidelines [20]. Ethical approval for the study protocol and the informed consent form were obtained from Comité de Ética, Colegio Médico de Salta, Salta, Argentina dated 19 March 2015, and Johns Hopkins University (IRB number 6199) dated 30 April 2015. All participants provided written informed consent prior to sample collection. Parents or guardians provided informed consent on behalf of minor participants. All members of these communities were invited to participate and received anthelmintic treatment free of charge based on the results of stool analysis.
Prior to the data collection, we performed a Monte Carlo simulation study that generated datasets of n = 400 observations 2000 times using a latent class analysis model in (Mplus 7 [21] [22]). The sample size was based on these simulations. In the model, based on an earlier study [15], we assumed that the true Strongyloides prevalence was 30%. Stool examination sensitivity was estimated to be 30%-40%. DNA and serological test sensitivity and DNA detection sensitivity was estimated to be 95% and 85%, respectively, based on previous work [23]. In the model we considered simulations involving four different tests. We assumed that the true prevalence of Strongyloides infection was 30% with stool examination sensitivity 70%, DNA and serological test 95% and antigen capture sensitivity 85% [24]. In this same model we also assumed stool examination specificity 100%, DNA specificity 98%, and serology specificity 75% and antigen capture specificity at 80% [24]. All parameters and standard error biases did not exceed 10% for any parameters in the model.
In the absence of a gold standard for diagnosis of this infection and to take into account the inherent data misclassification, we fitted a latent class model using a Bayesian approach to assess the performance of the three diagnostic procedures used in this study. The basic idea behind the Bayesian approach is that all unknown quantities/model parameters such as the true sensitivity and specificity of each diagnostic test as well as the prevalence of the infection are believed to have a distribution that captures uncertainty about these parameter values. This uncertainty is captured by a distribution that is defined before observing the data and is called the prior distribution or prior. Prior data are derived from previous information from publications or experience in the field [7, 15]. Next, the observed evidence (i.e. the actual data) is expressed in terms of the likelihood function of the actual data. The actual data likelihood is then used to weigh the prior and this product yields the posterior distribution. Thus, the posterior distribution is a parameter comprised of the prior distribution and the likelihood function. Such a process allows simultaneous inferences to be made on all model parameters [29].
In our study, for each model parameter, the particular beta prior density was selected by matching the center of the range of the mean of the beta distribution according to Joseph et al. [29]. For stool sensitivity and specificity, we assumed a range of 20–40% (mean = 30%; beta parameters a = 1.9, b = 4.444) and a range of 95–100% (mean = 97.5%; beta parameters: a = 420.3, b = 10.7), respectively [7, 30]. For NIE-ELISA serology, we assumed a priori sensitivity of 81–88%, (mean = 84.5%; beta parameters a = 45.7, b = 15.1) and a specificity of 71–81%, (mean = 76%; beta parameters a = 68.2, b = 21.5) [26, 31]. For the diagnostic sensitivity and specificity for PCR DNA as well as prevalence by age groups we assumed non-informative priors which correspond to beta parameters a and b = 1. To account for age in the model for the prevalence of Strongyloides infection, <15 years old were considered children and ≥15 years represented adolescents and adults. As a sensitivity analysis, we also changed input values by 10% in each afore-mentioned prior, to evaluate the impact of priors on model outputs. As the examined tests in the present study are based on different biological measurements, we have assumed that they are not correlated to any substantial extent and thus that they are conditionally independent on the latent infection status (i.e. the latent class in the fitted model). The software we used to fit such a model was WinBUGS [32]. Multiple chains were run and results examined to ensure convergence.
The percent total agreement between PCR and NIE-ELISA serology results was calculated and Cohn’s kappa statistic was used to assess the overall agreement in results [33]. Analyses were done using the ‘irr’ package in R. (https://cran.r-project.org/web/packages/irr/irr.pdf)
Table 1 outlines the distribution of the 359 participants in the study by age group (<15 years and ≥15 years), sex, and environmental context (rural vs urban). In the patients examined, 222/359 (62%) were positive for one or more of the diagnostic tests (Table 2). Serology and the transrenal DNA detection assays defined prevalence of 38% and 31%, respectively. In contrast, stool examination identified only ~8% of the participants as harboring an infection with S. stercoralis. There were no significant differences in infection status by any of the demographic categories used in this study.
The concordance in the assay outcomes between stool examination, serology, and the transrenal DNA test was evaluated (Fig 1). While ~53% of the participants were seropositive or transrenal DNA positive, only ~15% (52/359) were double positive for antibodies and transrenal DNA. Of the 359 samples examined total percent agreement between DNA and serology was only 61%. The kappa statistic was 0.131 with p = 0.0122 indicating the poor agreement between the two methods. Of the 30 patients who had detectable levels of parasites in their stool samples, 20 (66%) and 22 (73%) were positive by serological or transrenal DNA analysis, respectively. Over half of the stool-positive patients were also positive for serology and transrenal DNA (~4% of all patients). Therefore, nearly 70% of the seropositive participants tested negative for detectable amounts of parasite DNA in their urine and ~60% of the patients who were DNA positive had no detectable antibodies that bound to epitopes on the 31 kDa S. stercoralis L3 antigen.
Table 3 contains the results from the Bayesian LCA model estimates (i.e. posterior medians and 95% Credible Intervals (CrI), which are the Bayesian analogs of confidence intervals) for sensitivity and specificity for each of the three diagnostic tests and the prevalence of S. stercoralis infection for the two age groups. The S. stercoralis infection prevalence in persons <15 years was estimated as 13.5% (95% CrI 5.9–24.8) and for age ≥15 years this was estimated to be 19.8% (95% CrI 10.7–34.2). These estimates are based on Bayesian latent class modeling of collective values between the three diagnostic tests having taken into account associated measurement error from each test, not on the results of any one single test, and thus they are more accurate than the empirically calculated prevalence in Table 2. The estimate of sensitivity for serology slightly exceeded the estimate of the diagnostic sensitivity for urine-based PCR, but their 95% corresponding credible intervals overlapped, suggesting that the diagnostic performances of these two tests were similar. Specificity of urine based-PCR was estimated to be slightly higher than that estimated for serology, but, again, the 95% credible intervals overlapped with the corresponding estimate for serology. There was no substantial change in these results when the priors were altered by 10%. The Bayesian modeling results confirm the low sensitivity (43.6: 95% CrI: 25.7 to 70.4) and high specificity (97.9; 95% CrI: 96.5 to 98.9) of stool examination for S. stercoralis infection.
Soil-transmitted and other helminth infections are of increasing global importance and are the focus of several wide spread mass drug administration efforts to reduce the level of morbidity inflicted on endemic populations by these parasites [2, 3, 34]. As these programs progress and the prevalence and intensity of infection declines because of these interventions, it is imperative to employ diagnostic strategies with increasing sensitivity and specificity to monitor and identify lingering infections. Decisions to prematurely suspend regional intervention efforts that are made based on the results of diagnostic tests that provide inaccurate assessments of prevalence and intensity are likely to undermine both short-term and long-term programmatic goals. Indeed, models indicate that helminth control programs that terminate prior to a solid control of transmission will result in reemergence and spread of the parasite into susceptible populations with detrimental public health consequences [35, 36]. Given the limited sensitivity of many of the standard methods used to monitor the prevalence of helminth infections, it is time to revise the diagnostic strategies for these parasites.
The goal of the work presented here was to determine if the detection of parasite-derived transrenal DNA has the potential to enhance the sensitivity of diagnosing S. stercoralis infection over an established and widely used serological assay or the standard parasitological stool analysis. Although it is clear from this work that detection of transrenal DNA and serology have an advantage over conventional stool analysis for the identification of infection, the relative merits of transrenal DNA and serological analysis are more difficult to conclude. While each test identified approximately the same number of participants as infected with S. stercoralis, only about 22% (53/243) were positive for both assays. It is tempting to conclude that direct detection of a S. stercoralis-derived molecule (transrenal DNA) is superior to the indirect measure of detecting antibodies that recognize a restricted set of epitopes associated with a single, stage-restricted parasite protein. However, in the absence of a ‘gold standard’ test, or set of reagents against which the accuracy of these two tests can be measured, such a determination cannot be made. The impact that a lack of gold standard tests has had on the development of molecular-based parasite diagnostics has been expertly reviewed elsewhere [29, 37–39]. The absence of a gold standard has prompted us [23] and others [29, 37] to employ Bayesian latent class modeling to generate estimates of specificity and sensitivity for parasite diagnostic tests. For the S. stercoralis diagnostic tests used in this study, latent class analysis confirms the low sensitivity of stool examination and concludes that the diagnostic performance of the NIE ELISA and transrenal DNA tests were similar in terms of diagnostic sensitivity and specificity (Table 3).
The limited concordance of the results from the serological and transrenal DNA tests can be of significant importance when MDA control efforts are evaluated. The low concordance may be due, in part, to the single molecule focus of these two assays. The NIE ELISA uses a recombinant form of a 31 kDa molecule expressed by infective S. stercoralis larvae [27] and was chosen for its favorable sensitivity and specificity profile [26] as well as its performance in clinical settings [7, 31]. The demonstrated utility of the NIE ELISA notwithstanding, both the sensitivity and specificity of this assay would likely benefit from the strategic inclusion of additional parasite molecules expressed by somatic cells of adults or released components of the parasite’s excretory/secretory products. Likewise, the transrenal DNA assay targets a single repeat sequence, the absence of which does not infer a negative diagnosis [23]. While it is possible that the clinical and/or parasitological status of certain patients preclude the passing parasite-derived transrenal DNAs, it is also likely that Strongyloides DNA was present in the urine but derived from other regions of the parasite’s genome. Identifying these additional transrenal sequences would provide an opportunity to devise a multiplex assay that amplifies several transrenal DNAs to enhance the diagnostic sensitivity and specificity of this approach.
The estimates for the half-life of cell-free DNA in the blood of humans range between 4 minutes and 12 hours (reviewed in [40]). Assuming that the proximate source of transrenal DNAs is the cell-free DNA in the blood, this short half-life indicates that detection of Strongyloides-derived DNA in the urine is measuring an ongoing infection. This rapid decay in the blood also suggests that testing for the presence of transrenal DNAs could be a sensitive tool to measure the efficacy of chemotherapeutic elimination of the parasite. In support of the utility of using transrenal DNAs as a marker of successful chemotherapy, Ibironke et al. [41] demonstrated that Schistosoma haematobium transrenal DNA was no longer detectable 14 days after treatment with praziquantel.
Diagnostic approaches that can accurately assess changes in disease burden and the impact of chemotherapeutic/public health for programs that are at different levels of control (breaking transmission, elimination, or post-elimination) are critical for strategic decision making. Following multiple rounds of treatment, MDA programs require highly sensitive assays to identify hot spots of residual transmission. In most cases, there is an unmet need to replace microscopy, which is not sufficiently sensitive to detect these low-level residual infections. At this time, no single nucleic acid, antigen detection or antibody approach appears to be able to provide an appropriately high-resolution picture of infection status. Given this, it may be time to consider coordinating the results of two or more molecular based assays for the diagnosis of STH’s, including strongyloidiasis. The results presented here suggest that the combined use of assays that detect transrenal DNA and antibodies may be a useful approach.
|
10.1371/journal.pntd.0001696 | Oral Administration of GW788388, an Inhibitor of Transforming Growth Factor Beta Signaling, Prevents Heart Fibrosis in Chagas Disease | Chagas disease induced by Trypanosoma cruzi (T. cruzi) infection is a major cause of mortality and morbidity affecting the cardiovascular system for which presently available therapies are largely inadequate. Transforming Growth Factor beta (TGFß) has been involved in several regulatory steps of T. cruzi invasion and in host tissue fibrosis. GW788388 is a new TGFß type I and type II receptor kinase inhibitor that can be orally administered. In the present work, we studied its effects in vivo during the acute phase of experimental Chagas disease.
Male Swiss mice were infected intraperitoneally with 104 trypomastigotes of T. cruzi (Y strain) and evaluated clinically. We found that this compound given once 3 days post infection (dpi) significantly decreased parasitemia, increased survival, improved cardiac electrical conduction as measured by PR interval in electrocardiography, and restored connexin43 expression. We could further show that cardiac fibrosis development, evaluated by collagen type I and fibronectin expression, could be inhibited by this compound. Interestingly, we further demonstrated that administration of GW788388 at the end of the acute phase (20 dpi) still significantly increased survival and decreased cardiac fibrosis (evaluated by Masson's trichrome staining and collagen type I expression), in a stage when parasite growth is no more central to this event.
This work confirms that inhibition of TGFß signaling pathway can be considered as a potential alternative strategy for the treatment of the symptomatic cardiomyopathy found in the acute and chronic phases of Chagas disease.
| Cardiac damage and dysfunction are prominent features in patients with chronic Chagas disease, which is caused by infection with the protozoan parasite Trypanosoma cruzi (T. cruzi) and affects 10–12 million individuals in South and Central America. Our group previously reported that transforming growth factor beta (TGFß) is implicated in several regulatory aspects of T. cruzi invasion and growth and in host tissue fibrosis. In the present work, we evaluated the therapeutic action of an oral inhibitor of TGFß signaling (GW788388) administered during the acute phase of experimental Chagas disease. GW788388 treatment significantly reduced mortality and decreased parasitemia. Electrocardiography showed that GW788388 treatment was effective in protecting the cardiac conduction system, preserving gap junction plaque distribution and avoiding the development of cardiac fibrosis. Inhibition of TGFß signaling in vivo appears to potently decrease T. cruzi infection and to prevent heart damage in a preclinical mouse model. This suggests that this class of molecules may represent a new therapeutic tool for acute and chronic Chagas disease that warrants further pre-clinical exploration. Administration of TGFß inhibitors during chronic infection in mouse models should be further evaluated, and future clinical trials should be envisaged.
| Chagas disease, caused by the intracellular kinetoplastid parasite Trypanosoma cruzi, is a widely spread distributed debilitating human illness, affecting 10–12 million people in Central and South America. It is a major cause of mortality and morbidity, killing 15,000 persons each year [1], [2]. Chagas disease presents an acute phase of infection that is characterized by mild clinical symptoms (fever and malaise) and high parasitemia, but is often unmarked. Due to a potent specific immune response which control parasitemia, patients usually attain the indeterminate stage of the infection, with low-level of parasite persistence that can last from 10 to 40 years. About one in three infected individuals develops the symptomatic chronic stage of infection, which is characterized mainly by myocardiopathy or/and intestinal megasyndrome. A century has passed since the discovery of Chagas disease and the development of an efficient drug is still a challenge. As other neglected diseases, it has not received much attention of the pharmaceutical industry and present available therapies are insufficient [3]. Nifurtimox and benznidazole, the only two trypanocide drugs available, have toxic side effects, are not effective for all parasite strains and the effect in human chronic phase is still under clinical trial [4]. Moreover, no therapeutic approach targeting Chagas disease heart fibrosis is presently available.
Transforming Growth Factor ß1 (TGFß1) is the prototypic member of a family of polypeptide growth and differentiation factors that play a great variety of biological roles in such diverse processes as inflammation, fibrosis, immune suppression, cell proliferation, cell differentiation, and cell death [5], [6]. TGFß is also involved in many direct and indirect interactions between infectious agents and their hosts [7]. Several studies have demonstrated that TGFß plays a major role in the establishment and pathogenesis of T. cruzi infection (reviewed in [8]). Moreover, significantly higher circulating levels of TGFß1 have been observed in patients with Chagas disease cardiomyopathy [9] and in a culture system of cardiomyocytes infected by T. cruzi [10]. In order to establish its biological functions, TGFß must be activated into a mature form mainly by proteases, allowing its interaction with a specific transmembrane receptor called TGFß receptor-II (TßRII), which phosphorylates and stimulates the serine/threonine kinase activity of TßRI, also called activin receptor-like kinase 5 (ALK5). Upon activation, ALK5 phosphorylates the cytoplasmic signaling proteins Smad-2 and -3, which then associate with Smad-4, translocate into the nucleus as a multiprotein complex, and stimulate the transcription of TGFß-responsive genes, thereby inducing specific biological responses.
We have recently described that the ALK5 inhibitor, 4-(5-benzo[1,3]dioxol-5-yl-4- pyridin-2-yl-1H-imidazol-2-yl)-benzamide (SB431542) reduces the infection of cardiomyocytes by T. cruzi in vitro [11] and we could further show that it also inhibited T. cruzi infection in vivo and prevented heart damage in a mouse model [12]. This work therefore clearly demonstrated that blocking the TGFß signaling pathway could be a new therapeutical approach in the treatment of Chagas disease heart pathology. However the limitation of this compound was the preclusion to oral administration and some toxic effects. To reinforce the prove of concept, the aim of the present work was therefore to test, in the same parasite-mouse model of experimental Chagas disease, another inhibitor of the TGFß signaling pathway, 4-(4-[3-(Pyridin-2-yl)-1H-pyrazol-4-yl] pyridin-2-yl)-N-(tetrahydro-2Hpyran-4-yl) benzamide (GW788388) which can be orally administered and that has an improved pharmacokinetic profile [13], [14]. We found that GW788388 added 3-day post infection (dpi) decreased parasitemia, increased survival, prevented heart damage, and decreased heart fibrosis. Very importantly, we also demonstrated here for the first time that when added after the end of the intense parasite growth and consequent metabolic shock phase at 20 dpi, GW788388 could still decrease mortality and heart fibrosis.
Bloodstream trypomastigotes of the Y strain were used and harvested by heart puncture from T. cruzi-infected Swiss mice at the parasitemia peak, as described previously [15].
Mice were housed for at least one week before parasite infection at the Animal Experimentation Section at the Laboratory of Innovations in Therapies, Education and Bioproducts-IOC/FIOCRUZ under environmental factors and sanitation according to “Guide for the Care and Use of Laboratory Animals”. Animal studies adhered to the International guidelines (National Research Council. 1996, National Academy press, Washington, DC). This project was approved by the FIOCRUZ Committee of Ethics in Research (protocol number 028/09).
The aim of the present work was to evaluate whether the compound GW788388, which is an ATP-competitive inhibitor of the kinase activity of ALK5, could have a beneficial effect in vivo in an experimental model of mouse acute infection by T. cruzi and whether it could protect infected mice from parasite-induced alterations of cardiac functions and fibrosis when administrated early (3 dpi) and late (20 dpi).
In the first set of experiments, the inhibitor GW788388 was orally administered to male Swiss mice infected with 104 bloodstream trypomastigotes of the Y strain (day 0), at the 3rd dpi. We first performed a dose-response study by administering different doses of GW788388 (0.3, 3, 6 and 15 mg/kg) and analyzed parasitemia and survival rate. The results showed a dose-dependent inhibition of parasitemia at 8 dpi from 0.3 to 15 mg/kg of GW788388 (Methods S1 and Fig. S1A). On the other hand, the survival rate was increased with 3 or 6 mg/kg of GW788388 but unaltered at 0.3 and 15 mg/kg, suggesting some toxicity of the drug at this largest dose (Fig. S1B). For the subsequent studies, the dose of 3 mg/kg was chosen since it was the lowest GW788388 concentration that significantly affected parasitemia without worsening mortality. The choice for 3 mg/kg GW788388 administration was further reinforced by the assays performed by Gellibert and collaborators [13], who showed in a model of kidney fibrosis that doses as low as 3 mg/kg/mice of GW788388 significantly inhibited collagen type I mRNA levels. The control group received the vehicle buffer in which GW788388 was diluted (4% DMSO, 96% [0.5% Hydroxypropylmethylcellulose (HPMC), 5% Tween 20, 20% HCl 1 M in NaH2PO4 0.1 M]) and could be considered as the placebo group. The responses of DMSO-treated infected mice were not significantly different from those of untreated infected mice, excluding any sham or placebo effect (data not shown). In our model of acute infection, as previously described [12], parasitemia peaked at 8 dpi (Fig. 1A). We found that GW788388 administration at 3 dpi significantly reduced the blood parasitemia peak (Fig. 1A). Further, as previously described with the compound SB421543 [11], we could demonstrate that in vitro administration of GW788388 on cardiomyocytes impaired T. cruzi replication in host cells (Fig. S2) supporting the decreased parasitemia peak found in vivo. On the other hand, no effect of GW788388 on trypomastigote forms of T. cruzi viability could be observed after direct incubation of the drug with the parasites (unpublished result). We also showed that GW788388 administration significantly increased survival rates at 30 dpi (65% in the treated-group versus 34% in the untreated group, Fig. 1B). The infection induced a loss of body weight at 14 dpi [12], which was not modified by the administration of GW788388 (data not shown). To investigate whether GW788388 treatment would also affect myocardial parasitism and infiltration of inflammatory cells, we analyzed mouse infected heart sections collected at 15 dpi using histochemical techniques. Non-infected animals showed no inflammatory infiltration in the myocardium (data not shown). Myocardial sections from the T. cruzi-infected sham-treated group (Y DMSO) had many amastigote nests (Fig. 1C, open arrows) and large inflammatory foci (Fig. 1E, filled arrows) that were frequently associated with fibrotic areas. GW788388 treatment significantly decreased the number of amastigote nests (Fig. 1D and 1G). GW788388 administration also significantly decreased the area invaded by inflammatory infiltrates (Fig. 1F and 1H). A more detailed count of the number of cells per inflammatory foci showed that GW788388 treatment more particularly decreased the number of large inflammatory foci within the myocardium (larger than 20 or 50 cells per inflammatory infiltrates) (Table 1).
T. cruzi infection induces a strong hepatitis during the acute phase of Chagas disease [17]. We therefore analyzed several parameters of the liver in sham-treated versus GW788388-treated mice. Analysis of liver sections at 15 dpi revealed the presence of large inflammatory infiltrates in DMSO-treated animals (Fig. 2A, arrow). GW788388 administration significantly decreased the number of these infiltrates (Fig. 2B and C). We also measured two circulating markers of hepatic function which are induced by T. cruzi infection: AST (aspartate aminotransferase) and ALT (alanine aminotransferase). We found that GW788388 administration significantly decreased the serum levels of AST and ALT (Fig. 2D and E). We also measured urea, which reflects the renal functional status. Urea level was significantly increased at 15 dpi in DMSO-treated animals while GW788388 administration significantly reduced it (Fig. 2F).
We next analyzed electrocardiograms (ECG) of the different groups of mice at 15 dpi. As expected, analysis of the ECG demonstrated an atrial ventricular block with PR interval higher than 40 ms, leading to sinus bradycardia in sham-treated T. cruzi-infected animals as compared to the non-infected control group (495.8 and 774.2 bpm, respectively, Figure 3 and Table 2). GW788388 administration significantly limited the bpm decrease at 15 dpi, with a mean heart rate of 554.3 (Fig. 3 and Table 2). The other parameters analyzed demonstrated that infected mice had higher QT, PR and QRS intervals compared to non-infected mice (Table 2), and that GW788388 administration (3 mg/kg) also significantly decreased the QT intervals to 25.3 ms as compared to 29.6 in the infected DMSO-treated group (Table 2). A possible cause of this worsening in heart electrical conduction after the infection could be the direct effect of TGFß in heart cells. It has been already proposed that elevated TGFß levels during T. cruzi infection disorganize gap junctions, possibly contributing to abnormal impulse conduction and arrhythmia in Chagas disease [12]. To test this hypothesis, we measured connexin 43 (Cx43) expression in the different groups of mice. Heart sections from at least three mice per group at 15 dpi were immunostained for Cx43. We observed by confocal microscopy that non-infected hearts presented a dense structure of gap junction plaques (Fig. 4A, green staining). A drastic change in Cx43 expression was observed in the infected hearts of vehicle-treated mice, with an important decrease in Cx43 expression and a disruption of gap junction plaques (Fig. 4B). We found that GW788388 treatment reduced Cx43 disassembly and prevented the dissolution of gap junctions, preserving organized plaque distribution (Fig. 4C). The mean number of Cx43 plaques and their mean length were significantly lower in the heart of infected mice at 15 dpi as compared to the non-infected group (Fig. 4D and E). GW788388 treatment protected infected-mice from this loss as the decrease in the mean number of plaques was only reduced by 30% versus 45% in non-treated mice (Fig. 4D) and the mean length was similar to the non-infected mice (Fig. 4E). Immunoblotting analysis of Cx43 expression from heart ventricles confirmed these data (Fig. 4F and G).
One of the best established biological function of TGFß is the stimulation of extracellular matrix (ECM) protein deposition. Therefore, we checked whether GW788388 treatment would affect heart fibrosis that occurs in response to T. cruzi infection. Left ventricular heart tissues were obtained from each group and the deposition of ECM proteins was studied by immunostaining for collagen type I and fibronectin at 15 dpi. We observed an interstitial fibrous heart with high levels of both collagen type I and fibronectin deposition, as observed in red on Figure 5A and C, respectively. Interestingly, we could show that oral administration of GW788388 significantly reduced collagen type I and fibronectin levels (Fig. 5B and D, respectively). These data were confirmed by immunoblotting analysis of collagen type I and fibronectin expression from heart ventricles (Fig. 5E, F and G).
We found that GW788388-treatment decreased the phosphorylation level of Smad2 in infected hearts, demonstrating that GW788388-treatment was related to TGFß dependent signaling in vivo (data not shown).
Because most of the beneficial effects that we observed here with the TGFß inhibitor (GW788388) might be due to the resulting decreased parasitemia due to the inhibitory effect of TGFß signaling inhibitors in host cell invasion and intracellular proliferation [11], [12], we next studied the effect of GW788388 oral administration after the parasitemia peak. We chose to add GW788388 at 20 dpi as by this time, only 18% of infected mice survived and 30% of them died at 24 dpi. Interestingly, we found that GW788388 administration at 20 dpi completely protected these mice (n = 12) from death until 24 dpi (Fig. 6A, inset). In the inset, 100 represents the percentage of survival rate calculated from 20 dpi. GW788388 administration still decreased the number of inflammatory infiltrates within the myocardium (Table 3). To verify if GW788388 treatment presented an effect in the reversion of installed fibrosis, we performed Masson's trichrome staining on heart cross-sections of infected untreated mice at 15 dpi (Fig. 6B), 20 dpi (Fig. 6C) and 24 dpi (Fig. 6D), and of infected GW788388-treated mice at 24 dpi (Fig. 6E). We observed a progressive increase in collagen deposition visualized as light blue staining, which followed fibrosis progression (from 15 to 24 dpi, Table 4). At 20 dpi, which corresponded to the day of GW788388 administration, we observed a fibrotic pattern on the heart of infected mice frequently associated to inflammatory infiltrates (Fig. 6C). Interestingly, four days after GW788388 administration (i.e. 24 dpi) we observed a decrease in collagen deposition (Fig. 6E) as compared to the untreated group (Fig. 6D, Table 4). Immunoblotting assays were performed to compare the expression levels of collagen type I between each group. We observed a significant increase in collagen type I expression in the DMSO infected group as compared to the non-infected group (Fig. 6F and G, 9 fold increase), while GW788388 administration to infected mice significantly decreased the expression levels of collagen type I (Fig. 6F and G).
We have recently demonstrated that in vivo inhibition of the TGFß signaling pathway can decrease infection and prevent heart damage [12], suggesting that this new class of therapeutic agents should be considered in association with trypanocidal compounds for the potential treatment of Chagas disease cardiomyopathy. In the present work, we demonstrated that a more potent inhibitor of the TGFß signaling pathway, GW788388, which can be orally administered, significantly decreased parasitemia, increased survival and restored cardiac function as measured by ECG heart frequency (increase in bmp) and atrial conduction (decrease in QT interval). When administered at 3 dpi, we observed that GW788388 treatment reduced parasitemia and its subsequent deleterious effects. Whether the protective effect of GW788388 results only from this sole anti-infectious effect remains to be established. However, the short half-life of GW788388 in vivo (plasma T1/2 = 1.3 hours; [13]) makes it unlikely that it is mediated by long-term effects on e.g. fibrosis or cardiac rhythm. In contrast, administration of GW788388 at 20 dpi to mice that survived the metabolic distress syndrome clearly resulted in improved survival, which correlated with decreased cardiac fibrosis and has probably no causal relationship with the anti-infectious effect of the drug. Given the recent availability of reliable mouse models for chronic chagasic cardiomyopathy [18], the present proof that orally administrated GW788388 is feasible and efficient in the acute phase will offer in the near future the possibility of testing TGFß inhibitors in the chronic phase in pre-clinical assays. Taken together, these data further support that blocking TGFß signaling could represent a potential new therapeutic approach for Chagas disease heart fibrosis treatment.
It is now well established that the involvement of the TGFß signaling pathway plays an important role in the development of Chagas disease [8]. TGFß has been shown to be involved during parasite-host cell invasion, proliferation and differentiation [19]–[22]. Moreover, significantly higher circulating levels of TGFß1 have been observed in patients with Chagas disease cardiomyopathy [9], [16]. These data incited us to test the possibility of treating the development of Chagas disease by blocking the TGFß signaling pathway.
Here, we show that oral administration of GW788388 kinase signaling inhibitor prevents parasitemia, mortality, and heart fibrosis to acutely T. cruzi-infected mice in comparison to untreated-infected experimental group of animals. In lack of demonstration of GW788388 direct killing effect upon T. cruzi, we postulate the protein kinase inhibitor used may induce intracellular parasite latency [23], [24], such as that involved with the Plasmodium sporozoites cell cycle inhibition of initiation factor-2alpha (elF2alpha) kinase (IK2); its down-regulation by removal of PO4 from elF2alpha-P gives rise to the latency [25], [26]. In this regard, ongoing investigations in chronically T. cruzi-infected mouse model will determine whether GW788388 beneficial effects can be explained by the drug-induced parasite latency and long lasting cryptic infections.
Several approaches have been developed to abrogate TGFß signaling. Antibodies directed against TGFß have been administered in diabetic rodents and this was shown to efficiently prevent glomerulosclerosis and renal insufficiency [27]. Antisense TGFß oligonucleotides were found to reduce kidney weight in diabetic mice [28]. Recently, a soluble fusion protein of TßRII was reported to reduce albuminuria in a chemically induced model of diabetic nephropathy in rats [29]. And finally, inhibitors of the kinase activity of the TßRI (ALK5) have been developed. These inhibitors interact with the ALK5 ATP-binding site, thereby preventing TGFß intracellular pathways [30]. The first ALK5 inhibitor described, SB431542, is an ATP-competitive kinase inhibitor [31]. SB431542 significantly reduced procollagen1alpha (I) in rat kidneys in a model of induced nephritis. It was also described that SB431542 triggers antitumor activity in vivo [32]. Our work also demonstrated that SB431542 reduced mortality, decreased parasitemia and prevented heart damage as observed by histological and ECG analysis during the acute phase of experimental Chagas disease [12]. However, the limitations of SB431542 were the need of intraperitoneal injection and the in vivo toxic effects that have been demonstrated. Recently, GW788388 was developed as an alternative to SB431542 with better in vivo exposure. GW788388 is orally active and has a good pharmacokinetic profile [13], [14], [30]. GW788388 administration reduced liver and renal fibrotic response in a model of chemically induced fibrosis in rats and in the db/db mouse model of spontaneous diabetic nephropathy [13], [14]. Treatment with GW788388 also showed efficacy for preventing the fibrotic response in a skin fibrosis model [33] and attenuated cardiac dysfunction following myocardial infarction [34]. These data prompted us to test this compound during the acute phase of experimental Chagas disease.
We found that oral administration of GW788388 at 3 dpi significantly reduced peripheral parasitemia and lowered parasite load in hearts of infected mice observed 15 dpi. This effect was achieved with lower administration doses (3 mg/kg) than the one we previously used for SB431542 (10 mg/kg) [12], and with a single oral administration. More importantly, oral administration of GW788388 also significantly improved mice survival (70% in GW788388-treated mice against 30% in non-treated infected mice at 30 dpi). This is probably due to the combined impairment of the second wave of T. cruzi parasitemia due the decrease of parasite burden and of the early inflammatory cytokines secretion balance. Infection with T. cruzi in the acute phase is followed by a strong mononuclear cell inflammation on target tissues such as heart and liver, which could cause tissue disruption, necrosis followed by fibrotic deposition and abnormalities in electrical impulse conduction. Our data showed less inflammation on both heart and liver tissues and, moreover, less mononuclear cells by inflammatory focus. An improved ECG profile was also observed after GW788388 administration, characterized mainly by the absence of sinus node dysfunctions and reduced sinus bradycardia. PR intervals larger than 40 ms suggested slower transmission of the electrical impulses and atrioventricular block (AVB), which is characteristic of acute T. cruzi infection [35]. We observed an improvement of the QT intervals following GW788388 administration, which represent the wave of ventricular recuperation and this could be related to the decrease of sudden death [36] and to the progression to a pathological chronic phase [35]. Heart failure and sudden death are the most common causes of death in patients with chronic cardiac Chagas disease [37] and altered ECG parameters correlates with increasing myocardial scar and decreasing myocardial function in these patients [38]. This results from disorganized gap junctions that could contribute to abnormal impulse conduction and arrhythmia that characterize severe cardiopathy in Chagas disease and heart fibrosis [10]. Gap junction Cx43 molecules are responsible for electrical impulse conduction in the heart [39] and are affected by TGFß [10], [40]. We observed that GW788388 treatment preserved a correct Cx43 plaque pattern in the heart and blocked the down-regulation of Cx43 expression commonly observed following T. cruzi infection. GW788388 treatment therefore favored a regular and correct electrical impulse transmission. TGFß is also a key factor in the generation of tissue fibrosis [41] and has been correlated to development of Chagas disease symptoms in cardiac chronic phase [8]. Our data showed that administration of GW788388 to T. cruzi-infected mice significantly prevented the increase of fibronectin and collagen type I, two important components involved in heart fibrosis. These data are consistent with previous studies showing that GW788388 reduced fibrosis markers in the kidney following chemically induced nephropathy [14], [42].
In the human acute phase of Chagas disease, symptoms are frequently mild and not noticed and it is therefore difficult to propose correct treatments with trypanocidal drugs. Therefore, in the present study, we also treated mice with GW788388 at the end of the acute phase, when there are scarce circulating parasites. Interestingly, we found that oral administration of GW788388 at 20 dpi completely protected mice from death (100% survival). Analysis of cardiac fibrosis by Masson's trichrome staining on heart cross-sections of T. cruzi-infected mice showed a strong increase of fibrosis from 15 dpi to 24 dpi (Fig. 5, Table 4). Interestingly, we found that mice treated with GW788388, in a single dose scheme at 20 dpi, reversed heart fibrosis observed four days later (24 dpi) as compared to untreated infected mice. The level of collagen type I was also restored in GW788388 treated mice versus untreated mice. Taken together these data demonstrated that blocking TGFß signaling could decrease an installed heart fibrosis. This important finding encourages further pre-clinical assays targeting fibrotic lesions that are always involved in the severity of the clinical picture observed in the chronic cardiac disease. The development of an efficient drug for Chagas disease is still a challenge and trypanocidal drugs such as nifurtimox and benznidazole are still the only drugs employed for specific Chagas disease treatment, although the observation of serious side effects. Treatment strategy approaching the reversion of fibrosis has been demonstrated here at the end of the acute phase of experimental Chagas disease. Still, further studies on a chronic experimental model are necessary previously to clinical assays. The inhibition of TGFß signaling pathway and its biological functions could then be considered as an alternative strategy for the treatment of the symptomatic cardiomyopathy found in the acute and chronic phases of Chagas disease, in synergy with current administered drugs, enabling lower dosages and avoiding toxic effects.
|
10.1371/journal.ppat.1003703 | Defective Viral Genomes Arising In Vivo Provide Critical Danger Signals for the Triggering of Lung Antiviral Immunity | The innate immune response to viruses is initiated when specialized cellular sensors recognize viral danger signals. Here we show that truncated forms of viral genomes that accumulate in infected cells potently trigger the sustained activation of the transcription factors IRF3 and NF-κB and the production type I IFNs through a mechanism independent of IFN signaling. We demonstrate that these defective viral genomes (DVGs) are generated naturally during respiratory infections in vivo even in mice lacking the type I IFN receptor, and their appearance coincides with the production of cytokines during infections with Sendai virus (SeV) or influenza virus. Remarkably, the hallmark antiviral cytokine IFNβ is only expressed in lung epithelial cells containing DVGs, while cells within the lung that contain standard viral genomes alone do not express this cytokine. Together, our data indicate that DVGs generated during viral replication are a primary source of danger signals for the initiation of the host immune response to infection.
| In infections with viruses well adapted to the host virus-encoded proteins that delay the cellular response allow the virus to replicate to high titers prior to host intervention. The mechanisms overcoming viral evasion of the immune system and leading to the production of the primary antiviral cytokine IFNβ are not well established. Here, we demonstrate that truncated forms of viral genomes that are generated in situ during virus replication are a primary source of danger signals for the initiation of the host immune response to respiratory viral infections in vivo. Defective viral genomes (DVGs) are able to function as triggers of the immune response even in the absence of type I IFN signaling and are strong triggers of the host response to infection while overcoming viral antagonism.
| The recognition of virus-specific pattern associated molecular patterns (PAMPs) is a pivotal event in the initiation of the host innate response to infection. In recent years, it has been established that most viral danger signals are derived from oligonucleotide structures exposed during the replication of the viral genomes [1], [2], [3], [4], [5], [6]. However, most viruses produce proteins that antagonize and effectively delay signaling by the primary viral oligonucleotide sensor molecules retinoic acid inducible gene I (RIG-I) and melanoma differentiation–associated gene 5 (MDA5), allowing the virus to replicate to high titers and produce large amounts of danger signals prior to host intervention [7], [8]. It is currently unclear how the host immune response overcomes viral evasion to initiate a protective antiviral response.
Defective viral genomes (DVGs) arise when the viral polymerase loses processivity during virus replication at high titers, thereby generating truncated versions of the viral genome that contain deletions and/or complementary ends (the later known as copy-back or snap-back genomes) [9], [10]. DVGs with the ability to interfere with standard virus replication were first described by Von Magnus in the early 1950s as the genomes of incomplete forms of influenza virus called defective interfering (DI) viral particles [11]. DVGs have been identified in multiple distinct viral families when the viruses are grown in the laboratory at high multiplicity of infection and span a broad range of hosts, from plants to mammals [12]. Importantly, DVGs are found in patients infected with hepatitis A [13], hepatitis B [14], [15], hepatitis C [16], HIV [17], dengue virus [18], and influenza virus [19]. However, the biological role of DVGs in the context of natural infections is not well understood.
We and others have shown that stocks of Sendai virus (SeV) with a high content of copy-back DVGs with interfering activity trigger enhanced production of cytokines in vitro and more potently induce antigen presentation by mouse and human dendritic cells than do virus stocks lacking this kind of DVGs [20], [21], [22], [23], [24], [25]. Our group has also demonstrated that in contrast to standard viral genomes, SeV copy-back DVGs induce the expression of MDA5 and of a number of other interferon-stimulated genes in the absence of type I IFN positive feedback [23], [26], [27]. Remarkably, SeV copy-back DVGs show this potent in vitro stimulatory activity even in the presence of functional viral encoded antagonists of the host response [23], [24]. Here, we demonstrate that DVGs that trigger a robust activation of the transcription factors IRF3 and NF-κB accumulate at a high rate in infected cells becoming the main source of viral PAMPs. These DVGs arise naturally during acute respiratory viral infections in mice and provide essential stimuli for the initiation of the antiviral innate immune response in the lung. These data demonstrate the generation of DVGs in vivo during acute respiratory viral infections and suggest a critical role of these kinds of viral genomes in determining the quality of the host response to infection.
To further investigate the cellular mechanisms responsible for the efficient activation of the antiviral response by SeV DVGs, we evaluated the phosphorylation of transcription factors that are critical for the expression of type I IFNs in cells infected with equivalent amounts of infectious particles of a SeV strain Cantell stock containing high levels of copy-back DVGs (SeV Cantell HD) or with SeV Cantell depleted of DVGs (SeV Cantell LD). Virus stocks were prepared from the same parental virus and their content of DVGs was determined by calculating the ratio of infectious particles to total particles (ratios are specified in the material and methods section). In addition, copy-back DVGs of these stocks were identified by PCR. One predominant copy-back genome was present in cells infected with SeV Cantell HD (amplicon of 278 bp), while no copy-back defective genome was detected in cells infected with SeV Cantell LD up to six hours after infection (Figs. 1A and S1). Cloning and sequencing of the 278 nt long amplicon confirmed that it corresponded to a previously described SeV Cantell copy-back DVG of 546 nt in length (DVG-546) [28]. Phosphorylation of IRF3 and of the NF-κB repressor IκBα in response to SeV Cantell HD occurred rapidly and was sustained even in type I IFN receptor KO cells (Ifnar1−/−) (Fig. 1B and C), while no phosphorylation of IRF3 or IκBα was observed for up to ten hours post-infection with SeV Cantell LD despite equivalent or higher expression of the viral protein Np (Fig. 1D). Corresponding with the strong activation of transcription factors, Ifnb mRNA was expressed in Ifnar1−/− cells infected with SeV Cantell HD (Fig. 1E). In contrast, type I IFN signaling was required for the cellular response to Newcastle disease virus (NDV), an avian virus that only partially inhibits the type I IFN pathway, triggering the expression of type I IFN and other cytokines in the absence of DVGs.
To further validate the role of SeV copy-back DVGs as triggers of type I IFN-independent antiviral responses, we cloned DVG-546 under the control of the T7 polymerase promoter and used this construct to prepare a SeV stock containing a single recombinant DVG (rDVG). For this purpose we used SeV strain 52 that normally does not produce highly immunostimulatory copy-back DVGs [24]. Equivalent infectious units of SeV 52 and SeV 52 plus rDVGs had similar levels of total RNA (Fig. 1F) but infection with virus containing rDVGs strongly induced the antiviral response while virus that lacked DVGs did not (Fig. 1G) confirming the DVG immunostimulatory activity. In addition, presence of rDVGs significantly reduced the expression of SeV Np mRNA, demonstrating their strong interfering capacity (Fig. 1G). Notably, mouse embryo fibroblasts lacking the type I IFN receptor expressed Ifnb mRNA in response to SeV 52 containing rDVG (Fig. 1H) and virus containing rDVGs triggered IRF3 phosphorylation independently of type I IFN feedback (Fig. 1I), mirroring the response to SeV Cantell HD. Altogether, this evidence conclusively shows that SeV copy-back DVGs confer potent immunostimulatory ability to SeV stocks, independent of type I IFN feedback. Notably, potent Ifnb mRNA expression in response to SeV DVGs was independent of IRF1, IRF5, and IRF8 while only partially dependent on IRF7 (Fig. S2). This response was maintained in a variety of cell types (Fig. S3).
To determine whether standard viral genomes and DVG RNAs have distinct intrinsic properties that explain their differential immunostimulatory activities, we compared naked RNA purified from a stock of SeV Cantell LD with in vitro transcribed DVG-546. RNAs were transfected into cells before or after treatment with phosphatase or with RNase A that cleaves 3′ of single stranded C and U residues, and/or RNase V1 that cleaves base paired nucleotides. Both genomic RNA and DVG RNA were susceptible to treatment with phosphatase, as well as to treatment with RNases (Fig. 2A), corresponding with the literature that demonstrates a crucial role for 5′-triphosphate-RNA in the induction of type I IFNs. While transfected DVGs induced stronger expression of Ifnb than an equivalent concentration of gSeV (Fig. 2A), transfection of equivalent molar amounts of genomic and DVG RNA resulted in higher immunostimulatory activity of genomic RNA compared to DVG RNA (Fig. 2B), demonstrating that SeV LD RNA can strongly trigger the host response to infection when delivered naked into the cells.
Paradoxically, cells infected with SeV Cantell LD alone failed to induce strong type I IFN production even when used at a 10 times higher infectious dose than SeV Cantell HD (Fig. 2C). Although the amount of gSeV RNA was significantly higher in cells infected with an moi of 15 of SeV Cantell LD compared with ten times less SeV Cantell HD at 6 h post-infection (Fig. 2C), DVGs were only detected in cells infected with SeV Cantell HD, confirming a strong correlation between the presence of DVGs in the infected cells and the induction of the host response to infection. To determine whether the amount of total input viral RNA affected the immunostimulatory activity of SeV Cantell LD and HD, we measured the RNA content in equivalent infectious doses of these stocks. SeV Cantell HD had less than two fold higher the amount of total RNA than SeV Cantell LD and total RNA levels were equivalent between SeV Cantell LD and HD when LD was at twice the infectious dose (Fig. 2D). Thus, differences in the net input amount of viral RNA cannot explain the more than >1000 fold difference in the expression of Ifnb mRNA between cells infected with equivalent infectious doses of SeV Cantell LD and HD.
DVGs have an increased rate of replication compared to standard viral genomes due to their shorter size and promoter properties [29]. To determine whether DVGs replicate faster than gSeV, we calculated the rate of replication of gSeV and DVGs in cells infected with SeV Cantell HD. Although at an early time point more copies of gSeV than DVGs were detected in the cells, DVGs dominated by 12 h post-infection (Fig. 2E) accumulating at a 4 times faster rate than gSeV (Fig. 2F). These data demonstrate that DVGs rapidly surpass the number of gSeV in infected cells, providing large quantities of pathogen associated molecular patterns.
Supporting previous observations that DVGs from SeV stimulate the cellular antiviral response through signaling by RIG-I like receptors (RLRs) [1], [23], [24], the essential RLR adaptor protein mitochondrial antiviral signaling protein (MAVS) was required for the activation of the transcription factors IRF3 and NF-κB and for expression of numerous antiviral and pro-inflammatory molecules upon infection with SeV Cantell HD. In contrast, MAVS was not required for the response to herpes simplex virus, which can trigger the host response independently of RLRs (Fig. S4). In addition, only DVG RNA, but not standard viral genomes, could be amplified from endogenous RIG-I and MDA5 complexes immunoprecipitated from infected cells (Fig. S4C), supporting published evidence that DVGs bind to RIG-I preferentially over the standard viral genomes in infected cells [1]. As predicted, association of DVGs with RLRs correlated with type I IFN induction, but not with the level of virus replication (Fig. S4D). Importantly, in addition to the primary role of RIG-I in the response to SeV DVGs, MDA5 participates in the induction of type I IFN in primary mouse lung fibroblasts infected with SeV HD (Fig. S4E), similar to what we have observed in DCs [23], [27].
Overall, these data demonstrate that DVGs are produced in the infected cells at a higher rate than genomic RNA and that DVGs are the predominant ligands for both RIG-I and MDA5 during SeV infection.
Based on the potent ability of SeV stocks containing a high content of copy-back DVGs to induce the host response to infection in vitro [23], [24], [25], [28] (Fig. 1) and on our prior reports of strong host responses to DVGs regardless of the presence of functional virus-encoded antagonists [23], [24], we hypothesized that DVGs that arise in situ during viral infections provide essential stimuli to initiate an antiviral immune response. To test this hypothesis, we first determined if SeV strains that accumulate copy-back DVGs early in infection induced faster Ifnb mRNA expression in vitro than viruses with delayed DVG accumulation. For these experiments we used SeV preparations that did not show immunostimulatory activity or evidence of copy-back DVG accumulation by 2 h post-infection and all the viruses were used at a multiplicity of infection of 1.5 TCID50/cell. While standard viral genomes of all the different SeV strains used were detected at all tested time points, copy-back DVGs of different sizes were detected starting at 6 h post-infection in cells infected with SeV Z and at later time points in cells infected with SeV 52, Enders, or Cantell LD in both murine lung epithelial cells (TC-1) and bone marrow-derived dendritic cells (BMDCs) (Fig. 3A and B and data not shown). Sequences of the starred PCR products confirming the amplification of copy-back DVGs are shown in Fig. S5. Remarkably, accumulation of DVGs was directly associated with phosphorylation of IRF3 (Fig. 3C) and with the expression of Ifnb mRNA (Fig. 3D), demonstrating that standard viral genomes alone are not sufficient to initiate this response during infection in vitro and strongly supporting a unique ability of naturally arising DVGs to initiate the cellular antiviral response.
To evaluate the impact of DVGs during SeV infection in vivo, we infected mice with SeV Cantell HD or LD. Mice infected with SeV Cantell HD showed diminished morbidity than mice infected with the same infectious dose of SeV Cantell LD (Fig. 4A) despite equivalent levels of virus in the lungs at early times post-infection (Fig. 4B), agreeing with reports of reduced virulence in virus stocks with a high content of DVGs [30], [31], [32], [33], [34], [35]. Reduced virulence of SeV Cantell HD was associated with a stronger stimulation of the host antiviral response as shown by the expression of Ifnb mRNA (Fig. 4C). To conclusively demonstrate the role of DVGs in diminishing virulence in vivo, we co-infected mice with SeV Cantell LD and purified viral particles containing DVGs (defective particles; DPs). Confirming their critical role, DVGs reduced the pathogenicity of SeV Cantell LD in mice, while UV-inactivated DP particles did not provide significant protection (Figs. 4D–F). Interestingly, infection in the presence of DPs resulted in reduced expression of SeV NP protein in the lung at day 7 post-infection, suggesting that in this system, DPs reduce virulence by interfering with virus replication.
To determine whether immunostimulatory DVGs were generated in situ in the lung during infection, we infected mice with SeV Cantell LD, and we followed the appearance of copy-back DVGs in the lung by PCR. SeV copy-back DVGs were detected in whole lung homogenates at the time of high viral replication (Fig. 5A). Notably, upon infection with SeV Cantell LD, a copy-back DVG of high molecular weight was detected at day 3 post-infection in the lung, while a DVG of low molecular weight (amplicon of 278 bp) that predominates in the parent stock of SeV Cantell HD (Fig. 1A) was only detectable at day 5 post-infection. Copy-back DVGs also appeared in the lung of mice infected with SeV 52 (Fig. S6), showing that DVGs naturally arise during infection in vivo independent on the virus strain. Interestingly, accumulation of copy-back DVGs during infection with SeV Cantell LD was associated with the expression of Ifnb and Il-6 mRNA in the lung (Fig. 5B).
To determine whether DVGs were necessary for the expression of antiviral cytokines in vivo, we took advantage of IFNβ-YFP reporter mice. To demonstrate that YFP expression serves as readout for DVG activity, we first infected BMDCs prepared from IFNβ-YFP reporter mice with SeV Cantell LD alone, or together with increasing doses of purified DPs. As shown in Fig. 5C, at 6 h post-infection, YFP was expressed only in the presence of DPs and in a dose-dependent manner, and the YFP expression was lost when UV-treated DPs were used. DPs alone were also able to induce YFP in a dose-dependent manner, albeit at much lower levels than during co-infection with SeV. These data agree with our previous reports that demonstrate that the immunostimulatory activity of DPs is greatly amplified during DVG replication by the cognate polymerase provided by co-infecting SeV [23] and validate the IFNβ-YFP reporter system as a readout for DVG activity.
We then infected IFNβ-YFP reporter mice with SeV Cantell LD and analyzed viral genomes in YFP+ cells. We focused our analysis on the CD45− (non-hematopoietic) cellular fraction of the lung as SeV replicates predominantly in the lung epithelium [36]. Although full-length viral genomes were detected in both YFP+ and YFP− CD45− populations sorted three days after infection, DVGs were only found in YFP+ cells (Fig. 5D), suggesting that the presence of DVGs promotes IFNβ production in response to virus infection in vivo. Together, these findings show that DVGs are normally generated in situ in the lung during respiratory infection with SeV, and that their accumulation is associated with the expression of IFNβ in the lung.
To determine whether type I IFNs produced early upon infection promoted the generation of DVGs in the lung, we infected wild type or type I IFN receptor deficient mice (Ifnar1−/−) with SeV Cantell LD and analyzed the lungs at different times post-infection. As shown in Fig. 5E, DVGs accumulated in the lung at a higher rate in mice unable to respond to type I IFNs compared with wild type mice, corresponding with the predicted enhanced rate of virus replication in the lack of type I IFN signaling and demonstrating that type I IFNs are not required for the generation of SeV copy-back DVGs in vivo.
To investigate whether the content of DVGs in IAV stocks affects virulence similar to SeV, we obtained IAV strain PR8 stocks with a high content of DVGs (HD) or lacking DVGs (LD). The stock of IAV PR8 HD produced two predominant DVGs derived from the PA and PB1 genomic segments in infected cells, while no DVGs were detected in cells infected with IAV PR8 LD (Fig. 6A) (Strategy for IAV detection and sequences for the IAV DVGs present in infected cells can be found in Fig. S7). Mice infected with IAV PR8 HD showed reduced morbidity compared to mice infected with IAV PR8 LD (Fig. 6B) despite similar levels of virus replication (Fig. 6C). Similar to SeV Cantell HD, reduced morbidity was associated with enhanced Ifnb mRNA expression in the lung (Fig. 6D).
To determine whether IAV DVGs were generated in situ in the infected lung, we tracked their appearance in mice infected with IAV PR8 LD. Accumulation of DVGs was clearly observed at day 3 post-infection (Fig. 6E). Representative sequences of starred IAV DVGs products are shown in Fig. S8. Similar to SeV infection, accumulation of DVGs corresponded with enhanced expression of mRNA for Ifnb and Il-6 (Fig. 6F) despite evidence of reduced genomic viruses at that time point (Figs. 6E and F). These data demonstrate that DVGs are generated de novo in the lung during infections with IAV, and suggest an important role of these types of genomes in promoting the host response to IAV in vivo.
We have shown that DVGs are naturally generated in the lung during infection with SeV and IAV and provide primary danger signals for the triggering of the host response to infection. The generation of DVGs during virus growth in tissue culture is a highly conserved phenomenon among viruses of different species and is tempting to speculate that DVGs provide an evolutionary advantage to the virus by contributing to the preservation of both the virus and the host. Interestingly, immunostimulatory DVGs result from drastic truncations in the genome of the virus that render it a dead end product unable to persist in the absence of helper virus. It will be relevant to determine how DVGs relate to viral “quasispecies” that result from mutations as a consequence of having a viral polymerase with a lower fidelity and processivity [37]. Viral quasispecies have been shown to be essential for viral fitness and virulence [37]. Whether DVGs a tradeoff of this viral polymerase characteristic that enables more rapid virus evolution but makes the virus more vulnerable to innate immune detection remains to be established.
Our data demonstrate that DVG recognition is not necessary for the response to NDV, while is required for the response to SeV. We speculate that the differential DVG requirement may be explained by the poor adaptation of the avian NDV to grow in mice while the murine SeV is fully adapted to grow in this species. A critical factor of this adaptation is the activity of the virally encoded V and C proteins that effectively block the induction of type I IFNs, as well as the type I IFN-mediated amplification of the type I IFN pathway [38], [39], [40], [41]. As NDV is adapted to grow in birds, its antagonistic proteins are not fully functional in mammalian cells [42] allowing unrestricted production and amplification of type I IFNs. In contrast, the SeV C and V proteins very effectively block the cellular response to SeV [40], [42], [43], [44] and no cellular response is observed unless DVGs are present. Interestingly, in the absence of type I IFN feedback (or of the IFN-inducible transcription factor IRF7) SeV DVGs induce a more potent cellular response compared to NDV, suggesting that DVGs have a unique ability to bypass both virus antagonism and the requirement for IRF7 for strong type I IFN production. We have reported that SeV Cantell HD, but not NDV, has the ability to induce the expression of the viral sensor MDA5 independently of type I IFN feedback [27], and that MDA5 is involved in the recognition of SeV DVGs ([23] and data in Fig. S4). Although it is unclear why this newly synthetized MDA5 is less susceptible to inhibition by the SeV V protein, preferential binding of DVGs to both RIG-I and MDA5 compared to standard SeV genomes, together with the availability of high levels of MDA5, may explain the strong activation of transcription factors and type I IFN expression in response to DVGs, regardless of type I IFN feedback.
Remarkably, DVGs arise in vivo independently of type I IFN feedback, demonstrating that DVGs do not appear in response to host pressure via type I IFNs. Notably, viruses containing DVGs have significantly reduced virulence. In a previous study, we reported that a SeV strain with a lower propensity to produce DVGs (SeV 52) persisted longer in the lung than a SeV strain able to produce high levels of DVGs (Cantell) [26]. Consistently, we observed higher levels of viral NP protein at day 7 post-infection in the lung of mice infected with SeV Cantell LD, compared to mice infected with SeV Cantell LD plus DPs (Fig. 4G). In additional studies, we have not observed significant differences in the rate of SeV-specific T cells in the lung of mice infected with SeV LD alone or in the presence of DPs (data not shown and [26]). Although interpretation of this observation is complicated by the reduced amount of SeV antigen (NP protein) present in infection with SeV LD plus DPs, we favor the hypothesis that DPs diminish virulence by competing for the viral polymerase, thus interfering with the replication of the standard virus, a well-defined characteristic of DVGs. Additionally, enhanced production of type I IFNs in response to DVGs likely contributes to dampened viral replication.
Highly immunostimulatory SeV DVGs are of the copy-back type. Intriguingly, during SeV infections in vivo, accumulation of DVGs of high molecular weight preceded the appearance of low molecular weight DVGs, suggesting that the smaller ones may be secondary to longer defective genomic products. Copy-back DVGs are not transcribed due to their promoter properties [10], thus their stimulatory activity likely derives solely from their genomic composition. IAV DVGs are truncated versions of one of the genomic segments that have natural complementarity among their 3′ and 5′ ends providing the theoretical capacity to form structures similar to copy-back DVGs. Notably, it is apparent that both SeV genomic and DVG RNAs have the potential to induce a host response when delivered naked into the cells. Based on our data, we predict that in the context of infection, DVG RNA is more available for detection due to their enhanced rate of replication compared to standard viral genomes (Fig. 2). Interestingly, we have shown that DVGs have the ability to bypass viral-encoded antagonists of the immune response even upon overexpression of viral antagonistic proteins [24]. It remains to be investigated what is the molecular mechanism behind this DVG property.
Notably, DVGs of different forms and compositions have been described in the sera of patients chronically infected with a number of different viruses [13], [14], [15], [16], [17] and DVGs of various viruses have been shown to promote persistent infections in tissue cultures [45], [46], [47], [48], [49], [50], [51], [52], [53] supporting a role for DVGs in the maintenance of chronic viruses. The role of naturally arising DVGs in promoting virus persistence in vivo remains to be investigated.
In summary, we have demonstrated that DVGs arise naturally during an acute respiratory virus infection and that they play a critical role in regulating the virus-host cycle in vivo. Importantly, the recognition of DVGs as stimuli for the onset of immunity has multiple practical implications, most directly: (i) DVGs represent novel determinants of virus pathogenesis that could be targeted for therapy, and (ii) DVGs are novel candidate biomarkers to predict the outcome of infections and the rate of virus spread in the population.
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 (803176) was approved by the Institutional Animal Care and Use Committee, University of Pennsylvania Animal Welfare Assurance Number A3079-01.
TC-1 cells (mouse lung epithelial cells, ATCC, #DR-L2785), LLC-MK2 cells (monkey kidney cells, ATCC, #CCL7), MDCK (Madin-Darby canine kidney cells, kindly provided by Dr. T. Moran, Icahn School of Medicine at Mount Sinai), Baby hamster kidney-21 (BHK-21) cells expressing the T7 RNA polymerase (BSR-T7) (kindly provided by Dr. C. Basler, Icahn School of Medicine at Mount Sinai), and WT and Ifnar1−/− mouse embryo fibroblasts (kindly provided by Dr. B. tenOever, Icahn School of Medicine at Mount Sinai) were cultured in DMEM supplemented with 10% fetal bovine serum, 1 mM sodium pyruvate, 2 mL L-Glutamine, and 50 mg/ml gentamicin. C57BL/6 mice were obtained from Taconic Farms, Inc. IFNβ-YFP reporter mice (B6.129-Ifnb1tm1Lky/J) were obtained from The Jackson Laboratories. Ifnar1−/− mice were a kind donation of Dr. Thomas Moran (Icahn School of Medicine at Mount Sinai) and were bred in our animal facility.
SeV strains Cantell, 52, Enders, and Z, and influenza A/PR8/34 virus were grown in 10 days hen embryonated eggs (SPAFAS; Charles River Laboratories). SeV Cantell was passaged to retain its original high DI particle content (HD) or to deplete it of DI particles (LD) as we previously described [24]. In brief, SeV 52, Enders, Z, and Cantell HD were grown in embryonated hen eggs inoculated with 30,000 medium tissue culture infectious dose (TCID50) for 40 h. SeV Cantell HD TCID50 was calculated by end point dilution in LLCMK2 cells in the presence of trypsin, as described below [24]. SeV Cantell HD total particles were calculated by end point dilution of hemagglutination of chicken red blood cells. SeV Cantell HD stocks had consistently an infectious:total particle ratio of 5,000–15,000. SeV 52 had an infectious:total particle ratio of 24,472. SeV Enders had an infectious:total particle ratio of 38,746. SeV Z had an infectious:total particle ratio of 391,553. SeV Cantell LD was prepared by inoculating embryonated hen eggs with 3 TCID50 for 40 h. Under these conditions only 33% of the eggs grew virus. Allantoic fluid from those eggs was pooled and diluted 1/106 for subsequent inoculation into embryonated hen eggs in a total volume of 100 µl. Allantoic fluid containing virus (80% of the inoculated eggs) was pooled and tittered as described below. SeV Cantell LD stocks had consistently an infectious:total particle ratio of 100,000–200,000. IAV strain PR8 (LD) was grown by inoculating hen embryonated eggs with 30,000 TCID50 obtained directly from infected lung homogenates. Allantoic fluid containing the virus was collected 40 h later. Egg's allantoic fluid was snap frozen in an ethanol/dry-ice bath and stored at −80°C. IAV PR8 containing high dose of DI particles (HD) was kindly provided by Dr. Laurence C. Eisenlohr, V.M.D., Ph.D (Thomas Jefferson University). IAV HD was grown by inoculating hen embryonated eggs with 10,000 pfu of egg-passed virus. Eggs were incubated at 35°C and allantoic fluid containing the virus was collected 48 h later. IAV PR8 HD and LD stocks were originated from the same parent stock but were extensively passaged in the different conditions described.
Permissive cells were infected with serial 1∶10 dilutions of lung homogenates or virus stocks in the presence of 2 mg/ml of trypsin to determine the medium tissue culture infectious dose (TCID50). LLCMK2 cells were used for SeV titration, while MDCK cells were used for IAV titration. After 72 h of incubation at 37°C, 50 µl of supernatant from each well was tested by hemagglutination of chicken red blood cells (RBCs) for the presence of virus particles at the end point dilution. To do this, 1∶4 dilutions of the cell supernatant were incubated in 0.5% chicken RBCs at 4°C for 30 min. Hemagglutination of RBCs indicated the presence of SeV or influenza virus particles.
BMDCs were generated as previously described [24]. Detailed procedure can be found in the Supplemental Information Material and Methods. BMDCs were infected after 4 days in culture with viruses at an multiplicity of 1.5 as we have previously described [24].
For SeV infections, mice were anesthetized with tribromoethanol (Avertin®; Acros Organics) and inoculated in the nostrils with 30 µl of PBS containing 104 or 105 TCID50 of SeV. For IAV infections animals were infected intranasally with 100 TCID50/mouse in a 30 µl volume. Lungs were extracted at different times post-infection, homogenized in 0.1% w/v Gelatin-PBS and snap frozen in dry-ice/ethanol for preservation.
Total RNA was extracted from cell lines or lungs with TRIzol (Invitrogen) according to the manufacturer's specifications and total RNA was reversed transcribed using the high capacity RNA to cDNA kit from Applied Biosystem. For sorted cells, 500 ng of RNA were reversed transcribed, for all other experiments 1–2 µg of RNA were reversed transcribed. cDNA was diluted to a concentration of 10 µg/µl and amplified with specific primers in the presence of SYBR green (Applied Biosystem). For the detection of DVGs, isolated total RNA was reverse transcribed using Superscript III without RNase H activity, to avoid self-priming by the DVGs complementary ends and recombinant RNase H (Invitrogen) was added later to the samples. For the detection of the standard virus genome, the negative strand of the full-length genome was reverse transcribed with Transcriptor First Strand cDNA synthesis kit (Roche). PCR detection for IAV was performed using as it has been previously described [54]. Primers and detailed PCR conditions can be found in the Supplemental Information Material and Methods.
Detailed primers and PCR conditions can be seen in the Supplemental Information Material and Methods.
Whole cellular extracts were prepared by lysing 3×106 of cells in a NP-40-based lysis buffer containing phosphatase inhibitors, proteinase inhibitors (Roche and Thermo Scientific), and 0.5 M EDTA. The concentration of protein was measured by Bradford assay (Themo Scientific). Samples (25 µg) were boiled for 5 min and resolved on 10% Bis-Tris pre-cast gels (Bio-rad). Resolved proteins were transferred to a polyvinylidene fluoride (PVDF) membrane (Millipore). The membrane was blocked with 5% non-fat milk and immunoblotted with the indicated antibodies. Anti-rabbit IRF3, anti-rabbit phospho-IRF3 (Ser396), anti-mouse IκBα, anti-mouse phospho-IκBα (Ser32/36), and anti-rabbit IgG (HRP-conjugated) were purchased from Cell Signaling. Anti-mouse GAPDH was purchased from Sigma. Anti-mouse IgG and anti-mouse IgG1 (HRP-conjugated) were purchased from Jackson Immunologicals. Lumi-Light western blotting substrate was used for HRP detection (Roche).
DP purification was performed as previously described [24]. In short, allantoic fluid from 100 infected hen eggs was pooled and concentrated by high-speed centrifugation. Pellets were suspended in 0.5 ml of PBS/2 mM EDTA and incubated overnight at 4°C in a 5–45% sucrose (Fisher) gradient that was prepared using a gradient maker (BioComp). Gradients were centrifuged at 4°C for 1.5 h at 28,000 rpm and fractions containing low-density viral particles were collected, pelleted, suspended and re-purified using the same procedure. Collected low-density fractions were concentrated by centrifugation at 4°C for 2 h at 21,000 rpm. Pellets were suspended in PBS, snap frozen, and stored at −80°C. The content of DI particles was determined by calculating the ratio of infectious over non-infectious particles as described above.
A 591 nt long product containing the sequence of the T7 promoter followed by the 546-nucleotide long copy back DVG from SeV Cantell, and flanked by the restriction enzymes SpeI and SapI at the 3′ an 5′ ends was synthetically synthetized (DNA 2.0) and clone into the pSL1180 vector (Amersham Pharmacia Biotech) containing the sequences for the hepatitis delta virus ribozyme and the T7 polymerase terminator. In order to optimize the transcription of the DVG, 3 G residues were introduced downstream of the T7 promoter by site-directed mutagenesis (Stratagene, CA) using the oligonucleotides 5′CCACTAGTTAATACGACTCACTATAGGGACCAGACAAGAGTTTAAGAG-3′ and 5′CTCTTAAACTCTTGTCTGGTCCCTATAGTGAGTCGTATTAACTAGTGG-3′. BSR-T7 cells were infected with a moi of approximately 66 of partially inactivated SeV strain 52. Virus inactivation was performed by exposing diluted virus to UV light (254 nm model MRL-58, UVP Upland, CA) for 53 sec at a distance of 9 inches from the light source. Virus inactivation diminished the virus replication rate, while allowing the expression of viral proteins necessary for the replication of DVGs. Cells were incubated at 37°C for 1 h before transfection of 3 µg of vector encoding DVG. Transfection was performed with XtremeGENE transfection reagent (Roche) according to manufacturer instructions. Cells were cultured in Dulbecco's modified Eagle medium (DMEM) supplemented with 1% bovine serum albumin, 2% NaCO3, 0.5 µg/ml trypsin (Worthington) and 0.1% penicillin-streptomycin (Invitrogen) and incubated in 7% CO2 at 37°C. Cells and supernatant containing SeV 52 and rDPs were harvested after 48 h and 200 µl of the suspension were inoculated in the allantoic cavity of 10-day embryonated hen eggs (B & E Eggs, Silver Springs, PA). After 40 h allantoic fluid was harvest and 200 µl of undiluted fluid were inoculate in 10-day embryonated eggs for virus growth and egg inoculation was repeated for three consecutive passages. Allantoic fluid from the third passage was quick-frozen in dried ice/ethanol and used for infections. Presence of recombinant DVG was confirmed by PCR. No other DVGs were detected.
DVG RNA was in vitro transcribed (Ambion) from the T7-DVG 546 plasmid. Standard genomic RNA was extracted from SeV Cantell LD stocks. To remove 5′-triphosphates, 1 µg of RNA was incubated with 10 U of Calf Intestinal Phosphatase (New England BioLabs) for 60 min at 37°C. To cleave single stranded RNA, 1 µg of RNA was incubated with 1 ng of RNase A (Ambion) for 15 min at room temperature. To cleave double stranded RNA, RNA was incubated with 0.1 U of RNase V1 (Ambion) for 15 min at room temperature. After treatments, RNA was purified using TRIzol or precipitation/inactivation buffer according to the manufacturer's specifications.
LLC-MK2 cells were transfected with 250 ng or indicated doses of DVG and genomic RNA using lipofectamin 2000 (invitrogen). At 4 hours post transfection, the cells were harvested and total RNA was isolated using TRIzol according to the manufacturer's specifications.
Infected IFNβ-YFP cells were collected at 6 h post-infection. Reporter mice were sacrificed 3 days post-infection. Lungs were collected and dissociated with collagenase (Roche), followed by suspension on 0.5 M EDTA and RBC lysis buffer. Single cell suspensions were then incubated with CD16/CD32 FcBlock for 20 min at 4C, followed by incubation with biotinylated mouse anti-CD45.2. Washed cells were incubated with anti-biotin microbeads (Miltenyi) for 20 min and passed through a magnetic column for negative selection. CD45− cells were sorted based on YFP expression using a FACS Vantage SE sorter.
Statistical analyses were performed as indicated in each figure. GraphPad Prism version 5.00 for Windows, GraphPad Software, San Diego California USA, www.graphpad.com, was used for analysis.
Genes NCBI ID numbers. tuba1b: 22143; rps11:27207; ifnb: 15977; ifnl2: 330496; illb: 16176; il12b: 16160; tnf: 2926; Il-6, 16193.
|
10.1371/journal.pcbi.1004919 | A Molecular Clock Infers Heterogeneous Tissue Age Among Patients with Barrett’s Esophagus | Biomarkers that drift differentially with age between normal and premalignant tissues, such as Barrett’s esophagus (BE), have the potential to improve the assessment of a patient’s cancer risk by providing quantitative information about how long a patient has lived with the precursor (i.e., dwell time). In the case of BE, which is a metaplastic precursor to esophageal adenocarcinoma (EAC), such biomarkers would be particularly useful because EAC risk may change with BE dwell time and it is generally not known how long a patient has lived with BE when a patient is first diagnosed with this condition. In this study we first describe a statistical analysis of DNA methylation data (both cross-sectional and longitudinal) derived from tissue samples from 50 BE patients to identify and validate a set of 67 CpG dinucleotides in 51 CpG islands that undergo age-related methylomic drift. Next, we describe how this information can be used to estimate a patient’s BE dwell time. We introduce a Bayesian model that incorporates longitudinal methylomic drift rates, patient age, and methylation data from individually paired BE and normal squamous tissue samples to estimate patient-specific BE onset times. Our application of the model to 30 sporadic BE patients’ methylomic profiles first exposes a wide heterogeneity in patient-specific BE onset times. Furthermore, independent application of this method to a cohort of 22 familial BE (FBE) patients reveals significantly earlier mean BE onset times. Our analysis supports the conjecture that differential methylomic drift occurs in BE (relative to normal squamous tissue) and hence allows quantitative estimation of the time that a BE patient has lived with BE.
| Barrett’s Esophagus (BE) is a metaplastic precursor to esophageal adenocarcinoma (EAC). When a patient is diagnosed with BE, it is generally not known how long he/she has had this condition because BE is asymptomatic. While the question of how long a premalignant tissue or lesion has been resident in an organ (dwell time) may not be of importance for cases where curative interventions are readily available (such as adenomas in the colon), for BE, curative interventions are either costly or carry patient risks. Knowledge of a precursor’s dwell time may therefore be advantageous in determining the cancer risk due to the stepwise accumulation of critical mutations in the precursor. In this study, we create a molecular clock model that infers patient-specific BE onsets from DNA methylation data. We show that there is considerable variation in the predicted BE onset times which translates, using mathematical modeling of EAC, into large variation in individual EAC risks. We make the case that, notwithstanding other known risk factors such as chronological age, gender, reflux status, etc., knowledge of biological tissue age can provide valuable patient-specific risk information when a patient is first diagnosed with BE.
| There is great interest in the molecular characterization of precancerous fields and lesions (e.g., colorectal adenomas or ductal carcinoma in situ (DCIS) in the breast) to quantify their neoplastic potential, although it is generally not known how long such lesions (or fields) have sojourned in a patient when they are discovered. This point is of particular importance in the case of Barrett’s esophagus (BE), a variable-length metaplastic precursor of esophageal adenocarcinoma (EAC) that has been shown to undergo a stepwise progression to cancer involving multiple rate-limiting events [1–3]. In spite of a generally low EAC progression risk of about 0.2–0.5% per year across BE patients [4], the progression risk is believed to be highly variable and dependent on age, gender, histopathological grade, and personal risk factors such as severity of gastroesophageal reflux disease (GERD), body mass index (BMI), and smoking status [5]. However, since the total number of BE patients who progress to EAC is generally low for most epidemiological studies (mostly due to limited follow-up), inter-individual variability in progression risk is difficult to specify other than by gross factors. Furthermore, the clinical assessment of the BE tissue is known to be fraught with uncertainty as only a small portion of the tissue is biopsied for pathology. Thus, there is a pressing need to develop more accurate markers (and risk stratifications) that identify BE that is more likely to progress to EAC in a person’s lifetime versus BE that is indolent or has low neoplastic potential.
Inter-individual variability in the EAC progression risk may depend on the duration of how long a patient has lived with BE (BE dwell time). In a large population-based study in Northern Ireland, Bhat et al. [6] found a significant increase of the annual progression risk with patient age (2-fold from age <50 to age 60–69) suggesting that the BE-to-EAC progression risk is not constant but rather increases with the age of the BE tissue due to the stepwise accumulation of genetic and epigenetic alterations that drive premalignant and malignant progressions in BE [1, 2, 7]. Thus, a longer dwell time for BE may increase the risk for neoplasia and cancer in an exponential manner consistent with the exponential increases observed in the age-specific incidence of EAC in the general population [8, 9]. Also, in an environment of chronic inflammation analogous to that which is caused by GERD within BE, patients with ulcerative colitis have a higher colon cancer risk that increases with earlier age of onset and disease duration [10, 11]. These risk factors unfortunately cannot be identified clinically in the case of BE because BE is asymptomatic. Yet, the use of mathematical modeling to quantifiy the waiting (or dwell) time of premalignant stages during carcinogenesis until the occurrence of cancer has been of considerable interest [12].
Recently identified age-related changes in DNA-methylation have led to the notion of a biological tissue age which, although highly correlated with chronological age, may differ significantly from it [13, 14]. It is generally believed that epigenetic drift (i.e., neutral changes in DNA methylation levels) is responsible for this process [15]. In this study we examine array-based methylation patterns of CpG-dinucleotides across the genome to determine whether CpGs that drift differentially between BE and normal tissue can be used to infer the relative biological age of a patient’s BE tissue. Specifically, we identify CpGs that undergo such ‘methylomic drift’ based on array data from formalin fixed paraffin embedded (FFPE) tissue samples from two groups of BE patients: one group of 10 patients each with 2 or more tissue samples that were obtained at least 5 years apart (data set D1). These samples provide longitudinal information at the individual level. A second group of 30 patients ranging in age from 21 to 88 (data set D2) had matched tissue samples obtained from Barrett’s esophagus and adjacent normal esophagus squamous epithelium (SQ), providing cross-sectional information as well as differential drift information between SQ and BE tissue. The combined statistical analyses of these two data sets, as described in Materials and Methods, suggest that numerous hypomethylated CpG sites undergo significant differential methylomic drift in BE versus SQ. Significantly, the observed patient-specific drift differentials appear relatively uniform across the set of identified 67 CpGs, giving rise to high correlations in the methylation differentials (against the mean drift) between CpGs. Thus, a hallmark of methylomic drift is that the associated methylation differentials between markers (across patients) are highly correlated, as are all clocks that keep time. We also validated the computed methylomic drift rates for the 67 selected CpGs in an independent data set of 10 additional BE patients (data set DV) each with samples at two time points.
To infer patient-specific BE onset times from the measured methylation levels of identified CpGs that drift differentially between BE and SQ tissues, we use a Bayesian model that accounts for (CpG-specific) random effects in drift rates, measurement error, and a patient-specific BE onset time. Furthermore, to gain insights into how the age of BE onset may influence EAC risk, we used a recently developed mathematical model for EAC incidence to compute standardized lifetime risks for the individuals in data set D2 given their predicted BE onset times [8, 16]. Additionally, we applied this methodology to methylation array data from 22 familial BE (FBE) patients (data set D3). The quantitative predictions of both BE onset times and inferred EAC risks for BE patients without neoplasia (D2) and familial BE (D3) suggest that BE onset is a useful event-marker of cancer risk. In the following we describe the data and methodologies that support this conclusion.
All CpG-methylation data for this study were generated with the Infinium HumanMethylation450 beadchip arrays (Illumina) [17, 18] that include over 485,000 CpG-methylation sites throughout the genome (covering 99% of Reference Sequence (RefSeq) genes (National Center for Biotechnology Information (NCBI), Bethesda, MD, USA). Data normalization was performed using the R Bioconductor minfi package, which includes background level corrections, color adjustments and Subset-quantile Within Array Normalization (SWAN) normalization. SWAN is specifically designed for HumanMethylation450 array data to account for systemic differences between the Infinium I and Infinium II probe designs [19]. Next we filtered out unreliable, gender bias, and noisy probes from downstream analysis, including probes having the average detection p-values across samples greater than 0.05, chromosome X-associated probes, and those containing at least one SNP with low minor allele frequency (MAF = 0) in the probe body [20, 21]. For linear regressions of the probe-specific methylation fractions on patient age we used M-values rather than β-values to better account for epigenetic drift that occurs at very low (<1%) and high levels of methylation. M-values are logit2-transformed β-values (computed using Illumina’s formula β = M/(M + U + 100)), allowing for non-linear saturation effects of methylation fractions with age at both ends of the methylation spectrum. Note, at the molecular level, CpG-methylation is essentially a binary variable (a CpG dinucleotide is either methylated or unmethylated). However, in a tissue sample, only cell population averages can be measured across all epigenomes in that sample.
The human tissues used for the analyses presented here were obtained from 72 patients with confirmed Barrett’s esophagus (BE). Written informed consent was obtained, signed by all participants, and conformed to institutional ethics requirements. IRB approval (protocol numbers 1989, 8137) was given by the ethical review board of the Fred Hutchinson Cancer Research Center.
We examined levels of DNA methylation at over 450,000 CpG sites in tissue samples from four groups of BE patients (see S1 Table for detailed patient information). The first data set (D1) is unique and consists of serial samples from 10 BE patients, ages 33–70 years at index biopsy (mean age = 51.2), with 2 or more tissue biopsies each that were collected at least 5 years apart to comprise a total of 29 samples. D1 patient data for two particular CpGs that show longitudinal drift for each of these 10 patients’ serial sample sets are shown in Fig 1.
The second, cross-sectional data set (D2) includes matched BE and normal squamous esophageal epithelium (SQ) tissue samples from 30 BE patients ages 21–88 years (mean age = 63.4) comprising a total of 60 tissue samples. While the D1 data provide some information on methylomic drift in BE tissue for each patient, the aggregated cross-sectional data also provide population-level information on the mean drift rate across all patients and ages. Although methylomic drift may depend on various factors, here we will focus on the influence of BE dwell time, which may be highly variable from patient to patient, even for patients of similar age. Fig 2 shows the probability densities of BE onset for two representative D2 patients’ ages at time of biopsy (a1 = 21, a2 = 80), and the theoretical consequence their ages will have on the statistical inference of their BE onset ages. The inter-individual heterogeneity in BE onset times will thus affect the methylation level data around the mean population drift. An illustration for a single CpG site j for the BE samples from D2 is shown in the insert of Fig 2. Note, for the cross-sectional group (D2), the matched BE and SQ samples originate from biopsies collected during the same endoscopic exam.
The third serial data set (DV) consists of 10 BE patients from Cleveland Clinic Foundation, ages 54–77 years at index biopsy (mean age = 51.2), with 2 serial tissue biopsies each, comprising a total of 20 BE samples.
The fourth data set (D3) includes BE tissue samples from 22 familial BE (FBE) patients ages 39–84 years (mean age = 62.8) with one sample per patient. Familial Barrett’s esophagus (FBE) was defined as having a first- or second-degree relative with long-segment BE, adenocarcinoma of the esophagus, or adenocarcinoma of the gastroesophageal junction whose diagnosis was confirmed by review of endoscopy and histology reports [22]. The data also include gender and age when the tissue biopsy was collected for each patient (see S1 Table).
Two concepts have so far emerged that relate alterations in DNA methylation to biological tissue age. The first is based on the discovery of sets of clock-CpGs that undergo age-dependent changes in methylation that in combination correlate strongly with chronological age [13, 14, 23]. The second concept relates to subtle changes in methylation levels due to epigenetic drift as a result of a semi-conserved replication process of DNA-methylation patterns [24–27]. Significantly, some CpG-islands that show very low (hypo-)methylation levels early in life are known to undergo gradual methylation over time, presumably as a result of sporadic de novo methylation events during DNA replication, a process commonly understood as epigenetic or methylomic drift [15, 24, 28–31]. Therefore, to narrow the number of CpG candidates that may serve as markers for differential tissue aging in the emerging metaplastic tissue of BE patients, we first identified CpGs that show significant longitudinal drift among the patients of our longitudinal study D1, as described below.
The following steps summarize our discovery pipeline in more detail.
Here we show how information about methylomic drift characteristic of BE and differential between BE tissue and normal squamous (SQ) tissue can be combined with individual-level methylation data at a given age to predict when a patient developed BE assuming there is a single time point of origin for BE. Our model (described below) employs Bayesian inference to derive dates of BE onset via initial differential drift away from squamous methylation values, and in this way our method can be considered somewhat analagous to dating divergence times in phylogenies with a relaxed molecular clock [32]. In the following we assume that methylomic drift is essentially linear with age (at the logit scale), although there is also evidence that age-associated variation in methylation levels may be better modeled by a function of logarithmic age for younger individuals [23]. However, this approach has the flexibility to accommodate non-linear drift.
For patient i, i = 1, …, N, the data consist of measurements yBEi,j(ti) for BE clock CpGj (j = 1, …, 67) at observation time (age) ti = ai. We consider the following linear drift model for the conditional expected methylation values of variable YBEi,j(ti), taken from patient i at time ti for each clock CpG, given the onset of BE occurred at time si ≤ ti,
E [ Y B E i , j ( t i ) ] = α S Q j + b S Q j s i + b i , j ( t i - s i ) , (1)
for j = 1, …, 67. Thus, given the following parameters—the onset of BE at time TBE = si, the rate (bSQj) and intercept (αSQj) of the SQ population regression lines obtained from individuals with matched samples in data set D2, and the patient-specific, CpG-specific BE drift rate bi,j—we observe 67 independent measurements for N independent individuals. Furthermore, we used the linear regression slopes and intercepts provided by the ANCOVA procedure using the normal squamous sample group in D2 to impute αSQj and bSQj in D3 for each BE clock CpG, as implemented in the model shown in Eq (1). For this data set, we did not have matched SQ samples but because the methylation values in normal squamous tissue show little variation for our selection of BE clock CpGs, we assumed that the normal squamous tissues behave similarly for non-familial and familial patients. We show that this approach for imputing SQ M-values for non-matched samples is robust in a sensitivity analysis given in Results. Allowing for patient-specific drift rates for the BE clock CpGs, we explicitly model the inter-individual differences in BE drift rates between ‘slow’ and ‘fast’ aging BE tissues relative to the standard clock, which are measured from means and standard deviations of the serial samples.
Again, the observation from a single patient i, for i = 1, …, N, observed at time ti, is of the form
y i = { y B E i , j , j = 1 , ⋯ , 67 } . (2)
In the Bayesian BE clock framework defined by Eq (1), the likelihood contribution from a single patient observed at time ti is given by
∏j=167f(yBEi,j)=∏j=167fN(yBEi,j;μBEi,j=αSQj+bSQjsi+bi,j(ti−si),σBEi), (3)
where fN is the normal density function. For the Bayesian model we further assume uniform priors ps(si) for the BE onset times si (due to the fact that the distribution of BE onset times in the general population is essentially unknown), conjugate gamma priors pσ(σBEi) for the standard deviation σBEi of methylation measurement values using shape and scale parameters fitted to the distribution of non-drifting CpG measurements, and normal prior distributions pb(bi,j) for the drift rates bi,j, j = 1, …, 67, which were derived from the longitudinal data sets with empirical mean and standard deviation (see S1 Text for full expressions of prior distributions).
In order to ultimately simulate the BE onset times s1, …, sN from the corresponding patient-specific posterior distributions, let us define the vector Ψi = (si, bi,1, …, bi,67, σBEi) for patient i. Samples of Ψi under its posterior distribution for patient i will be obtained using Markov Chain Monte Carlo (MCMC). The posterior distribution of Ψi given the observation yi comprised of patient-specific data of the form in Eq (2), for i = 1, …, N, is given by
π ( Ψ i | y i ) ∝ likelihood · prior (4) = ∏ j = 1 67 f N ( y B E i , j ; μ B E i , j , σ B E i ) · p s ( s i ) · p b ( b i , j ) · p σ ( σ B E i ) . (5)
To estimate the model parameters of this Bayesian BE clock model we used MCMC with Gibbs sampling [33]. All the full conditionals are known distributions. Specifically, for each individual i, i = 1, …, N, we estimated the posterior means, medians, and other quantiles of the BE onset time si, patient-specific, CpG-specific drift rates bi,j, j = 1, …, 67, and patient-specific standard deviation of measurements parameter, σBEi. All MCMC simulations were run for 100K cycles and allowing 1K cycles for burn-in.
The Bayesian BE clock model requires specification of a prior distribution pb(bi,j) for the drift rates bj, j = 1, …, 67 of the BE clock. In the preselection pipeline described above (Step 1), we obtained mean drift rates (slopes) and standard deviations for each arrayed CpG in the longitudinal study D1. To illustrate the degree of variability and uncertainty in the estimated drift rates we show normal distributions with those means and standard deviations individually (in Fig 5, light dashed green curves) and aggregated as a single normal distribution (solid green curve). To validate the methylomic drift associated with these 67 BE clock CpGs in an independent longitudinal data set (denoted as DV), we used the procedure described in Step 1 to evaluate the drift rates (regression slopes) for each of the 67 CpGs. The results are shown in Fig 5, analogous normal distributions for each of the 67 CpGs in the clock set individually (light dashed purple curves) and in aggregate (solid purple curve) for the validation set DV. S3 Fig shows a scatterplot of mean drift rates between data sets D1 and DV. As expected, overall we observe slightly decreased means and increased variances in the drift rates of the clock CpGs in the validation set DV, a phenomenon commonly referred to as “winner’s curse”, reflecting the typical overestimation of effect sizes in discovery samples (see Fig 5). Ultimately, there was minimal effect of this bias conferred on posterior parameter estimates (see S1 Text).
In Results, we will apply the Bayesian BE clock model to estimate model parameters for 2 patient data sets independently—cross-sectional (D2) and FBE (D3). To formally assess differences between different patient groups, we use Bayes factors to statistically test if the BE onset ages estimated for one group si, i = 1, .., Nk, lead to BE dwell times that are significantly different from those of a second patient set with estimated BE onset ages s i ′, i = 1, .., Nl, for k, l ∈ {2, 3}. For two specified data sets Dk, Dl, we compare the average fraction of life until age at biopsy (ai) during which the patient harbored BE. This quantity is given for two data sets by the following variables,
γ k = 1 N k ∑ i = 1 N k a i - s i a i , γ l = 1 N l ∑ i = 1 N l a i ′ - s i ′ a i ′ . (6)
Thus, we are interested in testing hypotheses H0: γk > γl versus H1: γk ≤ γl. For this test, we consider data y ∼ = { y 1 , … , y N k , y 1 ′ , … , y N l ′ } comprised of patient-specific observations of the form in Eq (2) and compute the Bayes factor
B 01 = Pr [ y ∼ | H 0 ] Pr [ y ∼ | H 1 ] = Pr [ H 0 | y ∼ ] / Pr [ H 0 ] Pr [ H 1 | y ∼ ] / Pr [ H 1 ] = Pr [ H 0 | y ∼ ] / Pr [ H 0 ] ( 1 - Pr [ H 0 | y ∼ ] ) / ( 1 - Pr [ H 0 ] ) (7)
to quantify the evidence in favor of the null hypothesis H0 and against the alternative H1 [34]. To compute Pr[H0|y∼], we apply the ergodic theorem and approximate the posterior probability by the fraction of MCMC samples satisfying γk > γl. The prior Pr[H0] is computed similarly except we sample onset times si for the two groups of patients being compared directly from the uniform prior distributions si ∼ Uniform(0, ai).
The methods outlined in this section are implemented by the Bayesian BE clock model. All necessary tools to employ this model via the Gibbs sampler are available in documented R code at https://github.com/yosoykit/BE_Clock_Model.
First, we used the Bayesian BE clock model to obtain posterior estimates of parameters for data set D2 (size N2 = 30 patients) with the BE clock set of 67 CpGs. See Materials and Methods for modeling details and CpG selection. Fig 6 depicts the wide inter-individual variability in the predicted BE onset ages among the 30 patients in D2, with interquartile and 95% credible intervals (CIs) denoted by box and whisker, respectively, for each Markov Chain Monte Carlo (MCMC) parameter estimate of BE onset age si, i = 1, .., N2. For these 30 patients, median MCMC estimates for BE onset ages ranged from 2.0 to 59.0 years of age, with a median of 33.6 years of age. The model also estimates CpG specific drift rates bi,j, j = 1, .., 67 for the BE clock set and a measurement standard deviation parameter, σBEi for each individual i (see Materials and Methods for details).
The BE onset age estimates for the patients in D2 were obtained using prior pb(bi,j) derived from data set DV (purple curve in Fig 5). We provide MCMC results when using this prior because 1) the estimates of BE onset times si, i = 1, …, N, using the DV prior are very similar to those when using the D1 prior, and 2) the DV prior introduces no bias (i.e., more realistic overall population drift distribution) because it was not used for the BE clock CpG marker set selection.
To quantify the aggregation of BE and EAC in families, Chak et al. performed a study with 411 patients with BE and/or its associated cancers, and found that familial BE (FBE) can be determined in 7.3% of patients, comprising 9.5% of EAC cases [22]. One hypothesis is that FBE patients have a stronger predisposition to develop BE compared to non-familial individuals, possibly due to inherited susceptability gene(s). We estimated the Bayesian BE clock model parameters for the independent data set D3 (size N3 = 22 patients) with FBE, with age range 39–84 at time of biopsy (mean age = 62.7). Fig 7 depicts the posterior median BE onset ages estimated for the 22 patients in D3, with interquartile and 95% credible intervals denoted by box and whisker, respectively. For these 22 patients, median MCMC estimates for BE onset ranged from 0 to 46.4 years of age, with a median of 26.1 years of age. The youngest FBE patient is shown to have onset at birth due to the incongruence of the standard clock drift rate distribution with his methylation values for the molecular clock set and thus we were unable to obtain positive posterior estimates of his onset age.
Because a younger age of disease onset is often considered a surrogate marker for a genetic or environmental predisposition, we tested the hypothesis that the FBE patients of data set D3 had been living with their BE for longer than the general BE patients in data set D2, which in our notation translates to H0: γ3 > γ2 (see Materials and Methods for details). The Bayes factor (see Eq (7)) was conservatively estimated to be 100K. This result provides decisive support in favor of the hypothesis that the FBE patients indeed harbored BE (relative to their ages when biopsies were removed for analysis) longer than the general BE population harbored BE (see left panel of Fig 8 for violin plot depicting this result).
With the BE onset predictions provided in the previous results, we are in a position to associate a patient-specific risk of developing EAC before a certain age. We computed the cumulative risk of developing EAC for each patient before age 88 (age of the oldest patient in our data sets) by using tissue age biomarker data to inform the modeling of the neoplastic progression to EAC. Such an integrated perspective for cancer risk management has recently been suggested by Li and colleagues [35]. To this end, we employ a mathematical model for EAC progression, termed the multistage clonal expansion for EAC (MSCE-EAC) model, that was previously calibrated to EAC incidence in the US by birth cohort, to obtain EAC risk estimates for each patient assuming that all patients share similar risk factors (e.g., unknown dysplasia status at time of biopsy) for EAC progression [8, 16]. Specifically, for each BE patient who has not been diagnosed with EAC by age a, given estimated BE onset time TBE = s, we computed the following risk
Pr [ T E A C < 88 | T B E = s , T E A C > a ] = S M S C E ( a - s ) - S M S C E ( 88 - s ) S M S C E ( a - s ) , (8)
where SMSCE is the EAC survival probability for the multistage clonal expansion (MSCE) model after BE initiation (see S1 Text for a derivation and S1 Fig for a model schematic) [8, 16, 36]. Alternatively, we may use summary (constant) risk estimates of progressing from non-dysplastic BE to EAC using published annual risk estimates across individuals of different age and different BE onsets. Note, however, for general s < a our mathematical EAC model implies the following inequality,
Pr [ T E A C < 88 | T B E = s , T E A C > a ] ≠ Pr [ T E A C < 88 | T B E < a < T E A C ] , (9)
which demonstrates that a patient’s BE onset adds information to refine blanket risk stratifications that do not consider this information.
As a demonstration, we used this model to compute the patient-specific risk of developing EAC by age 88 assuming a standardized 1950 birth cohort, allowing for gender-specific model parameters, by inputting the BE onset age estimate s for each patient into Eq (8). See S1 Table for the MCMC BE onset median estimates (with 95% credible intervals) of the 2 BE data set groups. Fig 8 shows the distributions of median MCMC estimated BE onsets for the 2 patient data sets (green violin plots) and their age-at-biopsy distributions (grey boxplots), alongside the corresponding EAC risk estimates for these onset ages. Of the two patient groups, the FBE patients in data set D3 have a significantly higher predicted median EAC risk estimate of 0.47 compared to the sporadic BE population with a median risk of 0.11. Because EAC risk is predicted by our model to increase monotonically with BE dwell time for patients of the same age, the correlation between estimated BE onset age and predicted EAC risk by age 88 is very high across patients (corr = .92 for data set D2, corr = .97 for data set D3, see S5 Fig).
A fundamental problem in predicting the risk of esophageal adenocarcinoma (EAC) in patients with BE continues to be the difficulty in assessing the neoplastic potential of BE, which is considered the premalignant field in which EAC arises. Several lines of evidence and theoretical considerations support the notion that both BE segment length and the duration of how long BE has been present in a patient (i.e., BE dwell time) are important determinants of EAC risk in addition to environmental and genetic risk factors [16, 37, 38]. While endoscopic surveillance with systematic biopsy sampling is the standard clinical care to screen BE patients for dysplasia and early cancer, most BE patients never develop esophageal cancer in their lifetimes. Priority has therefore been given to novel approaches to identify the molecular signatures of EAC progression and biomarkers in an attempt to more precisely define EAC risk at an individual level. However, because chronological age is recognized as one of the strongest predictors of cancer risk, renewed attention has been given to exploring the roles of biological tissue-age and cellular senescence in the progression to cancer [39].
Unfortunately, a clinical determination of when a patient first developed BE is presently not possible because BE is mainly asymptomatic (over 90% of EAC cases do not present with a prior history of BE [40]). For this reason we made an attempt to validate our BE onset predictions indirectly through two lines of evidence. First, we validated the longitudinal drift rates with an independent data set (DV). Although the drift rates for the BE clock set were generally lower in the validation set DV compared with the rates seen in set D1 (which we attribute to selection bias in D1), we found very similar estimates of the BE onsets using either drift-rate prior distribution in our Bayesian model. Secondly, we considered previous efforts to identify tissue-based indicators that accurately reflect the biological age of a tissue using regularized regression techniques by directly regressing age on the levels of methylation at a large number of CpGs to identify subsets of CpGs that are predictors of chronological age [13, 14]. Although we cannot use these techniques in this context because the BE onset times are unknown, we find that our predictions are at least broadly consistent with the straightforward application of these clock models to estimate absolute tissue-age differences between BE and SQ tissue. Specifically, using the published elastic net coefficients by Horvath [14] and by Hannum et al. [13] we computed the predicted biological age of the BE tissue and subtracted the predicted biological age of the normal squamous (SQ) esophageal tissue to arrive at crude estimates of the BE dwell time for the 30 patients in D2 (the cross-sectional cohort of patients). By subtracting these estimates from the chronological ages of the patients we obtained corresponding BE onset times that correlated well with our predictions (r = 0.77 for the Horvath 110 clock-CpG model, r = 0.84 for the 89 clock-CpG model by Hannum et al.).
Finally, we tested our clock model using methylation array data from 22 familial BE patients (set D3). Patients from both groups D2 and D3 have similar age distribution (see Fig 8 and S1 Table). However, compared to the onset ages estimated for the patients in data set D2, the familial group show increased BE dwell times; Bayes factor testing for the FBE study suggests that the inferred BE onset times, although heterogeneous (Fig 8), tend to occur significantly earlier in life for FBE patients compared to nonfamilial BE cases implying a possible heritable predisposition to develop BE metaplasia. Given that the predictions of BE onsets among FBE cases are significantly earlier than the predictions for the sporadic cases, it is perhaps surprising that the age distribution for the familial cases is not dissimilar to the age distribution for the sporadic cases (see grey boxplots in Fig 8). One possible explanation is that, next to symptomatic reflux, heartburn and other common risk factors, family history may not have been an indicator for referral to endoscopy as familiarity of this disease was only discovered in the past couple decades [22]. Therefore, if reflux frequency and other indicators for referral are similar for familial and non-familial patients, we expect the mean ages of BE diagnosis to be similar between the two groups. Specifically, we found the median estimates of BE onset age for the FBE patients to be 7.4 years earlier on average than the sporadic BE cases in study D2. This result is consistent with the result of a large study by Chak et al. that concluded that multiplex FBE families (multiplex being defined as having at least 2 confirmed FBE cases among family members) develop EAC at an earlier age compared with nonfamilial EAC cases [38]. Similar to the conclusions drawn by these authors, our result suggests that FBE patients may need earlier and possibly more frequent endoscopic screening for neoplastic lesions in BE tissue before EAC develops.
Given the theoretical implications of our proposed model of BE initiation and progression to EAC, we propose that once a patient’s BE onset has been estimated from his/her methylomic drift profile, his/her risk of developing EAC can be estimated more precisely. We have used a previously validated multistage clonal expansion model for EAC incidence which explicitly considers the uncertainty of the timing of BE onset in the general population and describes, conditional on when BE develops, the stochastic process of neoplastic progression from metaplastic to dysplastic tissue to cancer [8, 16]. These theoretical predictions show a strong dependence of EAC risk on the BE dwell time (see Fig 8). Importantly, we found that the lifetime risks for the individuals in study D2 vary widely, with an interquartile range of 0.01 to 0.44. It is important to recognize that these EAC risk predictions do not consider the effects of interventions and therefore may be overestimates. Although this predicted variability in risk stands unconfirmed, our median risk prediction of 0.11 for the D2 patients (see Fig 8) is consistent with empirical estimates of the EAC lifetime risk in BE patients found in the range 0.07–0.13 [41]. Therefore, the finding that the lifetime risks for the individuals in study D2 vary widely with an interquartile range of 0.01 to 0.44 translates into relative EAC risks (for the 4th quartile relative to 1st quartile) of > 40, assuming an otherwise homogenous population. For comparison, we found positive correlations between our D2 EAC risk predictions based on BE onset and D2 EAC risk estimates using previously reported risk factors based on gender (corr = 0.57, p = .001), histopathological grade (corr = 0.53, p = .003), and chronological age (corr = 0.49, p = .006) [6]. However, each of those risk factor estimates led to much lower relative EAC risks of <3. This suggests that BE onset, as determined by methylomic drift, can be considered a potential biomarker for EAC risk, although further validation via properly powered prospective studies or case-control studies in BE patients are needed to confirm this. Such studies may provide the requisite data to further test how well BE tissue-age performs in identifying individuals that likely progress to HGD or EAC in their lifetime so that endoscopic surveillance and available interventions can be utilized more effectively.
|
10.1371/journal.pgen.1004147 | Re-sequencing Expands Our Understanding of the Phenotypic Impact of Variants at GWAS Loci | Genome-wide association studies (GWAS) have identified >500 common variants associated with quantitative metabolic traits, but in aggregate such variants explain at most 20–30% of the heritable component of population variation in these traits. To further investigate the impact of genotypic variation on metabolic traits, we conducted re-sequencing studies in >6,000 members of a Finnish population cohort (The Northern Finland Birth Cohort of 1966 [NFBC]) and a type 2 diabetes case-control sample (The Finland-United States Investigation of NIDDM Genetics [FUSION] study). By sequencing the coding sequence and 5′ and 3′ untranslated regions of 78 genes at 17 GWAS loci associated with one or more of six metabolic traits (serum levels of fasting HDL-C, LDL-C, total cholesterol, triglycerides, plasma glucose, and insulin), and conducting both single-variant and gene-level association tests, we obtained a more complete understanding of phenotype-genotype associations at eight of these loci. At all eight of these loci, the identification of new associations provides significant evidence for multiple genetic signals to one or more phenotypes, and at two loci, in the genes ABCA1 and CETP, we found significant gene-level evidence of association to non-synonymous variants with MAF<1%. Additionally, two potentially deleterious variants that demonstrated significant associations (rs138726309, a missense variant in G6PC2, and rs28933094, a missense variant in LIPC) were considerably more common in these Finnish samples than in European reference populations, supporting our prior hypothesis that deleterious variants could attain high frequencies in this isolated population, likely due to the effects of population bottlenecks. Our results highlight the value of large, well-phenotyped samples for rare-variant association analysis, and the challenge of evaluating the phenotypic impact of such variants.
| Abnormal serum levels of various metabolites, including measures relevant to cholesterol, other fats, and sugars, are known to be risk factors for cardiovascular disease and type 2 diabetes. Identification of the genes that play a role in generating such abnormalities could advance the development of new treatment and prevention strategies for these disorders. Investigations of common genetic variants carried out in large sets of research subjects have successfully pinpointed such genes within many regions of the human genome. However, these studies often have not led to the identification of the specific genetic variations affecting metabolic traits. To attempt to detect such causal variations, we sequenced genes in 17 genomic regions implicated in metabolic traits in >6,000 people from Finland. By conducting statistical analyses relating specific variations (individually and grouped by gene) to the measures for these metabolic traits observed in the study subjects, we added to our understanding of how genotypes affect these traits. Our findings support a long-held hypothesis that the unique history of the Finnish population provides important advantages for analyzing the relationship between genetic variations and biomedically important traits.
| Genome-wide association studies (GWAS) based on common single nucleotide polymorphisms (SNPs) have unequivocally demonstrated the contribution of thousands of loci to risk for common diseases and to variation in quantitative traits. However for most such complex phenotypes, the variants identified to date appear to explain only a fraction of heritable variation, suggesting an important role for variants not assessed in GWAS. In particular, the hypothesis that currently unidentified low-frequency genetic variants may have a major impact on complex phenotypes has stimulated extensive efforts to discover such variants through next-generation sequencing.
Over the next several years it will increasingly become feasible to conduct comprehensive variant discovery through exome or whole genome re-sequencing studies. Such studies have the potential to demonstrate the impact on complex phenotypes of genes, pathways, and networks that GWAS have not yet implicated in these phenotypes. However it is increasingly clear that identifying associations at genome-wide or exome-wide thresholds of statistical significance will require large samples, and thus these experiments remain very costly. Although targeted re-sequencing studies of large samples do not provide the same likelihood of implicating novel genes as do genome-wide or exome-wide sequencing, they offer an excellent opportunity to obtain an initial picture of the relative phenotypic impact of variants across the complete allele frequency spectrum, in regions of interest. Such studies require evaluation of a relatively limited number of variants and, if prior evidence indicates that variants within the targeted region contribute to the phenotype, require a less stringent statistical threshold.
Genes within loci for which GWAS have shown significant associations represent logical foci for investigations across the allelic frequency spectrum. Several genes are now known to harbor both rare variants responsible for Mendelian disorders and common variants associated with related phenotypes [1], [2]. Resequencing of such genes may suggest particular variants as contributors to the GWAS signal, and may identify variants whose association with the phenotype is independent of the GWAS signal. Together, such variants provide starting points to investigate the heritable component of biological processes underlying the associated phenotypes.
We therefore undertook a re-sequencing study of Finnish cohorts, targeting loci identified from GWAS of quantitative metabolic traits, including: fasting blood levels of lipids and lipoproteins (triglycerides, TG; high-density lipoprotein cholesterol, HDL-C; low-density lipoprotein cholesterol, LDL-C; and total cholesterol, TC), glucose (FG), and insulin (FI). Several of these traits (TG, HDL-C, and FG) are components of the metabolic syndrome, an aggregation of variables that increase risk for type 2 diabetes (T2D) and cardiovascular diseases [3]. We report here the results of such targeted re-sequencing of >6,000 individuals drawn from a population cohort (the 1966 Northern Finland Birth Cohort, NFBC; [4]) and a T2D case-control sample (the Finland-United States Investigation of NIDDM Genetics study, FUSION; [5], [6], which included 919 individuals with T2D and 919 normal glucose-tolerant controls). In these individuals, we sequenced the coding regions of 78 genes selected from 17 loci that showed genome-wide significant association to one or more of the designated quantitative metabolic traits in GWAS meta-analyses that included these studies [7], [8]. Details on how we selected loci and genes within loci for re-sequencing can be found in Text S1.
We focused on these Finnish cohorts for two reasons, both of which concern the relationships expected between population history and the distribution of rare variants within a study sample. First, when a founder population has expanded recently from severe bottlenecks, as in Finland, many variants may disappear from the population while others increase rapidly in frequency owing to subsampling and genetic drift. Thus, while the overall number of rare variant sites observed in sequencing studies of the Finnish population is smaller than in other European populations [9], some deleterious variants are observed at a much higher frequency in Finland than in other populations. These variants include the mutations responsible for about 40 rare Mendelian disorders, the so-called “Finnish disease heritage” [10], [11]. We hypothesized that some variants with a large effect on quantitative metabolic phenotypes would also have attained a relatively high frequency in the Finnish population, so that by re-sequencing Finnish samples we could identify novel associations that might be unfeasible to detect in comparably sized samples from most other populations.
Second, the availability of information specifying the birthplace of most members of the NFBC and FUSION cohorts (or their parents) addresses the recently raised concern that unidentified population substructure may pose a particular issue in association analyses of rare variants (e.g. those with frequency <1%) [12]. This concern reflects the expectation that such variants have generally arisen more recently than common variants and are therefore more likely to differ in frequency between study populations; this concern is mainly relevant in studies where the geographical origin of the subjects is unknown [12]. Indeed, previous studies in Finnish samples (including NFBC) have shown that the available birthplace data provide a highly accurate delineation of population substructure [7], [10].
Principal components analysis (PCA) using 122 k SNPs typed on genome-wide arrays revealed that the NFBC and FUSION samples overlap broadly in the first two PC dimensions (Figure S1). Phenotype distributions also overlap considerably between the cohorts (Table S1), and comparison of mean residual values after regressing the combined sample on age, age2, and sex showed no significant differences between NFBC and FUSION for any phenotype (p>0.77 for all comparisons; see Text S1), after excluding T2D cases from analysis of FG and FI.
We selected for re-sequencing the protein-coding regions and 5′ and 3′ untranslated regions (UTRs) of the genes within 17 loci that had previously demonstrated significant association (p<5×10−8) in GWAS to one or more metabolic phenotype (Table 1); TG (eight loci), HDL-C (nine loci), LDL-C (six loci), TC (nine loci), FG (six loci), and FI (one locus) [7], [8], [13]–[15]. The selection of the loci depended on the evidence from meta-analyses of several independent studies, but for eight of them, NFBC alone showed genome-wide significant association to one or more of the six phenotypes. We defined loci as the regions bracketed by the nearest recombination hotspots (>10 cM/Mb) on both sides of the reported GWAS SNPs. The numbers of genes included in the GWAS loci so defined ranged from one (four loci) to 50 (the MADD locus). As we did not have the resources to sequence all possible genes at each locus, we sequenced the genes nearest to the SNPs that showed genome-wide significant association with these phenotypes (see Text S1 for more detail), for a total of ∼270 kb of sequence.
We conducted targeted Illumina sequencing using 150 bp probes designed to capture primarily coding sequence, in whole-genome amplified (WGA) DNA from 6,958 individuals; 6,123 of these individuals (4,447 NFBC, 836 FUSION normal glucose tolerant controls, and 840 FUSION T2D cases) passed quality control procedures (Text S1). Mean depth of coverage (per bp per person) per gene ranged from 31×–285× (Table S2, Figure S2, and Text S1). On average, 96% of sequenced base pairs within a gene had genotype quality score ≥50 in ≥75% of subjects; some genes were covered at this level for as few as 60% of base pairs (Table S2). After this initial quality control process, we identified 2,221 variant sites, 1,779 (80%) with MAF<1%.
It is difficult to distinguish between low count variants and sequencing artifacts, and we reasoned that such artifacts might be increased in our study given that all DNAs had been whole-genome amplified (WGA). We therefore attempted to validate low count variants by PCR-amplification of the putative variant site in genomic DNA from variant carriers (or WGA DNA if genomic DNA was not available) and sequencing using a different platform (Roche 454 FLX). We sequenced all variants identified in ≤3 individuals in our sample and not reported in dbSNP version 135 (N = 1,104, Text S1), and considered validation for the sites as (1) their being variable and (2) the specific non-reference genotypes being correct as called.
Overall, we validated 89.5% of these 1,104 sites including 100% of the 91 sites with variants present three times and 271 of 273 (99.3%) corresponding non-reference genotypes; 205 of 207 (99.5%) of the 207 sites with variants present twice and 397 of 414 (95.9%) corresponding non-reference genotypes. Among singletons, we validated 691 of 806 (85.7%) non-reference genotypes; however, 336 of these validated only in WGA DNA (the only DNA source available for these samples). Conservatively, we excluded from further analyses these 336 WGA-only singleton sites, along with 104 singleton sites that were refuted (49 sites), not covered (20 sites), or found to be WGA artifacts (35 sites). Eleven additional singleton sites were found to be homozygous alternative when validated, bringing the number of retained singleton sites to 366 and the total number of retained sites (among the 1,104 for which validation was attempted) to 663.
After validation, we included a total of 1,780 variable sites for further analysis. The subsequently released dbSNP version 137 included 76 of our non-validated sites: our experiments had directly refuted four of these sites, we had not adequately covered five of them, and we had validated 67 sites only in WGA DNA. We re-included the 72 non-refuted sites, bringing the total number of validated polymorphic sites for subsequent analysis to 1,852 (Table S3).
To quantify the increase in rare variation information provided by sequencing compared with genotyping, we calculated the overlap between variants found in this study and those observed in a larger Finnish sample: 9,660 Finnish participants from the population-based Metabolic Syndrome in Men (METSIM) study [16] who were genotyped with the Illumina ExomeChip. The ExomeChip captured only 346 (19%) of the 1,852 polymorphic sites that we identified through sequencing.
The majority of sequence variants (1,114, 60%) were in coding sequence (37% non-synonymous [NS] and 23% synonymous) while 738 (40%) were in introns or UTRs (Figure S3). PolyPhen2 [17] predicted 236 variants to have a deleterious impact: 213 missense “probably damaging” and 23 nonsense variants. Of these 236 variants, 21 (19 missense and two nonsense) were present in homozygous form in at least one individual. For all 21 of these variants, the phenotype distributions for rare-allele homozygotes overlapped with the phenotype distributions of the common-allele homozygotes (Figure S4), suggesting these variants are not sufficient to cause extreme phenotypes. A total of 1,410 of the 1,852 validated variants (76%) had MAF<1%, including 486 (26%) singleton and 217 (12%) doubleton variants (Figure S5). Nucleotide diversity, as estimated by Watterson's measure θW = 7.1×10−4 was larger than the pair wise heterozygosity estimator θπ = 3.5×10−4, reflecting the abundance of singleton sites.
We observed less overall variation than that seen in earlier sequencing studies of individuals of European descent; one variant site in every 147 bp sequenced, as compared to every 21 bp [9], 57 bp [18] or 83 bp [19]. While the sample size in the study of Nelson et al. [9] was larger (12,514 European Americans) than that of our study, the sample sizes in Tennessen et al. [19] and Fu et al. [18] were smaller (1,351 and 4,298 European Americans, respectively; note that the samples sequenced in the latter two studies represented two different data releases from the same dataset). Nelson et al. observed that in the Finnish samples in their study, the number of variant sites per kb of sequence, was about one-third that of similar sized samples from southern Europe. Thus, while differences in sequencing coverage and in the number of sequencing artifacts could partially account for our observation of reduced numbers of variant sites compared to other studies, the results of Nelson et al. suggest that the Finnish population bottleneck may have played a larger role.
The reduced variation observed in our study compared to the three previous studies, primarily reflects numbers of rare variants. Nelson et al. report that 95% of their variant sites were rare (MAF<0.5%), with 74% seen in only one or two copies. Similarly, Tennessen et al. report that 72% of variant sites were seen in ≤3 copies. In our study, by contrast, 72% of variants were rare, 38% were seen in one or two copies, and 44% were seen in ≤3 copies.
By down-sampling our data [20] to match the sample sizes of Tennessen et al. and Fu et al., and down sampling the data of Nelson et al. to match our sample size, we directly compared our site-frequency spectra (SFS) with those observed in these three studies. We caution against over-interpretation of these SFS, as they can be impacted by differences between studies in the choice of genes sequenced, variant ascertainment, and coverage. Nevertheless, in our sample, a substantially lower percentage of coding variants have MAF<1% than in any of the other three studies (Table S4). Conversely, in our sample we observe a higher proportion of so called “Goldilocks alleles”: variants with MAF 0.5–2%, a frequency sufficient for single-variant analyses of potentially large-effect variants [21]. For example, while Nelson et al. report that 1.1% of NS variants are Goldilocks alleles, we observe that 7.4% of NS variants fall in this frequency range.
While we observe fewer rare variants than these other sequencing studies, the proportion of NS variants among rare coding variants in our study (65%; 95% CI = 62%–68%) is similar to that seen in Nelson et al. (63%). The proportion of rare variants predicted to be functional is also roughly similar between our study and other studies. For example, Tennessen et al. report that almost 96% of SNVs predicted to be functional have MAF<0.5%, and state an odds ratio of 4.2 that such rare variants are functional compared to variants with MAF>0.5%. We find that 89% of SNVs predicted to be functional are rare, and estimate an odds ratio of 3 (95% CI = 1.98–4.52).
A total of 39 unique locus-phenotype combinations represent the previously reported associations between the 17 re-sequenced loci and one or more of the six metabolic phenotypes: 32 associations for lipid measures, six for fasting glucose, and one for insulin (Table 1). To follow up these previous findings, we conducted association tests on the combined NFBC/FUSION data (see Methods). We conducted single-variant tests (regression of phenotype residuals on an additively coded genotype, see Methods) to assess association in each of the 39 locus-phenotype sets for all validated variants with MAF>0.1%; tests under alternative genetic models did not reveal any additional association evidence. Since multiple independent association signals may be present at a locus, we evaluated the relation of each newly associated variant to the “array SNP,” the SNP genotyped in the combined NFBC/FUSION sample with smallest p-value in this sample in single-SNP association tests (Table 1). We then conducted single-variant analyses conditional on the array SNP, by including the array SNP genotype as a covariate in the linear regression.
We used gene-level tests to evaluate the collective impact of non-synonymous (NS) variants with MAF<1% for each of the 62 genes that harbored at least two such validated variants, considering only phenotypes that showed prior evidence of association to the locus (a total of 147 tests). We adopted this MAF threshold after determining that any higher MAF threshold simply recapitulated associations identified by the single-variant tests. Given different alternative models of interest, we performed two minimally correlated tests: CMC [22] which assumes the direction of effect for all rare variants is the same, and SKAT [23] which is better tuned to the setting in which the direction of effect of rare variants is mixed.
Taking the combined results from our single-variant and gene-level analyses, we evaluated to what degree re-sequencing of these 17 loci has advanced our understanding, beyond what was known from GWAS, of the phenotypic impact of genetic variation. We considered such an advancement to consist of either identification of additional, independent association signals, or the detection of association to rare variants.
For several of the lipid-associated loci, we were able to assess the evidence for multiple independent signals in relation to a similar analysis conducted on SNP data by Teslovich et al. 2010 [8]. This comparison has two limitations: our study and that of Teslovich et al. did not examine the same set of variants, and for five of the 13 lipid loci, our variant set did not contain a good proxy (r2>0.8) for the lead SNP of Teslovich et al. To counter these limitations, we used information imputed from NFBC data on pairwise LD between variants analyzed in the two studies, and assumed that any pair of variants with r2<0.2 in NFBC were effectively independent.
We used here a significance threshold of p<0.001 (approximately the cutoff obtained by applying the Benjamini-Hochberg [24] rule to control FDR at the 0.02 level across all the variants/genes and phenotypes tested, see Methods). For 27 of the 39 locus-phenotype combinations, the re-sequencing analysis essentially recapitulated the results from the GWAS. For the remaining 12 locus-phenotype combinations (at seven loci), we summarize below how re-sequencing has advanced our understanding of genotype-phenotype relationships; MAF, p-values, and annotations for all associated variants at these seven loci are presented in Table 2.
Large-scale re-sequencing has the potential to identify a comprehensive set of variants that are missed by imputation and chip based fine-mapping approaches. In more than 6,000 members of Finnish cohorts assessed for metabolic traits, we re-sequenced 78 genes implicated in prior GWAS of these traits, identifying 1,852 total variants, including >200 predicted-deleterious missense variants and 23 nonsense variants, 125 of which are not currently in the public database (dbSNP 137). Using single-variant analyses, we found associations at seven loci (six involving one or more variants with MAF<5%, Table 2) and demonstrated using conditional analyses that these signals are independent of previously reported GWAS SNPs. Using gene-level tests we found compelling association evidence for rare variants in two genes, ABCA1 and CETP. By comparison, Hunt et al. [26] in a large (>40,00 individuals) autoimmune disease case-control sample, found that targeted coding region re-sequencing of 25 GWAS risk genes provided minimal new information. Several differences between our studies could account for the apparent discrepancies in findings: First, the genetic architecture of quantitative metabolic traits may be simpler than that of the diseases investigated by Hunt et al. Second, we benefitted from the effect of Finnish population history, which has led to a larger proportion of variants in the Goldilocks allele range and a smaller proportion of rare variants (about 70% of the variants observed by Hunt et al. are present in one or two copies, compared to <40% in our study). Third, the genes for which we identify rare variant associations may be unusual in their tolerance for functional variation.
Our gene-level test results for ABCA1 agree with two previous lines of evidence that rare variants in this gene could have an impact on lipid phenotypes. First, recessive mutations in ABCA1 cause extreme reduction in HDL-C, termed Tangier Disease or hypoalphalipoproteinemia; several of these variants were discovered in Finnish families [27]. Second, previous studies in diverse populations found enrichment of NS ABCA1 variants in individuals with low HDL-C levels [21], [28]. Among the fifteen previously described rare NS variants observed in our data, ten have previously been implicated in metabolic phenotypes: Tangier Disease (n = 3), increased risk for heart disease (n = 2), or either reduced (n = 3) or elevated (n = 2) serum HDL-C levels (Human Gene Mutation Database).
Our results have enabled us to clarify genotype-phenotype relationships for eight of the 17 loci examined. By delineating multiple distinct association signals, and in some instances highlighting specific candidate alleles, they also suggest potential targets for functional investigations that could specify causal variants. For example, at G6PC2 we identified a Goldilocks allele at rs13872630 which has a predicted deleterious effect. This variant has a distinct signal from the array SNP, and appears to have a much stronger effect in lowering FG. As this effect may provide protection against cardiovascular disease [29], there may be great value in generating mice mutated for this His177Tyr missense variant, which occurs at a highly conserved site. Additionally, the relatively high frequency of this variant within Finland offers an unusual opportunity to evaluate its impact on a much wider range of phenotypes than we investigated here.
At the same time these findings also point to the difficulty in predicting the phenotypic impact of individual variants. Recessive mutations in several of the genes that we re-sequenced are causative for rare metabolic disorders (e.g. [27]). However the relatively modest effect on quantitative metabolic phenotypes that we observed for variants in these and other genes predicted to be deleterious (nonsense and missense) suggest two possibilities: 1) the genetic and/or environmental backgrounds in families demonstrating Mendelian metabolic disorders may differ from the backgrounds in individuals drawn for population samples, and 2) we must be cautious in assigning likely causality to variants on the basis of annotation alone.
The incomplete coverage obtained for several loci provides an additional reason for caution in our conclusions. Methods for capturing a targeted region have become more efficient since we completed our study, and therefore it is possible that implementation of such methods would provide more complete coverage at these loci and could identify additional novel variants with a large contributions to metabolic phenotypes.
Our prior hypothesis was that the process of genetic drift within a recently expanded founder population such as Finland should elevate the frequency of some deleterious alleles so that, even if they are subject to strong selective pressure, they may be observed at relatively high frequency [11]. In such populations, these variants may be sufficiently common for phenotype-associations to be detected using single-variant tests. As predicted by this hypothesis, our re-sequencing identified, in G6PC2 and LIPC, two missense variants predicted to be deleterious that are very rare outside Finland (MAF<0.002), but that were sufficiently increased in frequency (MAF>0.013) in our study sample for us to detect significant association in single-variant tests. A recent genome-wide survey of copy number variations has similarly demonstrated that a rare deletion, highly over-represented within Finland, is associated with neurodevelopmental disorders [30]. Taken together, these results suggest that exome-wide and genome-wide investigations of Finnish population cohorts will likely identify additional associations to complex phenotypes that may not be apparent in other populations.
We obtained genomic DNA samples processed at the Finnish Institute of Molecular Medicine (NFBC) and US National Human Genome Research Institute (FUSION). All NFBC and FUSION participants included in this study provided informed consent. The studies were carried out in accordance with the approvals of the Ethical Committee of the Northern Ostrobothnia Hospital District (for NFBC), and the University of Michigan Health Sciences and Behavioral Sciences Institutional Review Board (IRB-HSBS) and the Institutional Review Board of the National Public Health Institute (KTL; now part of the National Institute for Health and Welfare, THL) (for FUSION).
We constructed Illumina multiplexed libraries with 5 µg of whole genome amplified material (see Text S1 for description of amplification procedures) or 1 µg native genomic DNA according to the manufacturer's protocol (Illumina Inc, San Diego, CA) with the following modifications: 1) DNA was fragmented using a Covaris E220 DNA Sonicator (Covaris, Inc. Woburn, MA) to between 100 and 400 bp. 2) Illumina adapter-ligated library fragments were amplified in four 50 µL PCR reactions for eighteen cycles. 3) Solid Phase Reversible Immobilization bead cleanup was used for enzymatic purification throughout the library construction process and for final library size selection targeting 300–500 bp fragments. Samples were multiplexed using Illumina barcoded libraries pooled together in pools of 12 or 18 depending on the sequencing platform. We designed a custom targeted set of 150 bp probes (Agilent Technologies, Santa Clara, CA) and captured ∼270 kb of primarily coding sequence from 78 genes. The concentration of each captured library pool was determined through qPCR according to the manufacturer's protocol (Kapa Biosystems, Inc, Woburn, MA) to produce cluster counts appropriate for the Illumina GAIIx and HiSeq 2000 platforms.
Sample pools of 12 and 18 were loaded on GAIIx and HiSeq machines, respectively, using paired end 101 bp read lengths. We aimed to achieve a coverage metric of 80% of the targeted space covered at ≥20× depth. We aligned reads from each sample to the NCBI37/hg19 reference sequence using BWA [31]. Sample identity was confirmed by comparing sequence data (SAMtools consensus calls) with pre-existing genotype array data. Individuals with ≥70% coverage at 20× and ≥90% genotype concordance with 51 array SNPs were included in the analysis (6,123 of 6,958 individuals).
Details on sequencing and generation of center-specific genotype call sets can be found in Text S1. To generate a consensus call set, we pooled together all quality controlled sites discovered by any of the three centers (UCLA, University of Michigan, or Washington University) in the defined target loci (number of markers m = 2,306). We excluded multi-allelic sites or sites with different alternative alleles (m = 72). Each center then re-called SNP genotypes at the remaining sites (m = 2,234). Majority vote was used to generate variant calls. Genotypes concordant between at least two centers were included in the consensus data set; others were set to missing. The overall concordance rate between centers was 99.96% (99.99%, 99.94%, and 99.95% for homozygous reference, heterozygous, and homozygous alternative genotypes, respectively).
NFBC individuals were previously genotyped on the Illumina 370duo Chip, and all FUSION cases and 774 of 919 FUSION controls on the Illumina HumanHap300 BeadChip (version 1.0). After standard quality control procedures [6], [7], high-quality GWAS genotypes were available for 296,978 SNPs for all genotyped individuals. We used PLINK [32] to identify 122,644 SNPs with no more than moderate pair wise linkage disequilibrium (r2<0.5) which we used to calculate genetic principal components (PCs) with EIGENSTRAT [33].
|
10.1371/journal.pgen.1003856 | A Novel Intra-U1 snRNP Cross-Regulation Mechanism: Alternative Splicing Switch Links U1C and U1-70K Expression | The U1 small nuclear ribonucleoprotein (snRNP)-specific U1C protein participates in 5′ splice site recognition and regulation of pre-mRNA splicing. Based on an RNA-Seq analysis in HeLa cells after U1C knockdown, we found a conserved, intra-U1 snRNP cross-regulation that links U1C and U1-70K expression through alternative splicing and U1 snRNP assembly. To investigate the underlying regulatory mechanism, we combined mutational minigene analysis, in vivo splice-site blocking by antisense morpholinos, and in vitro binding experiments. Alternative splicing of U1-70K pre-mRNA creates the normal (exons 7–8) and a non-productive mRNA isoform, whose balance is determined by U1C protein levels. The non-productive isoform is generated through a U1C-dependent alternative 3′ splice site, which requires an adjacent cluster of regulatory 5′ splice sites and binding of intact U1 snRNPs. As a result of nonsense-mediated decay (NMD) of the non-productive isoform, U1-70K mRNA and protein levels are down-regulated, and U1C incorporation into the U1 snRNP is impaired. U1-70K/U1C-deficient particles are assembled, shifting the alternative splicing balance back towards productive U1-70K splicing, and restoring assembly of intact U1 snRNPs. Taken together, we established a novel feedback regulation that controls U1-70K/U1C homeostasis and ensures correct U1 snRNP assembly and function.
| The accurate removal of intervening sequences (introns) from precursor messenger RNAs (pre-mRNAs) represents an essential step in the expression of most eukaryotic protein-coding genes. Alternative splicing can create from a single primary transcript various mature mRNAs with diverse, sometimes even antagonistic, biological functions. Many human diseases are based on alternative-splicing defects, and most interestingly, certain defects are caused by mutations in general splicing factors that participate in each splicing event. To address the question of how a general splicing factor can regulate alternative splicing events, here we investigated the regulatory role of the U1C protein, a specific component of the U1 small nuclear ribonucleoprotein (snRNP) and important in initial 5′ splice site recognition. Our RNA-Seq analysis demonstrated that U1C affects more than 300 cases of alternative splicing in the human system. One U1C target, U1-70K, appeared to be particularly interesting, because both protein products are components of the U1 snRNP and functionally depend on each other. Analyzing the mechanistic basis of this intra-U1 snRNP cross-regulation, we discovered a U1C-dependent alternative splicing switch in the U1-70K pre-mRNA that regulates U1-70K expression. In sum, this feedback loop controls and links U1C and U1-70K homeostasis to guarantee correct U1 snRNP assembly and function.
| In eukaryotes accurate splicing is an essential step in gene expression, because most protein-coding genes contain introns, which have to be removed from the precursor messenger RNA (pre-mRNA) to join the exons to a continuous open-reading-frame. In alternative splicing it is the balance between accuracy and flexibility of splice site recognition that creates from a single transcript multiple isoforms with diverse, sometimes even antagonistic, biological functions [1], [2]. Splice site selection depends on multiple parameters, such as splice site strength, RNA secondary structures, and transcription kinetics, and is modulated by trans-acting splicing regulators that can act as activators or repressors.
In both constitutive and alternative splicing, intron removal is catalyzed by the spliceosome, a macromolecular RNA-protein complex that comprises five small nuclear ribonucleoprotein particles (snRNPs) and numerous non-snRNP proteins [3]. Spliceosome assembly is a highly coordinated process characterized by a dynamic RNA-protein network. It is initiated by the recognition of the 5′ splice site by the U1 snRNP, however U1 snRNA:5′ splice site base-pairing alone is not sufficient. This interaction is further stabilized by both U1 snRNP components and non-snRNP factors that contribute to 5′ splice site selection and spliceosome assembly [4]–[6]. In addition to the snRNA, the U1 snRNP contains the Sm protein heptamer and three specific proteins: U1-70K, U1A, and U1C. Besides their role in splicing, both U1A and U1-70K bind directly to the poly(A) polymerase and are thereby involved in U1 snRNP-dependent inhibition of polyadenylation [7]–[9]. This includes the auto-regulation of U1A expression by inhibiting 3′-end processing of its own mRNA [10], [11]. U1-70K and U1C functionally depend on each other: First, the presence of U1-70K is a prerequisite for the stable incorporation of U1C into the U1 snRNP [12], [13]; second, the interaction of U1-70K with SRSF1 (ASF/SF2) stimulates U1 snRNP binding to the 5′ splice site only if U1C is present [14], [15]. Hence, the two proteins strongly rely on each other to ensure correct 5′ splice site recognition by the U1 snRNP.
The U1C protein is particularly important for correct 5′ splice site recognition: Mutational analysis in yeast revealed that U1C is essential for pre-mRNA splicing in vivo [16], and U1C stimulates the formation of early splicing complexes by stabilizing the U1 snRNA:5′ splice site duplex [17]–[19]. Consistent with this, structural analyses of the U1 snRNP located U1C in close proximity to the 5′ end of the U1 snRNA and revealed that U1C directly contacts the minor groove of the snRNA:mRNA duplex [20]. Moreover, several studies indicate that U1C participates directly in 5′ splice site choice: Du and Rosbash [21] demonstrated that recombinant yeast U1C protein binds to 5′ splice site consensus sequences independently of the U1 snRNP; moreover, we have recently shown that U1C regulates a subset of alternatively spliced 5′ splice sites in the zebrafish [22].
Here we investigate the role of U1C as an alternative splicing regulator in the human system. Based on an RNA-Seq analysis in HeLa cells after siRNA-mediated knockdown of U1C, we identified a distinct group of target genes with specific U1C-dependent alterations in their splicing patterns. We focus on a particularly interesting target, U1-70K, because these two proteins coexist within the same snRNP and strongly depend on each other (see above). We discovered a conserved, intra-U1 snRNP cross-regulation, the mechanistic basis of which was further investigated, combining mutational minigene analysis, in vivo splice-site blocking by antisense morpholinos, and in vitro binding experiments. This revealed that recognition of an alternative, U1C-dependent 3′ splice site within intron 7 of the U1-70K pre-mRNA requires binding of intact, U1C-containing U1 snRNPs to downstream cryptic 5′ splice sites. Importantly, this mechanism describes a novel feedback-loop to control U1-70K and U1C homeostasis, linking the expression of these two U1 snRNP-specific factors via alternative splicing.
To investigate whether U1C plays a splicing-regulatory role in the human system, we performed siRNA-mediated knockdown of U1C in HeLa cells and analyzed alternative splicing patterns by high-throughput RNA sequencing (RNA-Seq). Western blot analysis of whole-cell lysates confirmed that U1C protein is no longer detectable after three days of knockdown in comparison to the control-treated cells (Figure 1A). Importantly, as shown by Northern blot analysis of total RNA, U1 snRNA steady-state levels were not affected under U1C knockdown conditions (Figure 1A). In addition, affinity purification of U1 snRNPs from both control- and U1C-knockdown cells demonstrated that the U1C-deficient particles are fully stable (Figure S1A). In vitro binding assays with substrates containing functional 5′ splice sites further showed that the lack of U1C slightly reduces, but does not abolish U1 snRNP binding efficiency (Figure S1B).
Deep-sequencing of poly(A)+-selected RNA from control- and U1C-siRNA treated HeLa cells yielded 56.9 and 52.0 million 105-bp single-end sequence reads, respectively. 56% for the control sample and 69% for the knockdown could be uniquely mapped to the human genome and annotated splice junctions. Approximately 30% (control: 32%; knockdown: 28%) of the uniquely mapped reads span a splice junction. We applied a data analysis procedure described previously [22] to predict U1C-dependent alternative splicing targets, resulting in these two major alternative splicing changes (summarized in Figure 1B):
First, cassette-type exons of which 169 targets were detected with increased exon skipping, and 37 targets with increased exon inclusion upon U1C knockdown (see Tables S1 and S2). Second, we found 111 targets with alternative 5′ splice sites, where usage of the proximal (downstream) site increased upon U1C knockdown, and 12 targets with increased distal (upstream) site usage (see Tables S3 and S4). In addition, there were only 34 cases of alternative 3′ splice sites, 17 of them each with increased usage of the proximal or the distal splice site.
A total of 33 predicted targets with increased exon skipping or increased usage of proximal 5′ splice sites were randomly selected for validation by semi-quantitative RT-PCR: We were able to positively validate 17 out of 19 exon skipping events and 11 out of 14 cases of alternative 5′ splice site usage, corresponding to a general validation rate of ∼85% (Figure S1C and D).
To further control for the U1C specificity of these alternative splicing changes, we also blocked base pairing between the U1 snRNA and the pre-mRNA 5′ splice site, using an antisense morpholino oligonucleotide (AMO) directed against the 5′ end of the U1 snRNA [23]. Efficient morpholino blocking of the U1 snRNA was confirmed by RNase H protection, using an antisense DNA oligomer binding to the 5′ end of the U1 snRNA and silver staining (Figure 1C). Total RNA was then subjected to RT-PCR analysis, using the same target-specific primers used to validate U1C-dependent alternative splicing changes. Figure 1D shows six selected targets, four for exon skipping and two for alternative 5′ splice site choice: Knockdown of U1C resulted in increased exon skipping in SNHG5 (exon 4), KCNAB2 (exon 3), URB2 (exon 6), and CARM1 (exon 15); for other targets, here exemplified by MARCH7 (exon 7) and UFM1 (exon 2), an alternative (proximal) 5′ splice site became activated in the absence of U1C. In contrast, the isoform ratio did not significantly change after AMO blocking of U1 snRNA base pairing (Figure 1D, compare lanes U1C kd and U1 snRNA blocking). We note that AMO blocking generally reduced mRNA levels, most likely due to a general splicing block by the AMO treatment. In sum, this direct comparison of the splicing patterns after U1C depletion and after AMO-directed U1 blocking confirmed the U1C specificity of the effects observed.
We conclude that most of U1C-dependent alternative splicing changes fall into two classes, cassette-type exons and alternative 5′ splice site usage. There is a striking bias towards increased exon skipping (169 versus 37), followed by distal-to-proximal 5′ splice site shifts (111 versus 12; Figure 1B); therefore U1C appears to play primarily an activating role in 5′ splice site recognition.
Among the most interesting targets of U1C-dependent alternative splicing we identified the U1-70K pre-mRNA. As described in the Introduction, these two proteins, U1C and U1-70K, are both specific components of the U1 snRNP, interact with each other in the U1 snRNP, and are important for its function in pre-mRNA processing.
Figure 2A shows the distribution of read coverage along the U1-70K pre-mRNA (NM_003089) obtained by RNA-Seq analysis of control- and U1C-knockdown HeLa cells. In the wildtype situation (control), we see a significant accumulation of sequence reads in intron 7, starting at position +643 and extending up to the 3′ splice site of exon 8 (Figure 2A and B). The RNA-Seq data analysis revealed an alternative 3′ splice site at the position where the intron reads start to accumulate, which is frequently used in comparison to normal, productive exons 7–8 splicing, but strongly depends on U1C (Figure 2C; 107 versus 4 junction reads for control and ΔU1C, respectively). Usage of this alternative 3′ splice site introduces a premature termination codon (PTC) into the U1-70K mRNA. Conversely, normal exons 7–8 splicing strongly increases after U1C depletion (718 versus 385 junction reads for ΔU1C and control, respectively). We noted several cryptic 5′ splice sites located closely downstream of the alternative 3′ splice site (referred to as A, B, C), as well as one more further downstream (site D; Figure 2C). However, usage of these 5′ splice sites, that means inclusion of the alternative “exon 7a”, is not significant, under both normal and U1C-knockdown conditions. We note that Cunningham et al. [24] had proposed a mechanism by which competing adjacent 5′ splice sites are simultaneously bound by U1 snRNPs and thereby splicing efficiency is reduced because of strong mutual inhibition. Since under normal conditions intron reads in this region are relatively high (Figure 2B, control), we conclude that, if the alternative 3′ splice site is used, the remainder of intron 7 remains largely unspliced, and the resulting transcript is expected to be degraded by NMD.
To validate our RNA-Seq data and to assess the NMD effect, alternative splicing of the U1-70K exons 7–8 region was analyzed by RT-PCR, using total RNA from control- and U1C-knockdown HeLa cells (Figure 3A). Specific primers were located in the constitutive exons 7 and 8, as well as immediately downstream of the predicted PTC, but upstream of the cryptic 5′ splice sites. Knockdown of U1C decreased recognition of the alternative 3′ splice site in intron 7; conversely, more functional U1-70K mRNA was produced as shown by the increase of the spliced exons 7–8 product (Figure 3B, lanes 1/2 and 5/6). Off-target effects were ruled out by comparing two different U1C-specific siRNAs (one located in the 3′ UTR and another one within the open-reading-frame; U1C vs. U1C*). To test whether exon 7a inclusion indeed results in NMD, the cells were additionally treated with cycloheximide for 5 hours (after three days of siRNA treatment) to block translation and thereby NMD (Figure 3B, lanes 3 and 4). We were able to detect exon 7a inclusion under these conditions using a primer pair located in exons 7 and 8; sequencing and analysis of the RT-PCR products by Bioanalyzer revealed that predominantly the cryptic 5′ splice sites A and B were used, with site A being more frequently used than site B; usage of 5′ splice site C, however, was not significant (for a detailed analysis of splice site usage, see Figure S3A). Notably, sites A and B are much weaker 5′ splice sites than site C (splice site scores: 5.29, 3.38, and 8.91, respectively [25]). Since we can detect exon 7a inclusion only after cycloheximide treatment, we conclude that activation of the alternative 3′ splice site of exon 7a does activate NMD, thereby efficiently removing the non-productive splice isoform of U1-70K.
In addition, after validating these effects on the U1-70K mRNA, we assayed for up-regulation of the U1-70K protein (Figure 3C). Western blot analysis of both whole-cell and nuclear extracts from HeLa cells after U1C knockdown demonstrated that indeed the U1-70K protein levels correlate with the mRNA levels: Upon loss of U1C we detected an increase of U1-70K protein (between 1.4-fold in nuclear extract and 1.6-/2.2-fold in whole-cell extract).
To confirm the U1C specificity of the alternative splicing changes described above, we combined knockdown of endogenous U1C expression and over-expression of Flag/HA-tagged U1C in HeLa cells (Figure 3D, left panel). Clearly, add-back of FLAG/HA-tagged U1C increased exons 7-7a splicing, although control levels were not completely restored; this may be due to inefficient U1 snRNP incorporation or function of the FLAG/HA-tagged U1C protein. At the same time normal exons 7–8 splicing decreased.
Taken together, we conclude that this unusual alternative splicing regulation of U1-70K expression specifically depends on U1C.
Intron 7 contains highly conserved regions, in particular the first 0.8 kb, which include the alternative 3′ splice site and the cryptic 5′ splice sites (Figure 2B and Figure S2A). Therefore, we investigated the conservation of the U1C-dependent effects observed on U1-70K alternative splicing in zebrafish and mouse. First, we used a zebrafish U1C knockout mutant and performed in vivo rescue as previously established [22]. In brief, in vitro transcribed ZfU1C cRNA was injected into U1C mutant zebrafish embryos at the one-cell stage. 2.5 days-post-fertilization rescued embryos were selected according to their phenotypic appearance, and restoration of ZfU1C protein expression was confirmed by Western blotting (Figure 3D, right panel). RT-PCR analysis of total RNA from single embryos showed that U1C knockout in zebrafish completely abolished exon 7a inclusion (Figure 3D, top panel: compare lanes 4 and 5). Add-back of U1C (rescued individual) reactivated the alternative 3′ splice site of exon 7a (7-7a; Figure 3D, top panel lane 6). Second, we performed siRNA-mediated knockdown of U1C in mouse myoblast cells (Figure S2B) and detected a strong decrease in exon 7a inclusion after U1C depletion. In summary, the alternative splicing switch of U1-70K expression appears to be conserved among different vertebrates.
In view of the known functions of U1C in 5′ splice site choice, it was rather unexpected to discover a case of U1C-dependent activation of a 3′ splice site. In order to study the mechanistic basis of this unusual regulation in more detail, we next examined whether the cryptic 5′ splice sites identified downstream of the alternative 3′ splice site are important for its activation. Minigene constructs of U1-70K exons 7 to 8 were generated, maintaining the most highly conserved regions in intron 7 (see Figure 2B), including the alternative 3′ splice site and the three cryptic 5′ splice sites downstream (called A, B, and C; see Figure 4A). Point mutations were introduced to inactivate the cryptic 5′ splice sites individually, in combinations of two, and all three of them (Figure 4A). To analyze the U1C-dependent splicing patterns of the different minigenes in vivo, they were transfected into HeLa cells after three days of U1C or control knockdown; 24 hours later total RNA was isolated for RT-PCR analysis (Figure 4B). In contrast to the endogenous U1-70K expression, where NMD is active, this minigene analysis allowed monitoring the activation of the alternative 3′ splice site by measuring exon 7a inclusion.
As we had observed for the endogenous U1-70K gene (see above), the alternative 3′ splice site was also in this minigene context partially used, so that both exon 7a inclusion and skipping isoforms were detectable. For the wildtype minigene, splicing consistently occurred through 5′ splice site A, and only to a minor extent through sites B and and C. After U1C knockdown exon 7a skipping strongly increased, reproducing the U1C-dependent use of the alternative 3′ splice site (Figure 4B, lanes 1 and 2).
Mutating 5′ splice sites A, B, or both in combination strongly increased exon 7a inclusion in the presence of U1C (A, B, and AB mutants: Figure 4B, compare lanes 1 with 3, 5, and 9); splicing used almost exclusively 5′ splice site C. Interestingly, mutants B and AB showed strong exon 7a inclusion even in the absence of U1C (lanes 6 and 10). After mutation of 5′ splice site C alone, or in combination with A or B, both exon 7a skipping and inclusion were detected (C, AC, and BC mutants: compare lanes 1 with 7, 11, and 13); after U1C knockdown those mutants showed no significant exon 7a inclusion. Only when all three sites were inactivated (mutant ABC: lanes 15 and 16), complete skipping resulted both in the presence and absence of U1C (for a detailed analysis of splice site usage, see Figure S3B).
Together this indicates that the three regulatory 5′ splice sites are particularly important for modulating the use of the exon 7a 3′ splice site in the context of minigene construct, maintaining the balance between skipping and inclusion and sensing U1C-containing versus -deficient U1 snRNPs. Specifically, the splice sites show a differential requirement for U1C in 3′ splice site activation: Splice sites A and B negatively regulate 3′ splice site activation and strongly depend on U1C. In contrast, 5′ splice site C acts positively and appears to be U1C-independent. Therefore we conclude there is a complex interaction network of the three 5′ splice sites, with the three sites contributing both positive and negative individual effects and differential U1C sensitivity.
Next, to investigate U1 snRNP binding to the cryptic 5′ splice sites, we carried out in vitro binding assays. Short RNAs (139 nt) spanning the exon 7a region of U1-70K (enlargement in schematic of Figure 4A) were incubated in HeLa nuclear extract, comparing the wildtype sequence and derivatives with the mutated cryptic 5′ splice sites described above. Western and Northern blot analyses of bound proteins (U1-70K, U1A, U1C) and U1 snRNA, respectively, showed that mutation of 5′ splice sites A or B alone slightly reduced U1 snRNP binding (Figure 4C, compare lanes 3–5); in contrast, mutating splice site C alone showed the same pulldown efficiency as the wildtype sequence (Figure 4C, lanes 1 and 5). However, all double-mutants (AB, AC, and BC) strongly reduced, and the triple-mutant (ABC) completely lost U1 snRNP binding capacity (Figure 4C, lanes 6 to 9). In sum, this suggests an additive behavior of the three cryptic 5′ splice sites in U1 snRNP binding.
We also studied in the endogenous context of the U1-70K gene how important U1 snRNP binding to the cryptic 5′ splice sites is for 3′ splice site activation, using in vivo splice site blocking experiments (Figure 4D). Antisense-morpholinos directed against the three cryptic 5′ splice sites were transfected into HeLa cells, blocking either sites A and B together or B and C (Figure 4D, lanes 1–3). In addition, cells were treated with cycloheximide to inhibit NMD and thereby stabilize RNAs where the exon 7a 3′ splice site had been used (lanes 4–6). Blocking of 5′ splice sites A and B (AB) strongly reduced 3′ splice site activation (spliced exons 7-7a), and exon 7a inclusion (7-7a-8) was undetectable (Figure 4D, compare lanes 1/2 and 4/5). In contrast, blocking 5′ splice sites B and C (BC) still allowed efficient recognition of the alternative 3′ splice site and – under cycloheximide conditions – exon 7a inclusion (Figure 4D, lanes 3 and 6). We conclude that also in the endogenous context 5′ splice sites A and B appear to be involved in and splice site A to be sufficient for 3′ splice site activation.
Surprisingly, AMO blocking of the two 5′ splice sites A and B inhibited 3′ splice site activation in vivo, whereas mutating sites A and/or B in the minigene resulted in strong exon 7a inclusion. Note that in the endogenous 70K gene the 5′ splice site C was not used (Figure 3B, lane 3); we assume that in the minigene context the 5′ splice site C-mediated 3′ splice site activation is favoured because of the construct design, in which the intronic sequence between 5′ splice site C and the downstream 3′ splice site was shortened dramatically. Thus, exon 7a inclusion using site C in the minigene may be induced by intron definition rather than by exon definition mechanism that occurs when the regulatory 5′ splice sites A and B are active.
Finally, we addressed the question whether base-pairing of the U1 snRNA is essential for 3′ splice site activation: HeLa cells were transfected with an AMO that blocks the 5′ end of the U1 snRNA to generally inhibit U1 snRNP binding (for a control of AMO binding to the U1 snRNA see Figure 1B. Again, exon 7a inclusion was observed only in cells that had been treated with the control morpholino under cycloheximide conditions (Figure 4D, right panel). However, inhibition of U1 snRNP binding nearly abolished recognition of the alternative 3′ splice site and resulted in efficient skipping of exon 7a. The minor decrease observed for U1-70K exons 7–8 splicing as well as for the β-actin control probably reflects the general splicing inhibition by the U1 snRNP blocking oligonucleotide. However, U1-70K exons 7-7a splicing is clearly much more severely affected by U1 snRNP blocking, most likely due to differential stability and turnover of these spliced products. In summary, our in vitro binding experiments demonstrated that a direct interaction between U1 snRNA and the cryptic 5′ splice sites is necessary but without U1C not sufficient for 3′ splice site activation.
After we had established that U1-70K levels are regulated by a U1C-dependent alternative splicing event, we asked whether a reciprocal reduction of U1-70K protein may affect U1C protein levels and/or the molecular composition of the U1 snRNP. We performed siRNA-mediated knockdown of U1-70K in HeLa cells and discovered that along with U1-70K protein, U1C protein levels were also strongly reduced in comparison to the control-treated cells (Figure 5A). Furthermore, affinity purification of the U1 snRNP from control- and U1-70K-knockdown cells confirmed that the U1 snRNP lacks both proteins (Figure 5B). However, U1A remained stably bound to the U1 snRNA, which itself appeared to be unaffected by the loss of U1-70K and U1C.
In order to investigate whether this U1-70K/U1C-deficient U1 snRNP has the same effect on U1-70K alternative splicing as observed for the U1C-deficient particle, we performed in vivo splicing assays in control- and U1-70K-knockdown HeLa cells, using our U1-70K wildtype minigene (as described above). RT-PCR analysis showed that exon 7a inclusion was hardly detectable after U1-70K knockdown (Figure 5C). These observations are clearly consistent with a U1C-dependent regulation of U1-70K alternative splicing: Loss of U1-70K reduces levels of total and U1 snRNP-bound U1C, which in turn shifts the alternative splicing balance from the non-productive (exons 7-7a) towards the productive isoform (exons 7–8), increasing U1-70K protein levels.
To address the question how U1-70K reduction resulted in U1C depletion, we examined U1C mRNA stability and potential alternative splicing of U1C pre-mRNA, using RT-PCR and total RNA from control- and U1-70K-knockdown HeLa cells. Figure 5D shows that U1C mRNA levels remained unchanged after U1-70K knockdown (detecting U1C exons 1–4, middle panel). We can also rule out that an annotated alternatively spliced mRNA isoform of U1C without exon 2 is produced (NR_029472), which would introduce a PTC and should therefore reduce mRNA and protein levels. We did neither see a change in the intensity of the exons 1–4 band (212 bp), comparing control- and U1-70K-knockdown samples, nor could we detect the exon 2 skipping product (169 bp). In conclusion, since U1C mRNA stability and alternative splicing appear not to be affected by U1-70K knockdown, the down-regulation of the U1C protein levels most likely occurs on the level of protein stability and/or translation.
Alternative splicing can produce various mRNA isoforms from one precursor transcript by modulating splice site usage in a tissue or developmental-specific manner [26]. In general, alternative splicing factors, such as SR- or hnRNP proteins, are responsible for regulated activation or repression of certain splicing signals [27], [28]. However, the availability of general splicing factors can also influence alternative splicing events as described [22], [29]–[31; this work].
Here we have investigated the regulatory role of U1C in the human system, and found that U1C knockdown in HeLa cells leads to specific alternative splicing alterations rather than a general block of splicing activity. Thus, our analysis gives a genome-wide overview on how a general snRNP protein, which participates in each splicing reaction, can regulate alternative splicing. The two main alternative splicing modes we identified after U1C depletion were increased exon skipping and changes in the usage of alternative 5′ splice sites. Weakly defined exons more strongly depend on accurate U1 snRNP binding and therefore loss of U1C, which generally promotes correct 5′ splice site recognition, is expected to induce exon skipping. In the case of competing 5′ splice sites, it is known that although they can be bound simultaneously by separate U1 snRNPs the downstream one is preferentially used for splicing [32]. In addition to factors known to regulate the choice between distal and proximal 5′ splice sites (such as hnRNP A1 and its antagonist SRSF1) [33]–[35], we have identified here U1C, which appears to be important to promote splicing at the upstream site, consistent with our earlier study in zebrafish [22]. Although we do not know the mechanistic basis of the U1C dependency, this allows the cell to respond to variable U1 snRNP levels by changes in alternative splicing patterns. In sum, our results demonstrate that U1C acts primarily as a splicing activator.
Since U1C is known to stabilize base-pairing between the U1 snRNA and the 5′ splice site, loss of U1C would be expected to impair general splicing activity. However, we found a distinct group of target genes that are affected rather than a general block of splicing activity. Additionally, most targets do not change their U1C-dependent alternative splicing patterns upon U1 snRNP blocking via antisense morpholinos. Therefore we suggest there are target-specific requirements that determine the level of U1C dependency in each case. In fact, our results raise the question, whether U1C is at all a “general” splicing factor or rather a spliceosome-associated alternative splicing regulator.
One U1C target, U1-70K, appeared to be particularly interesting: First, U1-70K and U1C proteins physically interact with each other in the U1 snRNP, with U1C incorporation depending on prior U1-70K binding to loop I of the U1 snRNA. Second, considering the predominant role of U1C in 5′ splice site recognition, it was surprising to discover a case of U1C-dependent 3′ splice site activation. Therefore we decided to analyze in more detail the mechanistic basis for this unusual intra-U1 snRNP cross-regulation. Our RNA-Seq analysis in combination with RT-PCR validation had revealed that under normal conditions an alternative 3′ splice site within intron 7 of the U1-70K pre-mRNA is frequently used, which introduces a PTC into the mRNA and thereby is expected to induce NMD. Previous work had reported for several alternative splicing factors and core splicesosomal components auto-regulatory feedback mechanisms that involve alternative splicing and NMD (for example, [36]–[38]). In contrast to that, however, we found that U1-70K alternative splicing regulation strongly depends on the presence of U1C, and therefore we propose an intra-U1 snRNP cross-regulation mechanism (see below).
What are the sequence requirements for this U1C-dependent alternative splicing process? Downstream of the U1C-dependent alternative 3′ splice site several cryptic 5′ splice sites are located, that turned out to be critical for the regulation we observed. Under normal conditions, those 5′ splice sites are only very rarely used for splicing (only detectable in the absence of NMD), most likely because of their close proximity to each other. Our in vitro binding and in vivo antisense-morpholino blocking experiments confirmed that these regulatory 5′ splice sites efficiently bind U1 snRNPs and that this interaction requires base-pairing with the 5′ end of U1 snRNA.
Mutational analysis of these regulatory 5′ splice sites demonstrated that only two of them (labeled A and B in Figure 2C) convey U1C dependency, with one of them being sufficient for 3′ splice site activation. In fact, these two sites have very low splice site scores (5.29 and 3.38 for site A and B, respectively, compared to 8.91 for site C), which may explain the stringent requirement for U1C to promote U1 snRNP binding to these weak 5′ splice sites. In addition, when analyzed separately by mutant minigenes (see Results and Figure 4B), sites A/B and C behave as negative and positive regulatory elements, respectively, suggesting an intricate network of these regulatory elements, sensing U1C-containing and –deficient U1 snRNPs. This may enhance U1 snRNP binding for efficient 3′ splice site activation, and second, it may contribute to rapid responsiveness and fine-tuning of the regulatory mechanism.
We noted that the sequence including the alternative 3′ splice site and the regulatory 5′ splice sites downstream are highly conserved among vertebrates (Figure S2A). Accordingly, we demonstrated that in zebrafish embryos as well as in C2C12 mouse cells usage of the alternative 3′ splice site and exons 7-7a splicing was strongly reduced after U1C depletion, indicating that the entire U1C/U1-70K cross-regulatory mechanism is conserved.
In sum we established the following model of U1-70K/U1C cross-regulation (Figure 6): Alternative splicing of the U1-70K pre-mRNA provides the central switch between a productive (exons 7–8) and a non-productive (exons 7-7a) splicing mode, whose balance is determined by U1C and U1-70K protein levels. If U1-70K mRNA and protein levels decrease, U1C is co-depleted and U1 snRNPs are assembled inefficiently. This co-depletion of U1C appears to be mediated on the protein level; for example, free U1C protein, which is not assembled into U1 snRNPs may be less stable (Figure 5A and B). The same disturbance of balanced U1-70K alternative splicing can be initiated by U1C knockdown, resulting in U1 snRNPs defective only in U1C. Neither U1C- nor U1C/U1-70K-defective U1 snRNPs are unable to activate the alternative 3′ splice site in U1-70K intron 7, shifting the balance towards productive U1-70K exons 7–8 splicing. As a result, more functional U1-70K mRNA and protein are produced, restoring normal U1 snRNP assembly. Binding of intact, U1 snRNPs to the regulatory 5′ splice sites activates again the U1-70K alternative 3′ splice site and exon 7a inclusion, thereby shifting the alternative splicing balance back towards the non-productive, NMD-inducing mode. The resulting reduction in U1-70K mRNA and protein levels closes the circle. At this point, we cannot rule out a more direct function of U1-70K in the regulation of its own expression, because the effects of U1-70K depletion alone cannot be tested. Mechanistically, the switch could be triggered by a failure of U1-70K to cooperate with SRSF1 to efficiently activate the cryptic 5′ splice sites, which in turn would activate the alternative 3′ splice site through an exon-definition complex. Thus U1-70K itself would be the direct trigger for the alternative splicing switch, and U1C depletion would mimic a lack of U1-70K, because U1-70K is not able to efficiently interact with SRSF1 in the absence of U1C [14], [15]. This situation is reminescent of the auto-regulatory feedback mechanism described for the minor-spliceosomal U11 snRNP: As described by Verbeeren et al. [39], the U11 snRNP can bind to tandem regulatory 5′ splice sites of the minor type, which activates a 3′ splice site in the pre-mRNAs for U11-48K and U11/U12-65K proteins. In contrast, we describe here that one particular component of the U1 snRNP, U1C, is necessary for efficient activation of a 3′ splice site within U1-70K intron 7, and that this is part of a regulatory circuit linking the expression of both U1C and U1-70K proteins. Notably, U11-48K has a similar role in 5′ splice site recognition for the minor spliceosome as U1C for the major spliceosome [39], [40]. Thus, both spliceosomes appear to regulate the expression of their intrinsic factors by a comparable mechanism to ensure correct 5′ splice site recognition.
Taken together we describe a novel and conserved intra-U1 snRNP cross-regulation mechanism that ensures U1-70K and U1C homeostasis and guarantees stoichiometrically correct U1 snRNP assembly. This provides a new paradigm for and mechanistic insight in molecular communication within the spliceosome. It opens up another emerging new question of wide general interest, how the biosynthesis of the more than 100 protein and RNA components of the spliceosome is coordinated.
siRNA duplexes were transfected into HeLa cells using Lipofectamine 2000 (Invitrogen) according to the manufacturer' instructions. For U1C knockdown, siRNAs specific for the human U1C mRNA located in the 3′ UTR (ΔC: 5′-AGGCCUUAUUGUAUCGGUU[dT][dT]) or the open-reading-frame (ΔC*: AAACAACGGCUGCAUUUCAAC[dA][dC]) were transfected at a final concentration of 40 nM. For U1-70K knockdown, a specific siRNA (ΔU1-70K: 5′- GAGAGGAAAAGACGGGAAA[dT][dT]) was transfected at a final concentration of 60 nM. For U1C knockdown in C2C12 cells, an siRNA against the mouse U1C mRNA was reverse transfected (at a final concentration of 140 pmol), using Lipofectamine RNAiMax Reagent (Invitrogen). An siRNA specific for the firefly luciferase mRNA (5′-CGUACGCGGAAUACUUCGA[dT][dT]) served as a control (ctr) in all cases. Knockdown efficiencies were determined by Western blot, using monoclonal antibodies against U1C (4H12, Santa Cruz), U1-70K (H111, Synaptic Systems), and, as a control, γ-tubulin (GTU-88, Sigma).
U1 snRNA 5′ end and U1-70K cryptic 5′ splice site blocking were achieved by antisense morpholino transfections as described elsewhere [22]: Briefly, transfections were performed, using the Nucleofector Solution R (Lonza) and the Nucleofector Programm I-013 according to the manufacturer's instructions. 1.5×106 HeLa cells were transfected with AMO (U1, 5′-GGTATCTCCCCTGCCAGGTAAGTAT-3′, at 100 µM [23]; AB, 5′-ACAAACCCTTATACCAACCATACAC-3′ and AC, 5′- GATCTTACCCATGATACAAACCCTT-3′, at 50 µM each), and 15 hours later total RNA was isolated for further analysis. A control AMO (ctr, 5′-CCTCTTACCTCAGTTACAATTTATA-3′, [23]) was transfected at the same concentrations as used for the specific AMO. The efficiency of U1 snRNA inhibition was analyzed by an RNase H protection assay: Whole cell extracts were incubated with 5 µM antisense DNA oligonucleotide (5′-CAGGTAAGTAT-3′) and 1.5 U RNase H (Promega) for 30 min at 37°C. After phenol extraction the total RNA was analyzed on a 10% denaturing polyacrylamide gel followed by silver staining.
For NMD inhibition the growth medium was supplemented with 50 µg/ml cycloheximide (+CHX), and cells were harvested 5 hours later for Western blot analysis and total RNA isolation.
For the expression construct the coding region of U1C was amplified from cDNA generated from HeLa total RNA. The N-terminal Flag- and the C-terminal HA-tags were introduced by PCR, followed by cloning into pcDNA3 (Invitrogen) between the HindIII and XhoI restriction sites.
For simultaneous U1C over-expression and knockdown of endogenous protein, first, the U1C expression vector (pcDNA3_Flag-U1C-HA) was transfected (TurboFect in vitro Transfection Reagent, Fermentas), and 24 hours later siRNAs (final concentration: 140 pmol) were reverse-transfected, using Lipofectamine RNAiMax Reagent (Invitrogen). Overexpression was detected by Western blot, using monoclonal antibodies against U1C (4H12, Santa Cruz) and Flag (Sigma).
For details on poly(A)+-RNA selection for Solexa high-throughput sequencing (GAIIx), the data analysis, and alternative splicing target selection, see Rösel et al. [22]. Total RNA from HeLa cells was prepared 72 hours after siRNA transfection by TRIzol reagent (Invitrogen) and RNeasy kit (QIAGEN). The 105-bp single-end sequence reads were aligned to human genome (hg19) and a junction sequence data constructed with Gene Annotations from ENCODE Version 11. RNA-Seq raw data and processed coverage data were uploaded to the GEO database at NCBI (GSE42485).
For target gene validation, 1 µg total RNA was reverse transcribed (iScript cDNA synthesis kit, BioRad) and subjected to PCR using specific primer sets that span the region of interest. Table S5 lists all oligonucleotides used.
Three days after siRNA transfection, U1-70K minigene constructs (5 µg per 6-cm dish) were transfected into HeLa cells using FuGeneHD (Promega), and 24 hours later total RNA was isolated using Trizol (Invitrogen) and treated with RQ1-DNase (Promega). Reverse transcription was performed using the minigene specific BGHreverse primer (qScript Flex cDNA Kit; Quanta Biosciences), and for PCR gene-specific primers were used to analyze alternative splicing patterns.
The wildtype minigene construct was amplified in several PCR steps from HeLa genomic DNA and cloned into pcDNA3 vector using BamHI and EcoRI restriction sites. The final three-exon-construct (7-7a-8) comprises the full sequence of U1-70K exon 7, intron 7 positions 1–844, 1,928–2,227, and 3,077–3,162, with the sequences in between deleted (1,083 nt and 849 nt), and the full sequence of exon 8. Based on the wildtype construct point mutations were introduced by PCR, substituting the GT of the cryptic 5′ splice sites by AC (Figure 4A). Only the second half of the construct was reamplified to insert the point mutation and cloned into the wildtype construct, using an endogenous XcmI restriction site and the EcoRI site introduced with the exon 8 reverse primer.
Ethidium bromide-stained bands were quantified using the GeneTools software provided with the G:BOX gel documentation system from SynGene.
For better resolution exon 7a inclusion products (primers from exon 7–8), 1 µl of selected reactions (Figure S3) were analyzed on an Agilent 2100 Bioanalyzer DNA 1000 Chip.
RNA spanning the U1-70K exon 7a region (starting 46 nt upstream of the alternative 3′ splice site until 53 nt downstream of the third cryptic 5′ splice site) was in vitro transcribed (T7 High Yield RNA Synthesis Kit, NEB) from a PCR-generated DNA template. Purified transcripts (184 nt) were chemically 3′-biotinylated [41], and 60 pmol were incubated with 50 µl HeLa nuclear extract (CILBIOTECH, Mons, Belgium) in a total volume of 400 µl binding buffer (20 mM HEPES/KOH pH 7.5, 100 mM KCl, 10 mM MgCl2, 0.01% NP-40, 1 mM DTT) for 1 hour at room temperature. Bound material was pulled down via NeutrAvidin agarose beads (Thermo Scientific) for 2 hours at 4°C, and after several washing steps (20 mM HEPES/KOH pH 7.5, 200 mM KCl, 10 mM MgCl2, 0.01% NP-40, 1 mM DTT) bound proteins were analyzed by SDS-PAGE and Western blot, and bound U1 snRNA was detected by Northern blot hybridization.
The affinity purification of U1 snRNPs from HeLa nuclear extracts or whole cell extracts (after U1-70K knockdown) was according to Palfi et al. [42]; the experimental procedure is basically the same as described above for the in vitro binding assays but using a 3′-biotinylated 2′-O-methyl antisense RNA oligonucleotide (5′-GCCAGGUAAGUAU-3′) directed against the 5′ end of the U1 snRNA. Affinity-selected U1 snRNA was detected by Northern blotting, and co-purified proteins were examined by Western blot.
ZfU1C-cRNA injection into Danio rerio embryos at the 1-cell stage was performed as described previously [22]. Phenotypically wildtype individuals were sorted into wildtype (wt), mutant (mut), and “rescued” (rsc) individuals. Single embryos were used to measure ZfU1C protein expression by Western blot, using specific antibodies against ZfU1C and γ-tubulin, as a control. Splicing patterns of zebrafish U1-70K were analyzed by RT-PCR using total RNA isolated from single embryos and specific primers against exons 7, 7a, and 8.
|
10.1371/journal.pntd.0006402 | A real-time PCR tool for the surveillance of zoonotic Onchocerca lupi in dogs, cats and potential vectors | The ocular onchocercosis is caused by the zoonotic parasite Onchocerca lupi (Spirurida: Onchocercidae). A major hindrance to scientific progress is the absence of a reliable diagnostic test in affected individuals. Microscopic examination of skin snip sediments and the identification of adults embedded in ocular nodules are seldom performed and labour-intensive. A quantitative real-time PCR (qPCR) assay was herein standardized for the detection of O. lupi DNA and the results compared with microscopic examination and conventional PCR (cPCR). The specificity of qPCR and cPCR was assessed by processing the most common filarial nematodes infecting dogs, skin samples from O. lupi infected (n = 35 dogs) or uninfected animals (n = 21 dogs; n = 152 cats) and specimens of potential insect vector (n = 93 blackflies; n = 59 mosquitoes/midges). The analytical sensitivity of both assays was assessed using 10-fold serial dilutions of DNA from adult specimen and from a pool of microfilariae. The qPCR on skin samples revealed an analytical specificity of 100% and a sensitivity up to 8 x 10−1 fg/2μl O. lupi adult-DNA and up to 3.6 x 10−1 pg/2μl of mfs-DNA (corresponding to 1 x 10−2 mfs/2μl). Only 9.5% O. lupi-infected skin samples were positive for cPCR with a sensitivity of 8 x 10−1 pg/2μl of DNA. Out of 152 blackflies and mosquitoes/midges, eight specimens experimentally infected (n = 1 S. erythrocephalum; n = 1 S. ornatum; n = 6 Simulium sp.) were positive by qPCR. The qPCR assay herein standardized represents an important step forward in the diagnosis of zoonotic onchocercosis caused by O. lupi, especially for the detection and quantification of low number of mfs. This assay provides a fundamental contribution for the establishment of surveillance strategies aiming at assessing the presence of O. lupi in carnivores and in insect species acting as potential intermediate hosts. The O. lupi qPCR assay will enable disease progress monitoring as well as the diagnosis of apparently clinical healthy dogs and cats.
| The diagnosis of zoonotic ocular onchocercosis caused by Onchocerca lupi (Spirurida: Onchocercidae) is currently based on microscopic examination of skin snip sediments and on the identification of adults embedded in ocular nodules. These methods are labour-intensive and require multiple steps to achieve the diagnosis. In this context, a novel quantitative real-time PCR assay (qPCR) has been herein standardized and analytical specificity and sensitivity assessed. The results indicate that the qPCR assay could represent an important step forward in the diagnosis of onchocercosis, in carnivores and in insect species acting as potential intermediate hosts.
| Within the genus Onchocerca (Spirurida: Onchocercidae), Onchocerca volvulus and Onchocerca lupi parasitize humans and carnivores, respectively [1–5], the latter being a zoonotic agent [6,7]. While O. volvulus is a well-known parasite of humans transmitted by blackflies (Simulium spp.) [8,9], the epidemiology of O. lupi is far from being understood, particularly because the information about insect species acting as vectors is lacking. Only Simulium tribulatum was suggested as the putative vector of this filarial worm in California (USA), but proof of its intermediate host competence is currently absent [10]. Onchocerca lupi belongs to the spirurids in the Nematode clade III [11] was first detected from a Caucasian wolf (Canis lupus) in Georgia [12], and, only recently, diagnosed in dogs and cats from Europe (Greece, Portugal, Spain, Germany, Hungary) and USA [13–20]. The reports of O. lupi infection are mainly based on the presence of ocular nodules on the eyelids, conjunctiva, and sclera [3,21,22], though the localization of adult worms in the retrobulbar area of the canine patients may impair the assessment of its distribution in endemic areas [23]. The detection of microfilariae (mfs) in skin snip sediments is the only available tool for the diagnosis of the infection when nodules are not apparent in the eyes. The retrieval and identification of mfs in skin snip samples is a rather invasive and time-consuming method, highly dependent on the anatomical location of skin biopsy and mfs density [24]. Again, the detection of mfs may depend upon the prepatent period, previous microfilaricidal treatments, and on the operator’s skills in examining skin sediments, as described for O. volvulus [25,26].
Conventional PCR (cPCR) amplification and sequencing of mitochondrial NADH dehydrogenase subunit 5 (ND5) and cytochrome c oxidase subunit 1 (cox1) genes are available for the molecular identification of O. lupi adults and mfs [7,27,28]. The cPCR, however, may be relatively labour-intensive and exhibit low sensitivity, mainly for mfs detection, limiting the establishment of large-scale epidemiological studies in vertebrate hosts and putative vectors.
Here, we developed a quantitative real-time PCR (qPCR) assay based on the hybridization probe to detect O. lupi DNA in host and putative vector samples. The diagnostic validity of qPCR assay was compared with microscopic examination and cPCR methods.
All dogs’ and cats’ skin samples were collected in previous studies [17,29] and approved by the ethical committee of the Department of Veterinary Medicine of the University of Bari (Prot. Uniba 1/16) and by the ethical committee of the Faculty of Veterinary Medicine, Universidade Lusófona de Humanidades e Tecnologias.
Genomic DNA of adult specimens of O. lupi (n = 3), as well as DNA from single (n = 7) or pooled mfs (n = 10), collected from dogs in different geographical locations (Table 1) were used as control. All specimens were previously identified based on morphological and molecular analyses [18,30].
Primers (O.l.F 5′-GGAGGTGGTCCTGGTAGTAG-3′; O.l.R 5′- GCAAACCCAAAACTATAGTATCC-3′) and a TaqMan-MGB hydrolysis probe (FAM-5’-CTTAGAGTAGAGGGTCAGCC-3’-non-fluorescent quencher-MGB; Applied Biosystems; Foster City, CA, USA), targeting partial cox1 gene (90bp), were designed by alignment of sequences from a wide range of closely related filarial nematodes available from GenBank database (Table 2), using Primer Express 2.0 (Applied Biosystems, Foster City, CA). Specificity of the primers and probe for O. lupi were confirmed in silico using the basic local alignment search tool (BLAST, GenBank, NCBI).
qPCR reactions were carried out in a final volume of 20μl, consisting of 10μl of IQ Supermix (Bio-Rad Laboratories, Hercules CA, USA), 7.1μl of Di-Ethyl Pyro-Carbonate (DEPC) treated pyrogen-free DNase/RNase-free water (Invitrogen, Carlsbad, CA, USA), 2μl of template DNA (except no-template controls), 5 pmol and 0.5 pmol for primers and probe, respectively.
The run protocol consisted of a hot-start at 95°C for 3 min, and 40 cycles of denaturation (95°C for 10 sec) and annealing-extension (64°C for 30 sec). All assays were carried out in duplicate and a no-template control was included in each run. The qPCR was performed in a CFX96 Real-Time System (Bio-Rad Laboratories, Inc., Hercules CA, USA) and the increase in the fluorescent signal was registered during the extension step of the reaction and analysed by the CFX Manager Software Version 3.1 (Bio-Rad).
To investigate the analytical specificity of the assay, genomic samples of Onchocerca spp. and of the most common filarial nematodes infesting dogs (Table 1) were used. The specificity of the assay was tested by using DNA from skin samples of naturally infected dogs, which were positive for O. lupi (n = 35) at microscopic examination [29]. Skin samples were divided in five groups (G1-G5) according to their mfs load (Table 3), being 14 also co-infected with Cercopithifilaria bainae and Cercopithifilaria sp. II. Skin samples (dogs n = 21; cats n = 152), which did not test positive to any mfs [17,29], were used as negative control.
Specimens of blackflies (n = 66) and mosquitoes/midges (n = 39) collected from 2011 to 2014 in Greece [31], and 27 blackflies and 20 Aedes albopictus (colony specimens) experimentally infected by intrathoracic microinjection with mfs of O. lupi (parasitic load of 20mfs/μl) were analyzed after death (i.e., from one to 10 days post infection) (Table 4).
The analytical sensitivity of the qPCR assay was assessed using 10-fold serial dilutions of DNA from adult specimen (i.e., ranging from 8 × 104 to 8 × 10−3 fg/2μl of reaction) and from a pool of 10 mfs (i.e., ranging from 10 to 10 × 10−3 microfilariae/2μl of reaction, corresponding to 3.6 ×10−1 ng/2μl to 3.6 ×101 fg/2μl of DNA). Ten replicates of each serial dilution were submitted to the same run for assessment of intra-assay reproducibility.
Genomic DNA was isolated from all skin samples and from O. lupi adults and mfs, blackflies, mosquitoes and midges specimens using the commercial kits DNeasy Blood & Tissue Kit (Qiagen, GmbH, Hilden, Germany), respectively, following the manufacturers’ instructions. The amounts of purified DNA were determined spectrophotometrically using the Qubit (Applied Biosystems, Foster City, CA, USA).
The analytical specificity and sensitivity of the cPCR for the specific amplification of cox1 gene fragment (∼689bp; [32]) was assessed by testing genomic DNA of: i) skin samples with different parasitic load of O. lupi (Table 3), ii) serial dilution of O. lupi mfs DNA (i.e., from 3.6 ×101 pg/2μl to 3.6 ×10−3 pg/2μl) and iii) DNA of adult specimens (i.e., from 8 ×101 ng/2μl to 8 x 10−3 fg/2μl).
All cPCR products were resolved in 0.5x GelRed stained (Biotium, CA, USA) agarose gels (2%), purified using enzymatic purification (Exo I-FastAP; Thermo Fisher Scientific, MA, USA) and sequenced in an automated sequencer (3130 Genetic Analyzer). All sequences generated were compared with those available in GenBank using Basic Local Alignment Search Tool (BLAST) [33].
All O. lupi naturally-infected dog positive at skin samples examination by microscopy, considered the gold standard method as true positives, were positive by the O. lupi qPCR herein assessed (specificity of 100%). Out of 21 skin samples microscopically and qPCR positive for O. lupi, two were positive by cPCR (parasite load of 8 and 25 mfs), revealing a low analytical cPCR specificity (i.e., 9.5%). None of cat’s skin samples were positive by qPCR.
A specific fluorescent signal was recorded for all O. lupi adult and mfs positive controls tested (Fig 1). No fluorescence was obtained for all other Onchocerca species or filarial nematodes examined as well as for skin samples used as negative control.
The analytical sensitivity of qPCR was confirmed by detection of up to 8 x 10−1 fg/2μl and 3.6 x 10−1 pg/2μl of DNA (i.e., corresponding to 1 x 10−2 mfs/2μl) of O. lupi adult worm and mfs, respectively (Fig 2A and 2B). qPCR efficiencies ranged from 108.7 to 115.3% with an R2 from 0.996 to 0.999 and Slope from -3.003 to -3.131, for both adult and mfs (Fig 2A and 2B). The mean parasite load detected for the positive skin samples, ranged from 1.9 to 96 mfs/2μl of reaction, corresponding to 6.1 x 10−2 ng/2μl (mean cycle threshold of 33.49) and to 3.4 ng/2μl DNA (mean cycle threshold of 27.52), respectively (Table 3). The results of mfs detection by qPCR overlapped the values obtained by the microscopic examination. The detection limit registered for cPCR was up to 8 x 10−1 pg/2μl for adult worms and up to 3.6 x 101 pg/2μl for mfs DNA (i.e., corresponding to 1 mf/2μl), respectively (Fig 3).
Out of 152 blackflies, mosquitoes and midges, eight Simulim spp. (n = 1 S. erythrocephalum; n = 1 S. ornatum; n = 6 Simulium sp.), experimentally infected and died from 1 to three days post infection, returned positive signal for O. lupi DNA (Table 4). All field-collected blackflies and mosquitoes were negative for O. lupi DNA using qPCR (Table 4). All blackflies positive for qPCR scored positive also for cPCR.
Sequences derived from all amplicons of cPCR matched with 99–100% nucleotide identity appropriate reference sequences of O. lupi available from GenBank (accession numbers KC686702, KC686701).
A qPCR assay has been developed for the detection of O. lupi in animal skin snip samples and potential vectors and proved to be a sensitive and specific tool for the diagnosis of this parasite, with a mean detection limit as low as 1.9 mfs per reaction. In addition, the high sensitivity of the qPCR protocol has been demonstrated by detecting a small amount of DNA (up to 8 x 10−1 fg/2μl for adult and up to 3.6 x 10−1 pg/2μl for mfs), by the slope value of standard curve (−3.131), the efficiency (115.3%) and the coefficient of determination (R2 = 0.999). These features of the assay are due to the selection of a stable hydrolysis probe designed (100% specific for O. lupi DNA), as well as to the choice of the target gene used. Indeed, cox1 gene of the mitochondrial DNA has been well recognized as a “barcode” for filarial nematodes [34], with a high amplification efficiency, also due to the large copy numbers enabling the detection of minimum amounts of DNA [35–37]. Though few Onchocerca species DNA were herein tested, which may represent a limitation of the qPCR assay, this new tool provides an alternative to the labor intensive microscopic examination of skin snip samples and to cPCR for the diagnosis of O. lupi [38]. The qPCR assay was highly specific in revealing O. lupi DNA both in co-infected samples from dogs as well as in potential vector species, avoiding the sequencing confirmation needed using cPCR with filarioid generic primers [32]. Overall, the positive fluorescent signal from samples of O. lupi, from different geographical areas (i.e., Europe and USA), which displayed genetic intraspecific variability [18], indicates the usefulness of the qPCR also for the surveillance of O. lupi where the parasite has been reported [13,14,16,17,19,39–41]. Similarly, even if the qPCR cannot discriminate between viable and nonviable parasites or immature and infective larvae, the assay could be useful for detecting O. lupi in blackfly, mosquito and/or midge species, potentially involved in the transmission of this parasite. Indeed, the specificity of the qPCR to amplify exclusively the DNA of the pathogen in potential insect vectors herein tested, may ultimately assist in the quest to identify the elusive vector of O. lupi.
The newly designed assay represents an improvement in the diagnosis of onchocercosis, by the detection and quantification of low mf densities from tissue samples and could provide a contribution to disease progress monitoring and to the surveillance of O. lupi-infected dogs, avoiding the introduction and/or spread of this life-threatening parasitic nematode, as well as to the identification of apparently healthy animals [29, 42].
The qPCR may speed-up time of diagnosis and prompt treatments of infected animals, which may avoid the appearance of nodular lesions in the eyes or in other anatomical localizations [43].
A TaqMan-based specific and sensitive assay without sequencing is expected to assist high-throughput analysis of samples, eventually leading to improve disease monitoring under the frame of a Public Health perspective. This would be particularly relevant considering that, since its first description of its zoonotic potential [7], cases of zoonotic onchocercosis are being detected increasingly in people from Europe, Iran and the USA [44–47].
|
10.1371/journal.pgen.1008178 | Genetic and functional data identifying Cd101 as a type 1 diabetes (T1D) susceptibility gene in nonobese diabetic (NOD) mice | Type 1 diabetes (T1D) is a chronic multi-factorial disorder characterized by the immune-mediated destruction of insulin-producing pancreatic beta cells. Variations at a large number of genes influence susceptibility to spontaneous autoimmune T1D in non-obese diabetic (NOD) mice, one of the most frequently studied animal models for human disease. The genetic analysis of these mice allowed the identification of many insulin-dependent diabetes (Idd) loci and candidate genes, one of them being Cd101. CD101 is a heavily glycosylated transmembrane molecule which exhibits negative-costimulatory functions and promotes regulatory T (Treg) function. It is abundantly expressed on subsets of lymphoid and myeloid cells, particularly within the gastrointestinal tract. We have recently reported that the genotype-dependent expression of CD101 correlates with a decreased susceptibility to T1D in NOD.B6 Idd10 congenic mice compared to parental NOD controls. Here we show that the knockout of CD101 within the introgressed B6-derived Idd10 region increased T1D frequency to that of the NOD strain. This loss of protection from T1D was paralleled by decreased Gr1-expressing myeloid cells and FoxP3+ Tregs and an enhanced accumulation of CD4-positive over CD8-positive T lymphocytes in pancreatic tissues. As compared to CD101+/+ NOD.B6 Idd10 donors, adoptive T cell transfers from CD101−/− NOD.B6 Idd10 mice increased T1D frequency in lymphopenic NOD scid and NOD.B6 Idd10 scid recipients. Increased T1D frequency correlated with a more rapid expansion of the transferred CD101−/− T cells and a lower proportion of recipient Gr1-expressing myeloid cells in the pancreatic lymph nodes. Fewer of the Gr1+ cells in the recipients receiving CD101−/− T cells expressed CD101 and the cells had lower levels of IL-10 and TGF-β mRNA. Thus, our results connect the Cd101 haplotype-dependent protection from T1D to an anti-diabetogenic function of CD101-expressing Tregs and Gr1-positive myeloid cells and confirm the identity of Cd101 as Idd10.
| The complex interplay of environmental factors and genetic traits determines the susceptibility of an individual to autoimmune disease such as type 1 diabetes (T1D). Despite T1D being one of the most common and most studied polygenic autoimmune disorders, the mechanisms underlying the immune-mediated destruction of the insulin-producing pancreatic beta cells are still largely unknown. Genetic association studies identified many DNA sequence variants that confer risk to or protect from autoimmune disease. In this regard, we have identified a single gene, Cd101, as a T1D susceptibility locus. In accordance with our previous studies in which we reported an association of allelic Cd101 variants on T1D prevalence, we observed here that deletion of Cd101 perpetuated the expansion of pathogenic, pancreas-infiltrating immune cells and subsequently enhanced T1D incidence. The mechanisms by which Cd101 variants interfere with autoimmune responses will allow us to understand the regulation of molecules in autoimmunity in general as diabetes susceptibility loci have been associated with other autoimmune diseases. Consequently, our work will help to identify therapeutic approaches that can be used to guide the development of effective therapies for T1D, but also allows the identification of common targets in autoimmune disease for clinical intervention in the future.
| Type I diabetes (T1D) is a complex autoimmune disease driven by multiple genetic traits and facilitated by various immune cells infiltrating the pancreatic islets. In rodent models such as the frequently studied non-obese diabetic (NOD) mouse, myeloid cells including macrophages, dendritic cells (DCs) and neutrophils are the first cells to accumulate in the pancreas [1–4] and contribute to the initiation and perpetuation of the T cell-driven pancreatic islet destruction [5]. Myeloid-derived suppressor cells (MDSCs) and regulatory T cells (Tregs), in contrast, suppress diabetogenic immune cells and hinder the development of T1D [6, 7]. While the transcription factor FoxP3 identifies Tregs, the delineation of MDSCs from other myeloid cell subsets is challenging and requires a scrutinized investigation.
Tregs are pivotal for the maintenance of immune homeostasis. Their ability to suppress other immune cells maintains tolerance to self-antigens and prevents autoimmune disease. The depletion of Tregs promotes the development of T1D while their transfer or therapeutic enhancement exhibits protective effects [8–10]. Compared to other reference strains, some studies reported primary deficits in Treg numbers of NOD mice [8, 11–13], whereas others did not [14–19]. Although Treg frequencies in T1D patients appear normal in most studies, defects in the phenotype and the suppressive capacity of Tregs have been reported [20–23]. In mice, induced NOD-derived Tregs are also less effective in standard in vitro suppression assays and reveal subtle defects in the expression of distinct genes [18], although the relevance of these genes on Treg function and on the induction of T1D in NOD mice remains to be determined. An age-related decline in Treg function of NOD mice over time is also pivotal for T1D development in NOD mice [14–16].
Various myeloid cell populations and myeloid cell responses are frequently altered and impeded in NOD mice [24]. Thus, the development of precursors to dendritic cells (DCs) and macrophages, for example, is hampered in NOD mice [25–27] as well as the maturation of myeloid DCs [28]. In addition, the recruitment of neutrophils to sites of infection is severely impaired [24]. The molecular signals underlying these phenotypes, however, have rarely been identified.
CD101 is a negative costimulatory molecule expressed on subsets of myeloid and lymphoid cells [29–33]. Upon engagement of CD101 by agonistic antibodies myeloid cells are induced to be immunosuppressive in vitro [34]. In vitro, CD101+ Tregs are more suppressive than their CD101− counterparts [30]. In vivo, CD101 perpetuates the suppressive function of Tregs and reduces the development of T1D and chronic colitis [30, 31, 35]. Furthermore, CD101+ myeloid cells release more IL-10 than CD101− myeloid cells [35]. In addition, reduced CD101 expression is observed in T1D and Inflammatory Bowel Disease patients [35, 36]. Rare polymorphisms in the Cd101 gene have been suggested to underlie the reduced CD101 expression in some T1D patients [36].
We have identified Cd101 as a T1D candidate gene within the Idd10 region using multiple Idd10 congenic strains [37]. Susceptibility to T1D was correlated with genotype-dependent CD101 expression on multiple cell subsets, including Foxp3+ Tregs, CD11c+ dendritic cells, and Gr1+ myeloid cells [31]. To evaluate the impact of CD101 on T1D, we introgressed the Idd10 and Idd10/Idd18 regions from a B6 CD101 KO strain onto the NOD background and observed that T1D protection mediated by the B6-derived Idd10 and Idd10/Idd18 regions was lost in CD101−/− NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10/Idd18 mice. The loss of CD101 expression reduced the frequency of Tregs and transformed anti-inflammatory Gr1-expressing myeloid cells into an inflammatory, disease-promoting subset. Thus, our data further confirm the identity of Cd101 as Idd10 and provide cellular mechanisms by which the molecule mediates its protection from T1D.
Auto-reactive T cells are initially primed in the pancreatic lymph nodes of NOD mice beginning at the age of three weeks [5] and attracted by an infiltration of innate immune cells into the pancreatic islets one to three weeks later [38]. As we had observed that CD101 expression was genotype-dependent in bone marrow and spleen in NOD Idd10 congenic strains [31], we assessed the influence of Cd101 gene variation on the distribution of immune cells in peripheral lymph nodes as compared to spleen. In both lymph nodes and spleen CD101 expression was observed on a portion of T cells and myeloid cells (Fig 1A, S1A and S1B Fig). T cells constituted the majority of CD101-expressing cells in peripheral lymph nodes but not in the spleen (S1C Fig). A small portion of the CD101-expressing cells were neither T cells nor CD11b+ (S1C Fig). We observed an increased proportion of CD101-positive T cells and of CD101-positive Tregs which represent the largest CD101-positive subset within the T lymphocyte population [29–33] in the pancreatic lymph nodes of 4-8-week-old NOD and NOD.B6 Idd10 mice compared to other peripheral lymph nodes (Fig 1A and 1B; S2 Fig, S3 Fig). The percentages of CD101-positive T cells between NOD and NOD.B6 Idd10 mice were comparable (S3B Fig) while the mean fluorescence intensity for CD101 was often enhanced on T cells from NOD.B6 Idd10 mice (S3C Fig), a phenotype observed previously in splenic T cells [31]. CD101-expressing Tregs in pancreatic, but not popliteal, lymph nodes of NOD.B6 Idd10 mice were more frequent at the two later time points assessed compared to NOD mice (Fig 1C and 1D; S4A Fig), while Treg frequencies themselves in both organs remained comparable (Fig 1C and 1E; S4B Fig). The proportion of other CD101-expressing myeloid and lymphoid subsets was also similar between NOD and NOD.B6 Idd10 mice (S5 Fig). Thus, the introgression of the B6 Idd10 region not only promotes the expansion of Gr1-positive myeloid cells in the bone marrow [31] but also favors at later time-points an accumulation of CD101-expressing Tregs in the pancreatic lymph nodes.
Based on sequence comparisons of four Idd10 regions tested for T1D susceptibility and the observation that susceptibility to T1D correlated with genotype-dependent CD101 expression on multiple immune cell subsets, Cd101 is the prioritized gene candidate for the Idd10 region [31]. We therefore reasoned that if Cd101 is Idd10, elimination of CD101 protein expression should alter T1D susceptibility, a finding that would further strengthen our hypothesis that allelic variation in the structure or expression of CD101 influences T1D frequency in the context of the NOD background. Following deletion of a portion of the B6 Cd101 gene required for protein expression [31] the T1D-protective Idd10 and Idd10/18 regions carrying the Cd101 modification were introgressed onto the NOD background by genotype-selected backcrossing to generate congenic strains (Fig 2A). We wanted to examine the effect of eliminating CD101 expression both in the context of the B6-derived protective Idd10 region, but also in the context of the more complex Idd10/18 region where multiple Idd subregions have been defined [39, 40] and are depicted on Fig 2A. Therefore, as backcrossing of the B6 CD101−/− strain to the NOD strain occurred, we screened progeny for recombination events that most closely resembled those defining the boundaries of the regions in previous studies in order to later characterize the strains versus the congenic regions with intact B6 Cd101 alleles. As was observed on the B6 background [31], the engineered deletion within the Cd101 gene led to a lack of CD101 protein expression on all myeloid and T cells in CD101−/− NOD.B6 Idd10 (Fig 2B; S2 Fig) and CD101−/− NOD.B6 Idd10/18 mice (Fig 2C).
We have recently reported a correlation of CD101 expression on immune cells from four independent Idd10 haplotypes with the development of T1D [31]. Thus, to further establish the causative role of Cd101 in the pathogenesis of T1D, we evaluated T1D frequencies in our newly generated CD101−/− NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10/18 mice compared to parental CD101-expressing NOD.B6 Idd10 and NOD.B6 Idd10/18 controls. We observed that CD101-expressing NOD.B6 Idd10 and NOD.B6 Idd10/18 mice developed T1D substantially slower and with a reduced incidence than their CD101-deficient counterparts in two different animal facilities (Fig 3). Indeed, CD101−/− NOD.B6 Idd10 mice and CD101−/− NOD.B6 Idd10/18 mice had frequencies of diabetes equivalent to that of NOD mice housed in the same colony (Fig 3A, 3C and 3D). Contemporaneously, cohorts of CD101+/+ and CD101−/− NOD.B6 Idd10 mice as well as CD101+/+ and CD101−/− NOD.B6 Idd10 progeny from heterozygous CD101+/− NOD.B6 Idd10 intercross breeders were monitored for diabetes (Fig 3A and 3B). In the latter comparison CD101-replete and CD101 KO Idd10 homozygous progeny are part of the same litters and therefore exposed to the same micro-environment. Once again, mice homozygous for the CD101 KO Idd10 region had a higher frequency of diabetes than those having two doses of the intact B6-derived Idd10 region (Fig 3B); however, the difference in diabetes occurrence was less significant between the homozygous genotypes derived from heterozygous breeders than when the mice had been bred from homozygous parents (Fig 3A). CD101+/− NOD.B6 Idd10 heterozygous progeny had a diabetes frequency intermediate between those of CD101+/+ and CD101−/− NOD.B6 Idd10 progeny (S6 Fig). Protection from diabetes was associated with a significantly reduced infiltration of pancreatic islets by immune cells in CD101-expressing congenic mice at 8–10 weeks of age (Fig 3E, S7 Fig). Thus, these data confirm the protective role of the B6 Cd101 allele within the Idd10 region and strongly suggest that the gene encoding CD101 is Idd10.
In order to define the CD101-expressing cell subset(s) promoting protection from T1D, we assessed the distribution of myeloid and lymphoid cells in the organs of CD101-expressing and CD101-deficient NOD.B6 Idd10 and NOD.B6 Idd10/18 mice. The inflammatory infiltrate in the pancreata of CD101−/− NOD.B6 Idd10/18 and CD101−/− NOD.B6 Idd10 mice and respective CD101-expressing controls consisted mainly of T lymphocytes (Fig 4A) and few Gr1-expressing myeloid cells (Fig 5A and 5B). There was an enhanced proportion of CD4- over CD8-positive T lymphocytes within the TCRβ+ population in both CD101−/− strains compared to their CD101-expressing counterparts (Fig 4B–4D). While the NOD versus B6 Idd10 haplotypes revealed no differences in Treg percentages in the pancreatic lymph nodes (Fig 1E), Tregs were significantly reduced in the pancreatic lymph nodes (Fig 4E), but not the spleens (Fig 4F) of CD101−/− NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10/18 mice compared to CD101-expressing NOD.B6 Idd10 and NOD.B6 Idd10/18 mice. Popliteal lymph nodes revealed also comparable Treg percentages (S8 Fig). Thus, together with the improved function of the B6 over the NOD Cd101 allele the increased T1D frequency in CD101−/− NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10/18 mice as compared to their CD101+/+ counterparts is associated with a reduction of Tregs in pancreatic lymph nodes suggesting that CD101 acts locally at the site of T cell priming.
We have recently reported a correlation between the CD101-dependent distribution of Gr1-positive myeloid cells and the susceptibility to T1D [31]. Thus, we evaluated the distribution of different Gr1-expressing myeloid cell populations in CD101-expressing and CD101-deficient NOD.B6 Idd10 and NOD.B6 Idd10/18 mice. Similar to B6 CD101−/− mice [31], CD101−/− NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10/18 mice have a reduction of Gr1-positive cells in the bone marrow compared to CD101-expressing congenic controls (S9A Fig). A similar tendency was also observed in the spleen (S9B Fig). To further characterize the Gr1-positive population, we used additional markers and also investigated its distribution in the pancreatic islets and pancreatic lymph nodes. As previously reported for NOD mice [4] we observed myeloid immune cells in the infiltrates of the pancreatic islets of CD101-expressing NOD.B6 Idd10 and CD101-deficient NOD.B6 Idd10 mice. Strikingly, however, cells infiltrating the pancreatic islets of CD101−/− NOD.B6 Idd10 mice contained substantially smaller proportions of both CD11b+ Gr1− and CD11bhigh Gr1+ cells than their CD101-expressing counterparts (Fig 5A). In contrast, the CD11b+ Gr1− population was detected equivalently in the CD101−/− and CD101+/+ NOD.B6 Idd10 strains in the pancreatic lymph nodes (Fig 5B). The distribution of CD11c and F4/80 within this CD11b-positive subset in the pancreatic lymph nodes was also comparable in these two strains (S10 Fig). Similar to the islet infiltrating cells, the proportion of CD11bhigh Gr1+ cells, consisting of neutrophils and myeloid-derived suppressor cells (MDSCs) [41–43], were significantly reduced in the pancreatic lymph nodes of CD101−/− as compared to CD101+/+ NOD.B6 Idd10 mice (Fig 5B).
To study the function of CD101 on T cells in the T1D model we purified T cells from the spleens of CD101−/− and CD101-expressing NOD.B6 Idd10 mice and transferred the CD4- and CD8-positive T cell population into lymphopenic NOD scid recipients. The combined transfer of CD4- and CD8-positive T lymphocytes increased T1D frequency in recipients of donor T cells originating from CD101−/− NOD.B6 Idd10 mice compared to donor T cells originating from CD101-expressing NOD.B6 Idd10 controls (Fig 6A). The increased frequency was modest, just reaching significance (p = 0.038). The higher frequency of T1D was accompanied by a more rapid expansion of T cells from CD101−/− NOD.B6 Idd10 donors (Fig 6B and 6C). Furthermore, similar as observed in CD101−/− NOD.B6 Idd10 mice (Fig 4E), less FoxP3+ Tregs in relation to CD4-positive T cells accumulated in the pancreatic lymph nodes of NOD scid recipients from CD101−/− NOD.B6 Idd10 donors than from CD101-expressing NOD.B6 Idd10 donors (Fig 6D). Thus, these data support the hypothesis that CD101 expression on T cells reduces effector T cell expansion, a conclusion also reached in our T cell transfer colitis studies [35].
We had observed that CD101-expressing myeloid cells decreased upon transfer of naïve CD4+ T cells [35]. Thus, we evaluated the composition of myeloid cells and the distribution of CD101 expression in pancreatic islets of NOD scid recipient mice upon combined CD4+/CD8+ T cell transfer from CD101-expressing and CD101-deficient NOD.B6 Idd10 donors. Interestingly, significantly fewer Gr1-positive myeloid cells derived from NOD scid recipients accumulated in the pancreatic lymph nodes upon T cell transfer from CD101−/− NOD.B6 Idd10 donors as compared to T cells transferred from CD101+/+ NOD.B6 Idd10 donors (Fig 7A and 7B). Furthermore, when CD101−/− NOD.B6 Idd10 donor T cells were transferred fewer of the accumulating Gr1-positive myeloid cells expressed CD101 (Fig 7A and 7C) and the Gr1-positive myeloid cells produced less TGF-β (Fig 7D) and IL-10 (Fig 7E) than their CD101-expressing counterparts, as observed previously for CD101-expressing myeloid cells in the gut [35].
To further characterize the role of the B6 Cd101 allele on myeloid cells for the protection from T1D, we generated NOD.B6 Idd10 scid mice and assessed T1D incidence in these recipients as compared to NOD scid recipients upon T cell transfer from CD101+/+ NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10 donors. The development of T1D was significantly ameliorated (p = 0.007) when T cells from CD101-expressing NOD.B6 Idd10 donors were transferred into NOD.B6 Idd10 scid as compared to NOD scid recipients indicating that in the presence of the CD101 protein encoded by the B6 Cd101 allele in the T cell compartment, the status of the Cd101 allele expressed by myeloid cells determines the level of protection from T1D (Fig 8A). When CD101−/− NOD.B6 Idd10 donor T cells were used in the adoptive transfer, fewer NOD.B6 Idd10 scid recipients than NOD scid recipients developed T1D but the difference was not significant (p = 0.2; Fig 8B). In contrast to the results in Fig 6A where CD101−/− NOD.B6 Idd10 T cells mediated a modest increase in T1D upon transfer into NOD scid recipients as compared to CD101+/+ NOD.B6 Idd10 T cells, no difference (p = 0.35) was observed in a repeat of the same transfer combination (Fig 8A and 8B). A consideration of both results supports the conclusion that the effect of CD101 expression in T cells for influencing T1D progression is marginal when CD101 expression in the myeloid compartment is encoded by the NOD Cd101 allele rather than the B6 Cd101 allele. Overall our data support the hypothesis that the expression of the B6 CD101 molecule is not only important for limiting the aggressiveness of diabetogenic T cells, but also promotes the function of disease-limiting myeloid cell subsets.
Allelic variations within Cd101 have been previously associated with susceptibility to T1D [31, 37]. Here, we provide further evidence that Cd101 is Idd10 as the genetic deletion of CD101 within the introgressed B6 Idd10 region abolishes protection from T1D. Significantly reduced Treg frequencies in the pancreatic lymph nodes in these newly generated CD101−/− NOD.B6 Idd10 mice and an enhanced T cell expansion upon adoptive transfer accompanied the loss of protection from T1D observed in the CD101−/− NOD.B6 Idd10 strain. Furthermore, the proportion of CD101-expressing CD11b-positive myeloid cells in recipient scid mice was reduced upon transfer of CD101−/− donor T cells correlating with a reduced IL-10 and TGF-β mRNA. Myeloid cells from CD101-expressing NOD.B6 Idd10 mice also accumulated more frequently in pancreatic tissues than myeloid cells from CD101−/− NOD.B6 Idd10 mice. Thus, the interplay of CD101-expressing Tregs [30, 35] with CD101-positive myeloid cells appears to perpetuate an anti-inflammatory cytokine profile and limits the onset of T1D.
CD101 exhibits intrinsic effects on Treg differentiation and function and promotes the production of IL-10 by myeloid cells [35]. Here, we confirmed the pivotal role of simultaneous expression of the B6 Cd101 allele within the myeloid cell and T lymphocyte compartment for the most complete protection from T1D. Based on similar T1D frequencies in NOD and CD101−/− NOD.B6 Idd10 mice we hypothesize that the NOD CD101 allotype present in NOD mice may not function properly due to the 10 amino acid differences from the B6 CD101 protein [37], and thus, resembles the CD101 knockout. Furthermore, the B6 Idd10 allele enhances the expression of CD101 protein on Tregs and Gr1-expressing myeloid cells [31]. A greater understanding of the signaling and cellular interactions mediated by the CD101 protein is required to determine how the B6 and NOD Cd101 alleles mediate, or fail to mediate, their functions.
As Gr1-expressing myeloid cells in the bone marrow are precursors for multiple myeloid lineages in the periphery, we investigated the myeloid cell composition and the distribution of myeloid surface markers in pancreatic tissues and lymph nodes in more detail. However, with the exception of CD11b, we did not detect significant differences in the expression of F4/80, Ly6C or Ly6G between Gr1-expressing myeloid cells from CD101-expressing and CD101-deficient NOD.B6 Idd10 mice. As significantly more Gr1-positive CD11b-positive myeloid cells accumulated in CD101+/+ NOD.B6 Idd10 mice than in CD101−/− NOD.B6 Idd10 mice, CD101 might promote the maturation and function of this myeloid subset. In accordance with the anti-inflammatory cytokine profile, these Gr1-expressing cells might reflect MDSCs, which have been reported to suppress T1D [44]. However, the origin of this myeloid subset and its classification into granulocytes or MDSCs need to be assessed. Alternatively, these Gr1-expressing myeloid cells might consist of a plastic MDSC and neutrophil mixture which exert an anti-inflammatory cytokine profile dependent on the expression of CD101. Alternatively, CD101 might represent a functional marker separating inflammatory neutrophils from immunoregulatory MDSCs. Further studies are needed to delineate whether alterations in the development and/or generation of additional myeloid subsets are promoted by CD101-deficiency since NOD mice have been reported to have multiple alterations in DC subsets compared to B6 mice, for example [45, 46].
Our observation that CD101 affects Gr1-expressing cells is also of clinical relevance as reduced neutrophil counts in the peripheral blood and enhanced neutrophil activity have been reported in T1D patients [47–49]. Neutrophil infiltration and neutrophil extracellular trap formation are also detected in the islets of NOD mice as early as two weeks after birth, well before the onset of overt diabetes. The blockade of neutrophil activities or neutrophil depletion at these early stages reduces the development of insulitis and diabetes in NOD mice [4]. Thus, the reduced neutrophil counts in the periphery of CD101−/− NOD.B6 Idd10 mice might be a consequence of one or more of the following reasons: 1) impairment in the output of neutrophils from the bone marrow and/or the differentiation of neutrophils; 2) increase in peripheral consumption/destruction; 3) tissue sequestration. Based on our previous studies [31], we suspect that a bone marrow defect affects either the egress or the generation of myeloid cell precursors.
Decreased T1D frequency has been shown to be associated with enhanced CD101 expression on splenic Tregs of congenic NOD.B6 Idd10 mice compared to parental NOD mice [31]. In the current study, we observed that significantly more Tregs accumulated in the pancreatic lymph nodes of CD101+/+ NOD.B6 Idd10 compared to CD101−/− NOD.B6 Idd10 mice, confirming the positive effect of CD101 on Treg differentiation and function [30, 35, 50]. Thus, although NOD mice, similar to T1D patients, do not have a primary deficit in Treg percentages or numbers compared to other reference strains [17, 18, 51–53], our data imply that the B6 Cd101 allele primarily affects Tregs within the T cell compartment. These animal studies are also in line with a recent report on distinct T1D patient cohorts raising the possibility of CD101 being a susceptibility gene for human T1D [36]. In addition, since IL-2 administration acts via pancreatic Tregs [54] and CD101 sensitizes Tregs to IL-2 signals [35], the accumulation of CD101+ Tregs in NOD mice carrying the B6 Cd101 allele likely contributes to the protection from T1D.
CD101 expression restrains the accumulation and expansion of diabetogenic CD4- and CD8-positive T lymphocytes and reduces T1D frequency in lymphopenic recipient mice upon mixed CD4/CD8 T cell transfers from CD101+/+ NOD.B6 Idd10 mice. In particular, CD8-positive T cells might be interesting to study further since recent reports claimed altered functions between CD101-expressing CD8-positive T lymphocytes and their CD101-negative counterparts [55, 56]. Thus, in summary, our data clearly indicate that CD101 promotes the accumulation of anti-inflammatory lymphoid and myeloid cells and slows or halts disease in an autoimmune-prone background when sufficiently expressed.
The experiments were conducted according to the Institutional Animal Care and Use Committee guidelines of the Cincinnati Children’s Hospital (IACUC protocol number Protocol 8D02011) and approved by the Animal Welfare Committee of the local government (Regierung von Mittelfranken, Ansbach, Germany; protocol: 54–2532.1-30/10). Daily inspections were performed to minimize animal suffering. Mice with signs of discomfort or disease were euthanized immediately by C02 and cervical dislocation.
NOD/MrkTac (NOD) and NOD scid mice were obtained from Taconic Farms (Germantown, NY, USA). The development of the NOD.B6 Idd10 (N16) (Taconic line 3538) and NOD.B6 Idd10/18 (N10) (Taconic line 7754) strains were described previously [31, 39]. NOD.B6 Idd10 and NOD scid mice were intercrossed and F2 mice homozygous for the Idd10 and scid-containing regions selectively bred to develop the NOD.B6 Idd10 scid strain. B6 CD101−/− mice [31] were backcrossed onto the NOD background to develop the CD101−/− NOD.B6 Idd10 (N11) and CD101−/− NOD.B6 Idd10/18 strains (N10). Polymorphic markers near and within the Idd10 and Idd18 regions were used to define recombination events as similar as possible to the boundaries of these regions (S1 Table) as previously defined. The CD101−/− NOD.B6 Idd10 and CD101−/− NOD.B6 Idd10/18 strains were free of B6-derived genetic segments outside of the selected areas as defined by screening with a 1449 polymorphic marker panel as described previously [31]. All mice were raised and kept in a specific pathogen-free environment and used at 3–15 weeks of age for cellular and molecular analyses and for T1D frequency studies until the age of 200 days. The appropriate institutional review committee approved the T1D frequency studies performed at Taconic Farms.
Mice were euthanized and pancreata were perfused with a 1.5 mg/mL solution of collagenase P (Roche Molecular Biochemicals, Mannheim, Germany), dissected from surrounding tissues and cut into small pieces. The digestion buffer was supplemented with 1 mM PMSF, 100 μM leupeptin and 1 μM pepstatin A (Sigma-Aldrich, Taufkirchen, Germany). Pancreata were digested at 37°C for 10 min in a shaking water bath. The digestion was stopped by adding HBSS containing 5% FCS. The tissue suspension was washed three times and centrifuged through a discontinuous Ficoll gradient (23, 20.5 and 11%; Sigma-Aldrich, Taufkirchen, Germany) at room temperature. The purified islets were disrupted by adding 1 mL of cell dissociation buffer (GIBCO/Thermo Fisher Scientific, Waltham, MA, USA) for 10 minutes at 37°C. The obtained cells were washed, resuspended and used for analyses.
5x107 spleen cells enriched in TCRβ-positive T cells (consisting of about 2/3 CD4-positive T cells and of about 1/3 CD8-positive T cells) from 6-8-week-old donor female mice were transferred intraperitoneally into 6-8-week-old NOD scid or NOD.B6 Idd10 scid recipients. B cells, NK cells, DCs, granulocytes or macrophages were depleted using Auto-MACS (Miltenyi Biotec, Bergisch Gladbach, Germany) and PE- or APC-fluorescence coupled beads against CD19, NKp46, Gr1, CD11c or F4/80 from NOD mice before transfer following the manufacturer's instructions with purity control by FACS.
For cell division studies donor T cells were labeled with 5 μM CFSE prior to transfer according to the manufacturer's instructions (Molecular Probes/Thermo Fisher Scientific, Waltham, MA, USA and BD Pharmingen/BD Biosciences, San Diego, CA, USA).
Single-cell suspensions were prepared from the spleen, lymph nodes, pancreas and bone marrow. Red blood cells were not removed. Cell-surface expression of CD45.1 (clone A20), CD45.2 (clone 104), CD11c (clone N418), CD11b (clone M1/70), Ly6C (clone HK1.4), Ly6G/Gr1 (clone RB6-8C5), F4/80 (clone BM8), B220 (clone RA3-6B2), the β-chain of the TCR (clone H57-597), CD3 (clones 145-2C11 and 17A2), CD4 (clone GK1.5) and CD8a (clone 53–6.7) was detected using fluorescently labeled mAbs obtained from eBioscience (San Diego, CA, USA). PE-labeled anti-CD101 (clone 307707) was obtained from R&D Systems. Intracellular Foxp3 was detected with a staining kit following the manufacturer´s instructions (eBioscience). Cells were analyzed on an LSR II or a BD FACS Canto II (BD Biosciences, San Diego, CA, USA) with FlowJo software (Tree Star, Ashland, OR, USA). Gr1-expressing cells were purified on a FACS Aria II (BD Biosciences, Franklin Lakes, NJ, USA) (purity of > 98%).
All diabetes cumulative frequency studies were conducted using female mice. The presence of T1D was tested every 14 d beginning at 84 d of age by the detection of urinary glucose >500 mg/dl using Diastix (Miles, Elkhart, IN, USA). Overt diabetes was confirmed by a blood sugar level of >200 mg/dl. Studies were terminated at 196 d of age. Kaplan–Meier survival curves were plotted for each mouse strain, and these were compared using the log rank test (Prism4 software; GraphPad).
Pancreatic tissue was fixed in 10% buffered formalin, embedded in paraffin, and cut into 5 μm thick sections. Pancreas sections were deparaffinized, stained with H&E by the Department of Pathology and the Medical Department I of the FAU Erlangen-Nürnberg, and evaluated microscopically in a double-blinded manner. H&E–stained sections were scored for insulitis. At least 10 islets per mouse present on two or three non-adjacent pancreas sections were scored as either 0, no infiltration; 1, peri-insulitis; 2, mild-invasive insulitis; or 3, severe invasive insulitis. The average score of each pancreas was calculated and used for statistical analysis.
cDNA was synthesized using a High-capacity cDNA Reverse Transcription Kit (Applied Biosystems/Thermo Fisher Scientific, Waltham, MA, USA) following the manufacturer’s instructions. Quantitative (q) -PCRs were performed as described [33] using specific primers (Thermo Scientific) and pre-designed probes (Roche, Basel, Switzerland): TGF-β (forward: 5´- gtggtgtccccacacagg-3´; reverse: 5´-ccagggctgtaaccacttg-3´) and IL-10 (forward: 5´- cagagccacatgctcctaga-3´; reverse: 5´-gtccagctggtcctttgttt-3´). Gene expression was calculated relative to the house keeping gene HPRT (forward: 5´-tcctcctcagaccgctttt-3´; reverse: 5´-cctggttcatcatcgctaatc-3´ or Applied Biosystems assay Mm00446968_m1) using the ΔΔCt algorithm.
Samples were analyzed for normal distribution by a Kolmogorov-Smirnov test. According to the results, statistical significance in normal distributed samples were analyzed by one-way ANOVA with posthoc test (Bonferroni) and Student’s t-test, and samples failing the normal distribution test by Kruscal-Wallis Test with posthoc (Dunn’s multiple comparison) or Mann-Whitney U test as indicated in the respective experiments. A sample size of at least three (n = 3) was used for each sample group in a given experiment, and a p value of 5% (*; p 0.05), 1% (**; p 0.01), 0.1% (***; p 0.001) or 0.01% (****; p 0.0001) was considered significant to accept the alternate hypothesis. GraphPad Prism software was used for statistical analysis.
|
10.1371/journal.ppat.1003944 | Inactivation of the Host Lipin Gene Accelerates RNA Virus Replication through Viral Exploitation of the Expanded Endoplasmic Reticulum Membrane | RNA viruses take advantage of cellular resources, such as membranes and lipids, to assemble viral replicase complexes (VRCs) that drive viral replication. The host lipins (phosphatidate phosphatases) are particularly interesting because these proteins play key roles in cellular decisions about membrane biogenesis versus lipid storage. Therefore, we examined the relationship between host lipins and tombusviruses, based on yeast model host. We show that deletion of PAH1 (phosphatidic acid phosphohydrolase), which is the single yeast homolog of the lipin gene family of phosphatidate phosphatases, whose inactivation is responsible for proliferation and expansion of the endoplasmic reticulum (ER) membrane, facilitates robust RNA virus replication in yeast. We document increased tombusvirus replicase activity in pah1Δ yeast due to the efficient assembly of VRCs. We show that the ER membranes generated in pah1Δ yeast is efficiently subverted by this RNA virus, thus emphasizing the connection between host lipins and RNA viruses. Thus, instead of utilizing the peroxisomal membranes as observed in wt yeast and plants, TBSV readily switches to the vastly expanded ER membranes in lipin-deficient cells to build VRCs and support increased level of viral replication. Over-expression of the Arabidopsis Pah2p in Nicotiana benthamiana decreased tombusvirus accumulation, validating that our findings are also relevant in a plant host. Over-expression of AtPah2p also inhibited the ER-based replication of another plant RNA virus, suggesting that the role of lipins in RNA virus replication might include several more eukaryotic viruses.
| Genetic diseases alter cellular pathways and they likely influence pathogen-host interactions as well. To test the relationship between a key cellular gene, whose mutation causes genetic diseases, and a pathogen, the authors have chosen the cellular lipins. Lipins are involved in a key cellular decision on using lipids for membrane biogenesis or for storage. Spontaneous mutations in the LIPIN1 gene in mammals, which cause impaired lipin-1 function, contribute to common metabolic dysregulation and several major diseases, such as obesity, hyperinsulinemia, type 2 diabetes, fatty liver distrophy and hypertension. In this work, the authors tested if tomato bushy stunt virus (TBSV), which, similar to many (+)RNA viruses, depends on host membrane biogenesis, is affected by deletion of the single lipin gene (PAH1) in yeast model host. They show that pah1Δ yeast supports increased replication of TBSV. They demonstrate that TBSV takes advantage of the expanded ER membranes in lipin-deficient yeast to efficiently assemble viral replicase complexes. Their findings suggest possible positive effect of a genetic disease caused by mutation on the replication of an infectious agent.
| Positive-stranded (+)RNA viruses are important and emerging human, animal and plant pathogens. These viruses utilize cellular membranes and lipids during replication to build viral replicase complexes (VRCs) [1]–[4]. The subverted subcellular membranes are proposed to provide critical lipid or protein cofactors to regulate the function of the viral replicase, serve as scaffolds for VRC assembly, provide protection of the viral RNA against cellular nucleases, prevent recognition by the host antiviral surveillance system, or facilitate the targeting of the viral replication proteins to a particular microdomain in the membrane [1]–[9]. (+)RNA viruses induce membrane proliferation that requires new lipid biosynthesis as shown by several genome-wide screens, which identified lipid biosynthesis/metabolism genes [10]–[15]. Accordingly, several examples of virus-induced modification of cellular lipid metabolism and changes in lipid composition of membranes during virus replication are documented in the scientific literature [16]–[20].
Tombusviruses, such as tomato bushy stunt virus (TBSV) (Table S1), are among the best-characterized viruses [3], [12], [13], [21]–[25]. They belong to supergroup 2 (+)RNA viruses that include animal flaviviruses, and pestiviruses and plant luteoviruses, carmoviruses, and others. Tombusviruses code for five proteins including two replication proteins, termed p33 and p92pol [26]–[28]. Both p33 and p92pol are translated from the genomic (g)RNA and p92pol is the result of translational readthrough of the p33 stop codon [28], [29]. p92pol is the RNA-dependent RNA polymerase [30]–[32], while p33 is an RNA chaperone playing a role in RNA template selection and recruitment and in the VRC assembly [32]–[37].
(+)RNA viruses either target existing subcellular membranes or they extensively remodel membranes to support the VRC assembly. Interestingly, tombusviruses can utilize pre-existing membranes, but also remodel subcellular membranes by forming multivesicular body-like structures in infected cells [5], [6], [38], [39]. Tombusviruses are useful for membrane remodeling studies, because they can utilize peroxisomal membranes [e.g., TBSV and the closely related cucumber necrosis virus (CNV)] [5], [39], or they can switch to endoplasmic reticulum (ER) membranes in the absence of peroxisomes [40], [41], or replicate in ER or mitochondrial membranes in vitro [42].
Genetic diseases can alter critical cellular processes, which might affect pathogens that have to take advantage of cellular resources. The host lipins are particularly interesting because these proteins play key roles in cellular decisions about membrane biogenesis versus lipid storage [43]–[45]. Spontaneous mutations in the LIPIN1 gene in mammals, which cause impaired lipin-1 function, contribute to common metabolic dysregulation and several major diseases, such as obesity, hyperinsulinemia, type 2 diabetes, fatty liver distrophy and hypertension [44], [46], [47].
The yeast PAH1 (phosphatidic acid phosphohydrolase) gene is the homolog of the mammalian fat-regulating protein Lipin-1 [43], [45], [48]. Like the three mammalian lipin genes, the single copy yeast PAH1 codes for a phosphatidate phosphatase (PAP), which dephosphorylates phosphatidic acid (PA), yielding diacylglycerol (DAG) (Fig. 1A). Pah1p is involved in synthesis of DAG and triacylglycerol (TAG) storage lipids, and in the absence of PAH1, the ER/nuclear membrane expands considerably and the total phospholipid content of the cell increases by ∼2-fold [49], [50]. Thus, Pah1p sits at the crossroads between membrane biogenesis and lipid storage (i.e., the decision to store fat or build membranes) [45]. The mammalian or plant lipins can complement Pah1p function in yeast, demonstrating the functional similarity among these enzymes [46], [51]. Pah1p is the only yeast PAP protein involved in the synthesis of TAG and the regulation of phospholipid biosynthesis [52].
Since (+)RNA viruses likely depend on membrane biogenesis, we examined the relationship between the host lipins and tombusviruses, based on yeast model host. In this paper, we document that deletion of the yeast lipin gene, PAH1, whose inactivation is responsible for proliferation and enlargement of the ER membrane, facilitates robust RNA virus replication in yeast. Thus, surprisingly, a host gene whose homologs are involved in genetic diseases in humans, greatly affects virus replication.
Since TBSV and the closely related CNV (Table S1) induce membrane proliferation and they replicate by utilizing peroxisomal membranes for VRC assembly in vivo [5], [40]–[42] and ER membranes in vitro [42], we tested if deletion of PAH1 gene, which leads to ER membrane enlargement and proliferation [43], [45], [46], could alter TBSV and CNV replication in yeast cells. We found that the TBSV replicon (rep)RNA accumulated to ∼7-fold higher in the presence of TBSV replication proteins (Fig. 1B, lanes 4–6 versus 1–3) and ∼2.5-fold higher levels in case of CNV (lanes 10–12 versus 7–9) in pah1Δ yeast. These data suggest that the enlarged ER might provide favorable microenvironment for TBSV and CNV replication or tombusviruses might be able to take advantage of the increased phospholipid content of the cell. Interestingly, the levels of p33 and p92pol replication proteins were increased in pah1Δ yeast (Fig. 1B).
Moreover, expression of wt Pah1p protein in pah1Δ yeast had moderate inhibitory effect on TBSV replication (Fig. 1C, lanes 4–6) and over-expression Pah1p also inhibited TBSV replication (Fig. 1D, lanes 4–6). This moderate inhibitory effect by wt Pah1p could be due to phosphorylation and inactivation of the enzymatic function of the over-expressed Pah1p in yeast [50], [53]. Therefore, we also expressed/over-expressed a constitutively active, phosphorylation-deficient mutant of Pah1p, which indeed led to more pronounced inhibition of TBSV repRNA accumulation in pah1Δ or wt yeasts (by ∼40-to-50%, see mutant Pah1-7A containing alanine substitutions for all seven phosphorylation sites, lanes 7–9, Fig. 1C–D).
In addition, we tested TBSV accumulation in nem1Δspo7Δ yeast, which lacks the ER-associated phosphatase complex needed for ER association, dephosphorylation, and activation of Pah1p [50]. As expected, TBSV replication increased by ∼3-fold in nem1Δspo7Δ yeast (Fig. 1E, lanes 4–6 versus 1–3), further supporting the major role of Pah1p in TBSV replication.
Finally, we also tested the effect of over-expression of Dgk1p diacylglycerol kinase, which catalyzes the production of PA from DAG, the opposite reaction with Pah1p (Fig. 1A), on TBSV replication. Overproduction of the ER-localized Dgk1p induces the enlargement of ER-like membranes in yeast [54], [55]. We found that overproduction of Dgk1p in yeast led to increased TBSV replication (Fig. S1, lanes 4–6). Altogether, the above data support that the enlarged ER or the increased phospholipid content of the cell is a major advantage for TBSV and CNV replication.
To test if the high level of TBSV accumulation is due to increased TBSV repRNA replication, we measured the level of TBSV repRNA accumulation at various time points after induction of replication in pah1Δnem1Δ or pah1Δ yeast. These experiments revealed that TBSV repRNA accumulated to 2.5-to-4-fold higher level even at the early time points (Fig. 2A–B, 5 and 8 hour time points). Similarly, TBSV repRNA accumulation was ∼5-fold higher at an early time point in pah1Δ yeast (Fig. S2). These data suggest that more robust TBSV replication occurs earlier in pah1Δnem1Δ and pah1Δ yeasts than in the wt yeast, indicating that VRCs might be able to assemble faster in the mutant yeast cells.
Testing the replicase activity in the isolated membrane fraction containing the viral replicase/viral RNA complex from pah1Δnem1Δ (Fig. 2D–F) or pah1Δ yeast (Fig. S3A–B) also showed ∼4-to-5-fold increase over the replicase activity observed with the isolated membrane fraction from wt yeast at both early and late time points. The isolated membrane fractions were adjusted to contain comparable amount of the p92pol replication protein (Fig. 2E–F, S3B), thus the increased in vitro replicase activity in the samples from pah1Δnem1Δ or pah1Δ yeasts indicates that the tombusvirus replicase in pah1Δnem1Δ or pah1Δ yeasts is more active than in wt yeast. Interestingly, unlike p92pol, the amounts of the tombusvirus p33 replication protein, the Sec61p ER resident protein and the Ssa1p Hsp70 chaperone, which is co-opted for TBSV replication, all increased in the isolated membrane fraction from pah1Δnem1Δ (Fig. 2D–E) or pah1Δ yeast (Fig. S3B). The presence of elevated amounts of p33 [56] and Ssa1p [37], [57], [58] has been shown to increase TBSV replication, which is in agreement with the increased in vitro replicase activity in the isolated membrane fraction from pah1Δnem1Δ or pah1Δ yeasts.
Since p33 replication protein is an integral membrane protein [39], [58], the observation of increased p33 level in the isolated membrane fraction from pah1Δnem1Δ yeast suggests that p33 might be more stable in the mutant yeast than in the wt yeast. Indeed, estimation of the half-life of p33 revealed ∼4-fold increased stability of p33 in pah1Δnem1Δ yeast in comparison with wt yeast (Fig. 2G).
Based on the above data, it is possible that TBSV VRCs could assemble more efficiently in pah1Δnem1Δ yeast due to the presence of extended ER membranes and abundant amounts of phospholipids. To test this possibility, we utilized an in vitro tombusvirus VRC assembly assay based on purified recombinant replication proteins and cell-free extracts (CFE) obtained from pah1Δnem1Δ or wt yeast (Fig. 3A). In this assay, the tombusvirus (+)repRNA performs one cycle of asymmetrical replication supported by the tombusvirus VRCs assembled in vitro [37], [59]. Since we use recombinant viral proteins and repRNA in this assay, we can make sure that only the CFEs, involving the cellular membranes and possibly host factors, are different. The yeast CFEs were adjusted to contain comparable amounts of Pgk1p (a cytosolic protein) (Fig. 3B) and total proteins.
Interestingly, the in vitro RNA replication supported by CFE was ∼7-fold higher when assembled in CFE obtained from pah1Δnem1Δ yeast than from wt yeast (Fig. 3B, lanes 5–8 versus 1–4). These data strongly suggest that the tombusvirus replicase assembly in the CFE derived from pah1Δnem1Δ is more efficient than in the CFE from wt yeast.
To test if TBSV replication indeed includes a full cycle in the CFE from pah1Δnem1Δ yeast, we analyzed the (−) and (+)-strand RNAs in the CFEs (Fig. 3C). The membrane-fraction of the CFE at the end of the replication assay contains both single-stranded (ss)RNA [representing the newly made (+)-stranded progeny RNA] and dsRNA [representing the annealed (−) and (+)RNAs]. In addition, the soluble fraction contains the newly released (+)RNAs from the membrane-bound VRCs. We found that the amounts of both ssRNA and dsRNA were ∼2-fold higher in the membrane fraction and the ssRNA was ∼2-fold higher in the soluble fraction of the CFE prepared from pah1Δnem1Δ yeast than from the wt yeast (Fig. 3D). These results suggest that the CFE obtained from pah1Δnem1Δ yeast performs all the replication steps more efficiently than the CFE prepared from wt yeast cells.
To test if the membrane-fraction of the CFE from pah1Δnem1Δ yeast is important for the enhanced TBSV RNA replication, we separated the soluble and membrane fractions of the CFEs prepared from pah1Δnem1Δ and wt yeasts and then used various combinations of these fractions for in vitro TBSV replication (Fig. S4A). Interestingly, the CFE containing the mixture of the membrane fraction from pah1Δnem1Δ yeast and the soluble fraction from the wt yeast supported almost as efficient in vitro TBSV replication as the CFE consisting of both fractions from pah1Δnem1Δ yeast (Fig. S4B, compare lanes 3–4 and 7–8 with 1–2). Altogether, it seems that the membrane fraction when derived from pah1Δnem1Δ yeast was able to support ∼5-fold higher TBSV replication than the membrane fraction from wt yeast in the CFE-based replication assay. This is not surprising since PAH1 deletion is expected to dramatically change the ER membranes that could be utilized by TBSV for assembly of the VRCs.
To obtain direct evidence that the expanded ER structures and membranes in pah1Δnem1Δ yeast are utilized efficiently for TBSV replication, we took advantage of an isolated ER-based tombusvirus replication assay [42]. In this assay, the tombusvirus (+)repRNA can also perform full replication supported by the tombusvirus VRCs assembled in the ER membrane (Fig. 4A) [42]. As expected, the isolated ER preparations contained the Sec61p ER-resident protein, while lacked the cytosolic Pgk1p and the peroxisomal (Fox3p) proteins (Fig. 4B, lanes 1–3 versus 4–6, representing the total CFE). Larger amounts of the p33 replication protein were associated with the ER fraction obtained from pah1Δnem1Δ or pah1Δ yeasts than from the wt yeast (Fig. 4B, lanes 2 and 3 versus 1), suggesting that p33 utilized the ER membranes more efficiently in yeast lacking the PAH1 gene.
When we used similar amounts of isolated ER membranes (based on adjusted cellular Sec61p level) for TBSV replication, we observed that the replication of TBSV RNA was 2-to-4-fold higher in ER preparations obtained from pah1Δnem1Δ or pah1Δ yeasts than from wt yeast (compare lanes 5–6 with 4, Fig. 4C). These data suggest that the ER membranes derived from yeast lacking PAH1 are more efficiently utilized by the tombusvirus replicase than the ER membrane from wt yeast. On the contrary, when we adjusted the p33 levels in the isolated ER preparations, then we observed comparable levels of in vitro tombusvirus replication in ER membranes from all three yeast strains (Fig. 4C, lanes 1–3). This indicates that the relative activity of the tombusvirus replication protein expressed in these yeast strains is similar. Therefore, the higher activity of tombusvirus replicase in pah1Δnem1Δ or pah1Δ yeasts are likely due to the more efficient assembly of the tombusvirus VRCs, resulting in larger number of replicationally active VRCs than in wt yeast.
To test how efficiently the p33 and p92 replication proteins can utilize the expanded ER membranes in pah1Δnem1Δ yeast versus the wt yeast, we performed confocal laser microscopy with fluorescently tagged tombusvirus p33 replication protein. When we looked at the localization of YFP-p33 at an early time point (4 hours), we observed that a large portion of YFP-p33 was still cytosolic and a small number of punctate structures were forming in wt yeast (Fig. 5A). At the same time point, the YFP-p33 localized efficiently in the ER membranes in pah1Δnem1Δ yeast and only a smaller fraction of YFP-p33 showed cytosolic localization pattern (Fig. 5B). At the 6 hr time point, the YFP-p33 also showed mostly punctate pattern in wt yeast (Fig. 5C). A fraction of YFP-p33 was present in the ER membrane or was diffused in the cytosol, while other YFP-p33 molecules were likely localized in the peroxisomal membranes in wt yeast. In contrast, most YFP-p33 localized in the ER membranes in pah1Δnem1Δ yeast at the 6 h time point (Fig. 5D). At the late 24 h time point, as expected, most of the p33 was localized in punctate structures separate from the ER membranes [likely representing the peroxisomal membranes as shown previously [39]–[41]], although a small fraction of p33 did co-localize with the ER in wt yeast cells (Fig. 5E). In contrast, most p33 is localized in the ER membranes, forming large elongated structures in pah1Δnem1Δ yeast (Fig. 5F). Altogether, these data showed that p33 is rapidly localized to the expanded ER membranes in pah1Δnem1Δ yeast, while the localization of p33 from the cytosol to membranes is slower in wt yeast, and likely involves the peroxisomal membranes as shown before. It seems that a small fraction of p33 does target the ER membrane even in the wt yeast cells. Therefore, we conclude that the tombusvirus replication proteins efficiently exploit the expanded ER membranes in pah1Δnem1Δ yeast cells.
To test if peroxisomes are available for TBSV replication in pah1Δnem1Δ yeast cells, we used CFP-tagged Pex13p peroxisome membrane marker protein [41]. The distribution of Pex13p showed the characteristic punctate structures in pah1Δnem1Δ yeast cells (Fig. 6). However, the co-localization of Pex13p and the tombusvirus p33 differed in wt and in pah1Δnem1Δ yeast cells (Fig. 6A–B). While Pex13p showed remarkably good co-localization with the tombusvirus p33 in wt yeast (67% co-localization and 28% partial co-localization of Pex13p and p33 puncta and only 15% of p33 puncta were not co-localized with Pex13p puncta; Fig. 6A), the tombusvirus p33 was largely present in different compartment than Pex13p in pah1Δnem1Δ yeast cells (Fig. 6B). However, a small fraction of p33 (less than 10%) was completely or partially co-localized with Pex13p in pah1Δnem1Δ yeast cells, suggesting that peroxisome membranes are still utilized by tombusviruses in the mutant yeast. In addition, both Pex13p and the tombusvirus p33 were present in several large foci in wt yeast, indicative of membrane proliferation and peroxisome aggregation induced by the tombusvirus p33 [40], [41], [60]. In contrast, Pex13p was mostly present in several smaller foci in pah1Δnem1Δ yeast cells, suggesting lack of membrane proliferation and peroxisome aggregation (Fig. 6B). Altogether, these data is compatible with the model that the expanded ER membranes are more efficiently utilized by the tombusvirus replication protein than the peroxisomal membranes in pah1Δnem1Δ yeast cells, although peroxisomes are also available in these cells.
To test if the VRC assembly could indeed be more efficient in the pah1Δnem1Δ yeast cells, we used the CFE-based assay to assemble the membrane-bound VRCs [37]. After the assembly and activation of the VRCs in the CFEs, we solubilized and affinity-purified the tombusvirus replicase and tested the replicase activity on added RNA template (Fig. 7A). This assay depends on the efficiency of VRC assembly and the activation of the tombusvirus p92pol replication protein, which is originally inactive when expressed in E. coli, yeast or plants [31]. The activation of p92pol replication protein occurs in the membrane-bound VRCs and depends on many factors, including p33, cis-acting elements in the viral (+)RNA, host factors and cellular membranes [30], [32], [37], [61], [62]. We found that the CFE prepared from pah1Δnem1Δ yeast cells resulted in ∼3-fold higher replicase activity in vitro (Fig. 7B), suggesting that the VRC assembly and activation of p92pol replication protein is more efficient than that in the similar CFE from wt yeast.
To study if other RNA viruses that replicate in the mitochondrial membranes could take advantage of the expanded ER membranes or increased phospholipid synthesis in pah1Δ yeast, first we used Carnation Italian ringspot virus (CIRV) (Table S1), a tombusvirus closely related to TBSV [29]. CIRV replicates on the outside surface of the mitochondrial membranes in vivo and in vitro [42], [63], [64]. Interestingly, the level of CIRV replication was not changed in pah1Δ yeast (Fig. 8A, lanes 6–10). Also, the membrane-fraction obtained from pah1Δ yeast resulted in similar level of RNA replication supported by the CIRV p36 and p95pol replication proteins than the membrane-fraction from wt yeast (Fig. 8B, lane 2 versus 1). The accumulation levels of p36 and p95pol were comparable in pah1Δ yeast, while the accumulation of cellular Sec61p and Ssa1p was increased as expected (Fig. 8B). Based on these data, we suggest that CIRV replication is not affected in yeast lacking the PAH1 gene and the expanded ER membranes do not seem to be utilized by CIRV in yeast.
The second RNA virus tested was Nodamura virus (NoV) (Table S1), an insect RNA virus not related to tombusviruses. NoV RNA replicates on the outer mitochondrial membranes in yeast cells by expressing a single replication protein termed protein A [65]–[67]. We found similar level of NoV RNA accumulation in pah1Δ and wt yeasts (Fig. 8C), suggesting that NoV does not take advantage of the cellular changes caused by the deletion of PAH1.
To obtain evidence if the conserved lipin-like PAP phosphatidate phosphatase also plays a role in tombusvirus replication in plants, we over-expressed the Arabidopsis Pah2p protein in N. benthamiana leaves using an Agrobacterium-based expression system. Plants have two phosphatidate phosphatase-coding genes, PAH1 and PAH2, which are highly homologous with the yeast PAH1 gene and they can complement the phenotypes in pah1Δ yeast [68]–[70]. Also, deletion of PAH1 and PAH2 in Arabidopsis thaliana causes ER expansion and increased phospholipid synthesis, similar to the phenotypes observed in pah1Δ yeast [69].
We found that the over-expression of AtPah2p caused a 6-fold drop in the genomic RNA accumulation of CNV, a TBSV-like tombusvirus, which replicates in the peroxisomal membrane [39], in N. benthamiana leaves (Fig. 9A). The lethal necrotic effect of the CNV tombusvirus was also attenuated in N. benthamiana expressing the AtPah2p (Fig. 9B). The accumulation of TBSV RNA also decreased by ∼3-fold in N. benthamiana leaves over-expressing AtPah2p (Fig. 9C). On the contrary, the replication of another tombusvirus, CIRV (Table S1), which uses the mitochondrial membrane [63], was not changed in N. benthamiana leaves over-expressing AtPah2p (Fig. 9D). Based on these observations, we suggest that the plant phosphatidate phosphatase also plays a role in TBSV and CNV tombusvirus replication that can use peroxisomes and ER membranes, but does not affect the replication of the mitochondrial CIRV in plants.
To test if the inhibitory effect of AtPah2p over-expression is specific to peroxisomal tombusviruses, we studied the accumulation of red clover necrotic mosaic virus (RCNMV) (Table S1), which is a distantly related RNA virus replicating in the ER membranes of the host cells [71]. We observed that the accumulation of RCNMV RNA decreased by ∼3-fold in N. benthamiana leaves over-expressing AtPah2p (Fig. 9E). Thus, another plant RNA virus, RCNMV, is also affected by the plant lipin-like gene, suggesting that the effect of this host gene on ER membranes is critical for replication of RNA viruses taking advantage of the ER membrane to build replication complexes.
Tombusviruses, like many (+)RNA viruses, depend on host membrane biogenesis during replication. Thus, conditions that induce membrane biogenesis in cells might affect replication of some RNA viruses. Accordingly, the key finding in this paper is that tombusviruses (TBSV and CNV) can take advantage of expanded ER membrane surface, which is due to deletion of the cellular Pah1p PAP enzyme leading to massive enlargement of ER membranes and increased phospholipid biosynthesis in yeast [43], [49], [55], [72], to build VRCs efficiently. The evidence supporting this model is extensive and includes: (i) increased TBSV repRNA accumulation in pah1Δ (or in the functionally similar pah1Δnem1Δ lipin deficient) yeasts or Dgk1p diacylglycerol kinase over-producing yeast; (ii) increased accumulation and stability of tombusvirus p33 replication protein in pah1Δ yeast; (iii) the enhanced in vitro assembly of the tombusvirus VRCs in a CFE-based assay derived from yeast lacking Pah1p; (iv) detection of highly abundant p33 replication proteins in isolated ER fraction or using confocal microscopy in yeast lacking Pah1p; (v) the higher in vitro activity of the tombusvirus replicase purified from CFE obtained from yeast lacking Pah1p than from wt yeast; and (vi) the stimulatory role of the membrane fraction of pah1Δ yeast in the CFE-based TBSV replication assay. The efficient utilization of the expanded ER membrane is reflected by the rapid localization of the p33 replication protein to the ER membrane at an early time point in yeast lacking Pah1p. In addition, it appears that the tombusvirus replicase assembles faster in yeast lacking Pah1p than in wt yeast. Thus, our model proposes that TBSV and CNV efficiently subvert the expanded ER membranes and utilizes the abundant phospholipids generated in pah1Δ or pah1Δnem1Δ lipin-deficient yeast, leading to robust viral replication. These tombusviruses replicate in the expanded ER membranes in lipin deficient yeast although the peroxisomal membranes are also present (and some of these membranes are still utilized by TBSV and CNV), suggesting that the expanded ER likely contains favorable microenvironment (membrane microdomains) for these tombusviruses. Therefore, it seems that TBSV and CNV are flexible in utilizing peroxisomal and ER membranes depending on cellular conditions. Accordingly, TBSV is capable to assemble the VRCs on ER and mitochondrial preparations in vitro [42], indicating that TBSV could exploit a range of subcellular membranes in cells. Altogether, we document that deletion of the lipin gene results in strong stimulatory effect on TBSV replication. We also show that TBSV could readily switch to the vastly expanded ER membranes in lipin-deficient cells to build VRCs and support robust viral replication instead of utilizing the peroxisomal membranes as observed in wt yeast and plants [5], [39]. Thus, the increased phospholipid synthesis and the expanded ER membranes in lipin-deficient cells provide highly suitable environment for TBSV and CNV for efficient viral replication.
Similar to tombusviruses, other (+)RNA viruses subvert various intracellular membranes for construction of VRCs, and the ER membrane is often critical for these processes [4], [73]–[75]. Indeed, we find that the ER-based RCNMV replication in plant host cells is also affected by the lipin gene. Therefore, our findings with tombusviruses presented in this paper could be relevant for other RNA viruses of plants and animals. However, not all viruses could benefit from lipin mutations, since we find that CIRV tombusvirus and NoV insect RNA virus, both of which replicate on the mitochondrial membrane surfaces, could not take advantage of the expanded ER membranes generated in pah1Δ yeast. CIRV was also more restrictive in the CFE-based replication assay, utilizing the mitochondrial membranes more efficiently than the ER membranes [42].
The role of PAH1 in tombusvirus replication based on yeast model host was further supported by data obtained in natural plant host by over-expression of AtPah2p. The over-expression of the yeast Pah1p (especially the constitutively active Pah1-7A mutant) in yeast and the AtPah2p in N. benthamiana strongly inhibited tombusvirus replication. This inhibition is likely due to strong competition between cellular phospholipid pathways driven by the over-expressed PAP enzyme and the need for phospholipid-containing cellular membranes in tombusvirus replication [76]. We suggest that tombusviruses could find and/or induce phospholipid-rich microdomains in peroxisomal or ER membranes less efficiently when the over-expressed cellular PAP enzyme decreases the phosholipids level (by producing DAG) in cells.
Our findings might also influence the current view on genetic mutations and diseases versus pathogen infections. Mutations in cellular genes frequently alter cellular pathways and they can cause genetic diseases. The same gene mutations/changes might also influence pathogen-host interactions as well. For example, the cellular lipin, which is involved in a key cellular decision on using lipids for membrane biogenesis or for storage [43]–[47], could be one of those factors. We suggest that mutations in lipin genes, which are known to induce many genetic diseases in humans and animals [44], [46], [47], [77]–[79], not only change the physiology of the given organism, but they might affect pathogens and their interactions with the altered host. Although we used yeast as a model host, since it only has a single lipin gene (PAH1), the role of the PAP enzyme in phospholipid pathways are conserved in yeast, plants and animals. Deletion/mutations of the lipin gene is known to facilitate membrane biogenesis in all these organisms. Therefore, the effects seen with tombusviruses and RCNMV might also manifest with animal viruses possibly leading to exploitation of the expanded ER membranes during infections. Thus, it is possible that one disease state might facilitate the development of robust viral infection at the cellular level.
Saccharomyces cerevisiae strain RS453 (MATa ade2-1 his3, 15 leu2-3, 112 trp1-1 ura3-52), pah1Δ (pah1Δ::TRP1 derivative of RS453) and pah1Δnem1Δ (pah1Δ::TRP1 nem1Δ::HIS3 derivative of RS453) were published previously [53].
We found that the pah1Δnem1Δ yeast was more consistent than pah1Δ yeast in supporting high level of CNV and TBSV repRNA replication for yet unknown reasons. Therefore, most of the experiments were performed with both strains or with only pah1Δnem1Δ yeast strain.
The yeast expression plasmids, pYEplac181, pYEplac181-wt-Pah1 and pYEplac181-Pah1-7A have been obtained from Dr. George M. Carman [53]. The following yeast expression plasmids have been prepared before: LpGAD-CUP1-HisFlag-p92 (LEU2 selection), UpGBK-ADH-His-p33/GAL1-DI-72 (URA3 selection), HpESC-GAL1-His-p33/GAL10-DI-72 (HIS3 selection) [80]; HpGBK-CUP1-HisFlag-p33/GAL1-DI-72 (HIS3 selection) [81]; UpESC-CUP1-His-p92 (URA3 selection) [82]; UpYES-GAL-Hisp33 (URA3 selection) [76], UpESC-YFP-p33 (URA3 selection), LpGAD-Pho86-CFP (LEU2 selection) [39]; UpESC-GAL1-C36/GAL10-DI-72 (URA3 selection), HpYES-GAL1-C95 (HIS3 selection) [42]. The plasmids pMAL-33, pMAL92 and pET-His-MBP-p33 expressing CNV viral proteins in E. coli were described earlier [61], [83]. LpGAD-ADH::Pex13-CFP (LEU2 selection) [39]; UpESC-GAL1::C36/GAL10::DI-72 (URA3 selection), HpYES-GAL1::C95 (HIS3 selection), UpYES-GAL1::T92 (URA3 selection), HpESC-GAL1::T33/ GAL10::DI-72 (HIS3 selection) [42].
Yeast wt RS453 and pah1Δ strain was transformed with pESC-CUP1-His-p92 (URA3 selection), pESC-GAL1-His-p33/GAL10-DI-72 (HIS3 selection) and LEU2 based plasmids: pYEplac181 (vector control), pYEplac181-wt-Pah1p (for expression of wt Pah1p) or pYEplac181-Pah1p-7A (for expression of constitutively active phosphorylation-deficient Pah1p), respectively [53]. Yeast was pre-grown at 23°C overnight in 2 ml SC-ULH− medium containing 2% galactose and then the cultures were harvested after 24 h 50 µM CuSO4 induction.
To study the stability of p33 replication protein in yeast, RS453 and pah1Δnem1Δ strains were transformed with plasmid UpYES-GAL-Hisp33 expressing His6-tagged CNV p33 from the galactose-inducible GAL1 promoter. Yeast transformants were cultured overnight in SC-U− medium containing 2% glucose at 23°C. Yeast cultures were transferred to SC-U− medium supplemented with 2% galactose for 3 h at 23°C. Then, the cultures were shifted back to the SC-U− medium supplemented with 2% glucose and cycloheximide (at a final concentration of 100 µg/ml). The amount of p33 was detected by Western blotting with anti- His6 antibody at given time points after cycloheximide treatment. Each sample loading was adjusted based on total protein levels as determined by SDS-PAGE [39].
Replication assays were performed by measuring the accumulation of DI-72(+) repRNA relative to the accumulation of the cellular 18S rRNA. For the CNV replication proteins-based replication assay, RS453 (wt), pah1Δ or pah1Δnem1Δ yeast cells [53] were transformed with plasmid LpGAD-CUP1-HisFlag-p92 and UpGBK-ADH-His-p33/GAL1-DI-72. Then, yeast was pre-grown at 23°C overnight in 2 ml SC-LU− (synthetic complete dropout medium lacking leucine and Uracil) medium containing 2% galactose. Replication of TBSV repRNA was induced by adding 50 µM CuSO4 into the medium and, then, the samples were harvested at different time points. The TBSV replication proteins-based assay was similar to that described for CNV above, except yeasts were transformed with plasmids UpYES-GAL1::T92 and HpESC-GAL1:: T33/GAL10::DI-72 were directly grown in 2 ml SC-UH− medium containing 2% galactose for 24 h at 23°C. For CIRV replication assay, yeast strains were transformed with plasmid UpESC-GAL1-C36/GAL10-DI-72 and HpYES-GAL1-C95 and then the yeast was grown at 23°C in 2 ml SC-UH- medium containing 2% galactose for 2 days. Standard RNA extraction and Northern blot analysis was performed as described previously [31], [84].
The membrane fractions were prepared as described previously [85]. Briefly, yeasts were transformed and cultured as described in the main text for TBSV or CIRV repRNA replication in yeast. Cultures were collected by centrifugation to obtain membrane fractions containing the in vivo-assembled tombusvirus replicase complexes. Each membrane fraction preparation was adjusted by the relative amounts of His6-tagged p92 and comparable amounts of replicase from each preparation were used in the subsequent replicase assay. The replicase assay was performed as described [85]. The in vitro reaction (50 µl) contained 10 µl of the normalized MEFs preparations, 50 mM Tris–Cl pH 8.0, 10 mM MgCl2, 10 mM DTT, 0.1 U RNase inhibitor, 10 mM ATP, 10 mM CTP, 10 mM GTP and 0.1 µl of 32P-UTP (3000 Ci/mmol). Reaction mixtures were incubated for 3 h at 25°C, followed by phenol/chloroform extraction and isopropanol/ammonium acetate (10∶1) precipitation. The 32P-UTP-labeled RNA products were analyzed in 5% acrylamide/8 M urea gels.
CEFs from RS453 (wt) or pah1Δnem1Δ were prepared as described earlier [37] and adjusted to contain comparable amounts of cellular Pgk1p, a cytosolic protein marker. The in vitro reaction was performed in 20 µl total volume containing 2 µl of adjusted CFE, 0.5 µg DI-72 (+)repRNA transcript, 0.5 µg purified MBP-p33, 0.5 µg purified MBP-p92pol (both recombinant proteins were purified from E. coli), 30 mM HEPES-KOH, pH 7.4, 150 mM potassium acetate, 5 mM magnesium acetate, 0.13 M sorbitol, 0.4 µl actinomycin D (5 mg/ml), 2 µl of 150 mM creatine phosphate, 0.2 µl of 10 mg/ml creatine kinase, 0.2 µl of RNase inhibitor, 0.2 µl of 1 M dithiothreitol (DTT), 2 µl of 10 mM ATP, CTP, and GTP and 0.25 mM UTP and 0.1 µl of 32P-UTP. Reaction mixtures were incubated 3 h at 25°C, followed by phenol/chloroform extraction and isopropanol/ammonium acetate (10∶1) precipitation. 32P-UTP-labeled RNA products were analyzed in 5% acrylamide/8 M urea gels [37]. To detect the double-stranded RNA (dsRNA) in the cell-free replication assay, the 32P-labeled RNA samples were directly loaded onto the gel without heat treatment. Membrane and soluble fractions of these CFEs or in vitro reaction were separated by centrifugation at 35,000 g for 30 min and then mixed them in various combinations (described in the figure legends).
Yeasts were transformed and cultured as described in the main text for TBSV or CIRV repRNA replication in yeast. The ER fractions were prepared as described earlier [42], [86]. Briefly, yeast cells were made into spheroplasts by incubating with 5 mg/g (wet weight) Zymolyase 20T (Seikagaku), and then the spheroplasts were homogenized and lysed with a glass Dounce homogenizer in ice-cold HEPES lysis buffer (20 mM HEPES/KOH [pH 6.8], 50 mM potassium acetate, 100 mM sorbitol, 2 mM EDTA, 1 mM DTT and 1% [V/V] yeast protease inhibitor cocktail [Ypic]). The homogenized spheroplasts were then centrifuged at 1,000 g for 10 min at 4°C, and the supernatant was subjected to additional centrifugation at 27,000 g for 10 min at 4°C to obtain the membrane preparation. To further purify the ER fraction, the membrane preparation was subjected to centrifugation at 100,000 g on a sucrose step gradients (1.5 M and 1.2 M sucrose/HEPES). The purified ER fraction was recovered between the sucrose gradient interfaces and each fraction was adjusted by the amounts of the protein mentioned in figure legend. The in vitro TBSV replication assay was performed as described for in vitro replication assay with MEF except that using the purified ER fraction.
200 µl of the CFE-based replication assay was performed as described above, except that only rATP and rGTP were used. Also, the CFE assay contained the MBP and His6-tagged recombinant p33. At the end of the in vitro replicase assembly assay, the reaction mixture was diluted with 800 µl solubilization buffer, and the replicase complex was purified followed the procedure described previously [61]. In vitro RdRp activity assay was performed using DI-72 region I/III (−)RNA or region IV (+)RNA as template transcribed in vitro by T7 transcription.
To visualize the ER, Pho86-CFP was used as a marker, while peroxisomes were monitored with the help of Pex13-CFP (a peroxisome membrane marker) [39]. The yeast cells were transformed with UpESC-YFP-p33 and LpGAD-Pho86-CFP or LpGAD-ADH::Pex13-CFP. For the 24 h time point, transformed yeast cells were grown in SC-UL− medium containing 2% galactose at 23°C for 24 h and then sample were collected and analyzed by confocal microscopy as described [41]. For short time points, the transformed yeasts were pre-grown overnight at 23°C in SC-UL− medium containing 2% glucose and then transferred to media containing 2% galactose, and then, samples were collected for microscopy analysis at given time points.
Confocal laser scanning micrographs of yeast cells were acquired on an Olympus FV1000 microscope (Olympus America Inc., Melville, New York) as described [41]. ECFP was excited using 440 nm laser light, attenuated to 4.5% of the maximum laser power, while EYFP was excited using 515 nm laser line (3.5% of the maximum laser power). The images were acquired using sequential line-by-line mode in order to reduce excitation and emission cross-talk. The primary objective used was water-immersion PLAPO60XWLSM (Olympus). Image acquisition was conducted at a resolution of 512×512 pixels and a scan-rate of 10 µs/pixel. Image acquisition and exportation of TIFF files were controlled by using Olympus Fluoview software version 1.5.
To prepare the total protein sample for Western blotting, we followed a previous protocol [31], [39]. Briefly, 1 ml of yeast culture was harvested by centrifugation. Then, the samples were re-suspended in 200 µl of 0.1M NaOH and incubated at room temperature with shaking for 20 min. The supernatant was removed after a short centrifugation, and the pellet was re-suspended in 50 µl, 1X SDS-polyacrylamide gel electrophoresis (PAGE) buffer containing 5% β-mercaptoethanol and incubated at 85°C for 15 min. The supernatant was used for SDS/PAGE and Western blot analysis as described [87]. To detect the CNV or TBSV viral proteins, anti-His6 antibody was used as the primary antibody (Invitrogen) and the secondary antibody was alkaline-phosphatase conjugated anti-mouse IgG (Sigma). For cellular protein markers, the following antibodies were used: anti-3-phosphoglycerate kinase (anti-PGK), and anti-heat shock protein 70 (anti-Hsc70) (purchased from Invitrogen, CA). Sec61p antibody was provided by Tom Rapoport, Harvard Medical School. Fox3p antibody was provided by Daniel J. Klionsky, University of Michigan.
Transient expression of Arabidopsis thaliana Pah2p in N. benthamiana leaves was performed by agroinfiltration [88]. A. thaliana PAH2 (At5g42870) was amplified by PCR using primers #4630 (GCCGGATCCATGAATGCCGTCGGTAGGATC) / #4631 (CGGCTCGAGTCACATAAGCGATGGAGGAGGCAG) and genomic DNA as template. The obtained PCR product was digested with BamHI and XhoI, purified and ligated into pGD-L [6] previously digested with BamHI and SalI. The resulting plasmid was transformed into Agrobacterium tumefaciens C58C1 [6]. N. benthamiana plants were infiltrated with A. tumefaciens (OD600 = 0.8) carrying pGD-L-PAH2 or the control empty plasmid pGD. Two days later, the same leaves were infiltrated with A. tumefaciens (OD600 = 0.2) carrying pGD-CNV to launch CNV replication, or pGD-CIRV to launch CIRV replication [6]. Leaf samples were collected 5 days later and total RNA was extracted. CNV and CIRV RNA accumulation was analyzed by agarose gel electrophoresis and Northern blotting using 32P-labeled probes complementary to the 3′ end of the viral RNAs [6].
For studies with TBSV and RCNMV, N. benthamiana plants were infiltrated with A. tumefaciens (OD600 = 0.8) carrying pGD-L-PAH2 or the control empty plasmid pGD. Two days later, the same leaves were inoculated with infectious saps containing TBSV or RCNMV virions. Leaf samples were collected 3 days later and total RNA was extracted. RNA accumulation was analyzed by agarose gel electrophoresis and Northern blotting using 32P-labeled probes complementary to the 3′ end of the TBSV RNA or RCNMV RNA1 [6].
|
10.1371/journal.ppat.1006700 | Salmonella exploits the host endolysosomal tethering factor HOPS complex to promote its intravacuolar replication | Salmonella enterica serovar typhimurium extensively remodels the host late endocytic compartments to establish its vacuolar niche within the host cells conducive for its replication, also known as the Salmonella-containing vacuole (SCV). By maintaining a prolonged interaction with late endosomes and lysosomes of the host cells in the form of interconnected network of tubules (Salmonella-induced filaments or SIFs), Salmonella gains access to both membrane and fluid-phase cargo from these compartments. This is essential for maintaining SCV membrane integrity and for bacterial intravacuolar nutrition. Here, we have identified the multisubunit lysosomal tethering factor—HOPS (HOmotypic fusion and Protein Sorting) complex as a crucial host factor facilitating delivery of late endosomal and lysosomal content to SCVs, providing membrane for SIF formation, and nutrients for intravacuolar bacterial replication. Accordingly, depletion of HOPS subunits significantly reduced the bacterial load in non-phagocytic and phagocytic cells as well as in a mouse model of Salmonella infection. We found that Salmonella effector SifA in complex with its binding partner; SKIP, interacts with HOPS subunit Vps39 and mediates recruitment of this tethering factor to SCV compartments. The lysosomal small GTPase Arl8b that binds to, and promotes membrane localization of Vps41 (and other HOPS subunits) was also required for HOPS recruitment to SCVs and SIFs. Our findings suggest that Salmonella recruits the host late endosomal and lysosomal membrane fusion machinery to its vacuolar niche for access to host membrane and nutrients, ensuring its intracellular survival and replication.
| Intracellular pathogens have devised various strategies to subvert the host membrane trafficking pathways for their growth and survival inside the host cells. Salmonella is one such successful intracellular pathogen that redirects membrane and nutrients from the host endocytic compartments to its replicative niche known as the Salmonella-containing vacuole (SCV) via establishing an interconnected network of tubules (Salmonella-induced filaments or SIFs) that form a continuum with the SCVs. How Salmonella ensures a constant supply of endocytic cargo required for its survival and growth remained unexplored. Our work uncovers a strategy evolved by Salmonella wherein it secretes a bacterial effector into the host cytosol that recruits component of host vesicle fusion machinery-HOPS complex to SCVs and SIFs. HOPS complex promotes docking of the late endocytic compartments at the SCV membrane, prior to their fusion. Thus, depletion of HOPS subunits both in cultured cell lines as well as a mouse model inhibits Salmonella replication, likely due to reduced access to host membranes and nutrients by the vacuolar bacteria. These findings provide mechanistic insights into how this pathogen reroutes the host’s endocytic transport towards its vacuole, ensuring its own intracellular survival and replication.
| Salmonella enterica serovar typhimurium (hereafter Salmonella) is a Gram-negative facultative intracellular pathogen that causes gastroenteritis in a human host and a typhoid-like disease in mice. Salmonella replicates inside the non-phagocytic and phagocytic mammalian host cells in a unique membrane-bound vacuolar compartment known as the Salmonella-containing vacuole or SCV. Modulation of SCV association with the host endocytic machinery and reorganization of the host late endosomes and lysosomes is a major virulence strategy used by this pathogen. Salmonella invasion into the host cell and its replication inside the SCV is facilitated by bacterial effector proteins translocated into the host cytosol by its two type III secretion systems (T3SS)-1 and (T3SS)-2, encoded by the Salmonella pathogenicity island (SPI)-1 and -2 respectively [1,2,3].
During early time points of infection, SCV acquires markers of early endosomes including Rab5, EEA1 (Early endosome antigen 1), SNX1, and PI(3)P [4,5,6]. Within 30–60 min post infection (p.i.), early SCV matures into late SCV by loss of early endosomal proteins and simultaneous acquisition of selective late endosomal and lysosomal proteins including Rab7, lysosomal glycoproteins (lgps) such as, LAMP1 and LAMP2 and v-ATPases [5,7]. Although the SCV acquires characteristics of late endocytic compartments including acidification, it does not become bactericidal, due to reduced presence of lysosomal hydrolyases [8]. Onset of bacterial replication in host cells begins at 3–4 hr p.i. and coincides with the formation of a tubular membrane network that emanate from the SCV, known as Salmonella-induced filaments (SIFs) [9,10,11]. SIFs have been observed in both Salmonella infected epithelial cells and phagocytic cells, and are characterized by presence of lgps such as LAMP1 [12]. Recent studies have shown that early SIFs formed during 6–8 hr p.i. are highly dynamic structures that continuously acquire content from the late endosomes and lysosomes of the host cell [10,13,14]. Detailed ultrastructure analysis of SIFs has revealed that a subset of these are double membrane structures wherein the space (that harbors the bacteria) between the outer and the inner lumen is accessible to content from the host late endosomes and lysosomes [13]. Notably, this crosstalk with the host’s endocytic compartments is essential for supply of nutrients to the SCV. As previously reported, auxotrophic strains of Salmonella acquire external amino acids by inducing SIF formation and redirecting host vesicular transport to the SIFs and SCV membranes [14,15]. Moreover, as SIFs are a large interconnected network of tubules forming a continuum with SCVs, this is proposed to rapidly dilute the antimicrobial activities transferred to the bacterial vacuole upon content mixing with the host late endosomes and lysosomes, preventing degradation of the vacuolar Salmonella [14]. These studies signify a crucial role of Salmonella effectors that mediate SCV interaction with the host endocytic machinery and SIF formation required for the survival and replication of this pathogen within its intravacuolar niche.
Several T3SS-2 effectors including SifA, SseJ, SseG, SseF, SopD2, and PipB2 contribute to SCV maturation, vacuole integrity and SIF formation [16]. The most severe phenotype on the intracellular Salmonella growth is observed with strains lacking SifA that are highly attenuated in systemic infection and replication [17,18]. Functionally, SifA regulates the integrity of SCV and is essential for SIF formation [19]. SifA interacts with the host protein SKIP (SifA and Kinesin interacting protein)/PLEKHM2 (Pleckstrin homology and RUN domain containing protein M2) that in turn bind to the motor protein, kinesin-1 [8,20,21]. SifA-mediated SKIP recruitment on SCVs is thought to relieve the auto-inhibition of kinesin motor, which in turn promotes the microtubule-dependent extension of SIFs [22]. SifA also interacts with the host protein PLEKHM1 (Pleckstrin homology and RUN domain containing protein M1), which has similar domain architecture as SKIP, and regulates membrane biogenesis of the SCV compartment, and intracellular Salmonella proliferation [23]. Although components of the host late endosome-lysosome fusion machinery are known to localize to SCV and SIFs (such as Rab7 [24] and Arf-like (Arl) GTPase8b [25]), but their function in Salmonella replication and whether Salmonella modulates their recruitment for its own survival needs further exploration.
Lysosome fusion with other membrane-bound compartments requires the small GTPases Rab7 and Arl8b and their effectors; PLEKHM1 and tethering/docking factor HOPS complex, respectively as well as the SNARE proteins [26,27,28,29,30,31,32]. HOPS complex is an evolutionarily conserved multisubunit tethering complex (MTC) that mediates lysosome fusion with late endosomes, phagosomes, and autophagosomes [33]. HOPS is a hexameric complex where four of the six subunits namely, Vacuole protein sorting (Vps)11, Vps16, Vps18 and Vps33a form the core complex, and Vps39 and Vps41 are the accessory subunits [34]. The four core subunits of the HOPS complex are shared with CORVET (class C core vacuole/endosome tethering), an early endosomal MTC. Vps3/TGFBRAP1 (Transforming Growth Factor Beta Receptor Associated Protein 1)/TRAP1 and Vps8 are the accessory subunits of the hexameric CORVET complex [35,36]. In yeast and in mammalian cells, CORVET complex is recruited on to the early endosomal membranes by TGFBRAP1 binding to early endosomal Rab protein, Rab5 [37,38]. In yeast, but not in mammalian cells, Rab7 directly binds to the accessory subunits Vps39 and Vps41 to recruit HOPS complex to the vacuolar membranes, promoting their homotypic fusion [27,39]. Interestingly, in metazoans including C. elegans and in mammalian cells, Vps41 subunit of the HOPS complex binds to Arl8b, which then mediates assembly of HOPS complex on lysosomes [27,28,40]. Fewer studies have explored the role of mammalian HOPS subunits in maturation of pathogen-containing vacuole. Previous work has shown that HOPS complex plays an inhibitory function in regulating intracellular survival of the pathogen Coxiella burnetti by mediating fusion of bacterial phagosomes with lysosomes [41]. Consequently, C. burnetti-mediated phosphorylation of Vps41 subunit of the HOPS complex prevents its membrane localization, and thereby, its function in phagolysosome fusion. HOPS complex has also been shown to be essential for Ebola virus replication with loss of HOPS subunit expression preventing viral escape to the cytosol from host’s late endosomes/lysosomes [42,43].
Although SCV compartment is known to acquire content from the late endocytic compartments of the host cell [9,12], little is known if Salmonella employs HOPS complex to mediate fusion with host endolysosomes. Previous studies have shown that HOPS subunit Vps39 interacts with SKIP/PLEKHM2 and PLEKHM1, both of which bind to the Salmonella effector SifA [27,29]. Moreover, HOPS complex is an effector of the small GTPase Arl8b that localizes to SCVs and SIFs in Salmonella-infected HeLa cells [25,27]. A more direct evidence of HOPS function during Salmonella infection was shown where depletion of HOPS subunits (similar to PLEKHM1 depletion) altered SCV morphology with multiple bacteria present within a single enlarged vacuole [23]. However, Salmonella infection in these experiments was visualized after 20 hr p.i. while SCV interaction with host late endosomes/lysosomes and SIF formation is observed as early as 6 hr p.i. [9]. Thus, with regard to the role of HOPS complex in Salmonella infection, several important questions remain unanswered, for instance, whether HOPS complex regulates Salmonella replication, does it regulate SCV maturation and SIF formation and what are the bacterial and host factors required for recruitment of HOPS complex to SCVs and SIFs.
Here, we demonstrate an essential role of HOPS complex in mediating intracellular Salmonella replication in non-phagocytic and phagocytic cells and in a mouse model of Salmonella infection. Live-cell imaging experiments and transmission electron microscopy (TEM) studies revealed that SIF formation and fusion of mature SCVs with late endosomes and lysosomes were severely compromised upon depletion of HOPS subunits. Consequently, nutrient access to SCVs from the host late endocytic compartments was also impaired upon HOPS depletion. Notably, we found that bacterial effector SifA, in complex with the host protein SKIP, interact with HOPS complex and mediate HOPS localization to SCVs, enabling fusion with LEs/Lysosomes. Surprisingly, we did not find a role PLEKHM1 in mediating HOPS recruitment to SCVs although previous studies had shown that it independently binds to both SifA and Vps39 [23,29]. In conclusion, our results demonstrate that Salmonella recruits the host vesicle fusion machinery to gain access to nutrients and membranes from the late endocytic compartments to build its replicative niche inside the host cells.
To investigate the role of HOPS complex in SCV maturation and fusion with late endosomes/lysosomes, we first examined the time-dependent localization of HOPS subunits to SCVs and SIFs in Salmonella-infected human epithelial cell line (HeLa). Vps41 that recruits other subunits of the HOPS complex to lysosomes [27], showed weak association with early SCVs at 10 min p.i. Most SCVs were positive for the early endosomal marker-EEA1 at this time point (S1a Fig). Recruitment of HOPS subunits, Vps41 and Vps18, around the SCVs was observed starting at 1 hr p.i. (53±4% for Vps41) that became more evident by 3 hr p.i. where 74±1% SCVs were positive for Vps41 (Fig 1a, 1b and 1f; S1b and S1c Fig). By these time points, SCVs undergo maturation and as shown in the images acquire late endosomal/lysosomal marker- LAMP1 (Fig 1a and 1b; S1b and S1c Fig). Epitope-tagged Vps41 and Vps33a were similarly recruited to the mature SCVs in Salmonella-infected HeLa cells that were briefly treated with mild-detergent prior to fixation to remove the cytosolic signal of the overexpressed proteins (S1f–S1i Fig; quantification of HA-Vps41 SCVs shown in Fig 1g). Prior treatment with detergent resulted in non-specific nuclear staining, as observed in confocal micrographs of the transfected cells (S1f–S1i Fig). Notably, endogenous and epitope-tagged HOPS subunits also localized to LAMP1-positive vesicles, supporting their reported subcellular distribution to late endosomes/lysosomes (see red arrowheads in insets of Fig 1a and 1b). We were unable to observe the subcellular localization of the HOPS-specific subunit-Vps39 in these experiments, due to lack of an antibody against the endogenous protein and the fact that its overexpression results in striking coalescence of lysosomes into large aggregates [44]. Previous studies have shown that four of the six subunits of the HOPS complex (Vps11, Vps16, Vps18 and Vps33a) are shared with CORVET, which is an early endosomal tethering factor [35]. To determine whether CORVET complex also localizes to SCV membranes, we analyzed distribution of epitope-tagged CORVET specific subunit-TGFBRAP1 in Salmonella-infected cells at 10 min, 30 min, 1 hr, 3 hr and 6 hr p.i. (S2a–S2e Fig). As expected, TGFBRAP1 colocalized with EEA1 (see yellow arrowheads in insets of S2a Fig) but not LAMP1. Further, little or no recruitment of TGFBRAP1 on SCVs and SIFs was observed at different time points of infection (S2a–S2e Fig).
Beginning at 6 hr, but better visualized at 10 hr p.i., HOPS subunit Vps41 localized to >93±4% SCVs and also localized to SIFs, identified by co-immunostaining for LAMP1-a well-characterized marker for these tubular membranes that frequently extend from the surface of mature SCVs (Fig 1c and 1d). We also observed a similar striking localization of Vps18 subunit of the HOPS complex to SCVs and SIFs in infected cells beginning at 6 hr, but primarily at 10 hr p.i. (S1d and S1e Fig). Localization of HOPS subunits to SIFs was also verified by expressing epitope-tagged-Vps41 and -Vps33a in Salmonella-infected cells (S1j–S1m Fig). We observed that localization of Vps41 appeared to be discontinuous and in discrete domains along the length of the SIFs. Previous studies have described similar discrete distribution of LAMP1 on SIFs attributed to poor preservation of the tubular membranes in fixed cells [10,16]. To elucidate whether the punctate localization of HOPS subunits observed on the SIFs was due to fixation, we infected HeLa cells with Salmonella constitutively expressing monomeric DsRed (DsRed-Salmonella) followed by transfection with GFP-tagged Vps41 at 2 hr p.i. Live-cell imaging at 9–10 hr p.i. revealed Vps41 was present around the SCVs and on SIFs in a continuous manner rather than as discrete domains (S2k Fig and S1 Movie). To reduce the cytosolic signal that interfered with visualizing the membrane localization of overexpressed Vps41, we co-expressed small GTPase Arl8b, which recruits Vps41 to lysosomes (Fig 1e and S2 Movie) [27]. Moreover, as shown in a previous study [25] and as can be appreciated in S2f–S2j Fig, Arl8b itself localizes to SCVs starting at 1 hr p.i., and to SIFs at 6 hr and 10 hr p.i., and is an excellent marker to visualize these compartments. Notably, we found that GFP-Vps41 was completely cytosolic and failed to localize to SCVs in CRISPR/Cas9 Arl8b-knockout cells (S2l and S2m Fig and S3 Movie). Quantification of Vps41-positive SCVs at 10 hr p.i. in wild type (WT)- and Arl8b knockout-HeLa cells demonstrated an essential role of Arl8b in recruitment of HOPS complex to SCV membranes. (S2n Fig; mean percentage Vps41-positive SCVs in WT: 91±2% and Arl8b KO: 6±1%). Consistent with this role of Arl8b, we found a striking recruitment of GFP-tagged Vps41 to SCVs and SIFs in Arl8b co-expressing cells where Vps41 was present around the SCVs and SIFs in a continuous manner, with fewer Vps41-positive vesicles remaining in the cell (Fig 1e and S2 Movie). We also observed the dynamic extension and retraction of Vps41-labeled SIFs in these cells (see white arrowheads in Fig 1e and S2 Movie). Further, Vps41-positive vesicles were also observed to fuse with the existing tubules and vesicles that were moving along the length of the tubules (see red arrowheads in Fig 1e).
Next, we examined whether recruitment of HOPS subunits to Salmonella-associated membranes increased as a function of time. To analyze this, we resolved the Salmonella-infected homogenates by two-step density gradient ultracentrifugation and confirmed the presence of Salmonella by immunoblotting with antibodies against the bacterial protein-DnaK (Fig 1h, fractions 8–10 (labeled as SCV fraction)). Comparison of the Salmonella-infected cell homogenates processed at 3 hr and 8 hr p.i. demonstrated that endogenous HOPS subunits along with LAMP1 (a known SCV marker) were enriched in the SCV fraction from 3 hr to 8 hr p.i. (Fig 1h). As expected, the early endosomal marker-EEA1 was associated with the SCVs at 3 hr p.i. but not at 8 hr p.i. Similarly, CORVET-specific subunit- TGFBRAP1 was weakly associated with SCV fractions at 3 hr p.i. but not at 8 hr p.i., supporting our earlier results of little or no recruitment of TGFBRAP1 to early SCVs (Fig 1h). To verify HOPS enrichment on late SCVs and SIFs, we also employed a recently described method of SCV isolation by immunoprecipitation (IP) of SseF-an integral membrane SPI2-T3SS effector protein [45]. As shown in Fig 1i, HOPS subunits were specifically enriched in the SseF-IP eluate but not control IP with levels comparable to the known SCV markers, such as LAMP1 and Rab7. In contrast, little or no co-IP of GAPDH or Catalase with SseF was observed, substantiating the specificity of this approach for SCV isolation (Fig 1i). Our results indicate correlation between recruitment of HOPS complex with time points wherein SCV is known to acquire content from late endosomes and lysosomes [9]. Indeed, in a recent study by Santos et al., where proteomes of early SCV and late SCV were compared, enrichment of HOPS subunits Vps11, Vps16, and Vps18 was observed in the late SCV fractions [46].
To elucidate the significance of HOPS complex during Salmonella infection, we assessed the intracellular replication of Salmonella in cells depleted of various HOPS subunits. Western blotting and qRT-PCR analysis confirmed efficient depletion of HOPS subunits in HeLa cells (S3a–S3e Fig). Control- and HOPS specific-siRNA treated HeLa cells were infected with Salmonella and fixed at 2 hr and 10 hr p.i., and labeled with anti-Salmonella antibodies to enumerate the intracellular bacterial load by immunofluorescence microscopy (Fig 2a). At 2 hr p.i., both control- and HOPS-siRNA treated cells showed a similar bacterial load with ~35% cells containing 6–10 bacteria/cell, ~17–30% of cells containing 11–20 bacteria/cell, and ~2–7% of cells were containing >20 bacteria/cell. These results suggest that HOPS complex is not required for Salmonella invasion into the host cells (Fig 2a). In contrast to this early time point of infection, at 10 hr p.i., while ~73% of control siRNA treated cells had >20 bacteria/cell and <13% had 11–20 bacteria/cell, only 10–30% of HOPS siRNA treated cells showed a similar bacterial load with almost equal distribution of cells containing either 6–10 bacteria/cell (20–30%) or 11–20 bacteria/cell (30–35%) (Fig 2a). These results indicate a severe defect in bacterial replication upon depletion of HOPS subunits.
We corroborated these observations by determining the number of Colony Forming Units (CFUs) present in control- and HOPS siRNA-treated HeLa cell lysates at 2 hr and 10 hr p.i. As shown in Fig 2b (quantification of CFUs/well), we observed a ~3 fold increase in bacterial replication in control cells, while only ~1.09–1.4 fold increase was observed in HOPS-depleted cells. Consistent with our previous data depicting weak or no association of CORVET subunit TGFBRAP1 with SCVs and SIFs, a ~2.8 fold increase in bacterial replication was observed upon TGFBRAP1 depletion (Fig 2b and S3f Fig showing knockdown efficiency >70%), which was not significantly different from control cells. These results suggest that HOPS, but not CORVET complex, regulates intracellular Salmonella replication. Since HOPS complex is one of the components of the late endocytic vesicle fusion machinery, we compared bacterial burden in HOPS-depleted cells with cells depleted of small GTPases and SNARE proteins also required for late endosome-lysosome fusion. To this end, we analyzed bacterial replication in cells treated with siRNA against Rab7, Arl8b, and late endosomal/lysosomal SNAREs proteins-Vti1b, Syntaxin 8, Syntaxin 17, and Vamp7. Western blotting and qRT-PCR analysis were done to confirm efficient depletion of these proteins in HeLa cells (S3g–S3l Fig). Similar to HOPS depletion, only ~1.1–1.3 fold increase in bacterial replication from 2 hr to 10 hr was observed in Rab7 and Arl8b depleted cells (Fig 2c). Amongst SNAREs, Syntaxin8 showed the most significant decrease in bacterial replication (~1.15 fold; Fig 2d) followed by Vti1b and Syntaxin 17 (~1.75 and ~1.8 fold, respectively; Fig 2d) whereas bacterial replication was modestly (but significantly) decreased in Vamp7-depleted cells (~2.3 fold; Fig 2d). In addition to HeLa cells, we also verified that HOPS subunit-Vps41 is required for Salmonella replication in primary mouse embryonic fibroblasts, MEFs (S3p and S3r Fig; knockdown efficiency >70%; control siRNA: ~3.3 fold, Vps41 siRNA: ~2.5 fold increase in bacterial burden from 2 hr to 10 hr p.i.).
It is well understood that macrophages are the major reservoir of Salmonella in host organisms [47]. Accordingly, to determine whether HOPS complex is required for Salmonella replication in macrophages, we performed CFU assays in control-, Vps39- and Vps41-siRNA treated macrophage-like cell line-RAW264.7, a well-established cell line model for in vitro studies of Salmonella infection (S3m and S3n Fig; knockdown efficiency >80%). As compared to control where ~4 fold increase in bacterial burden was observed, we found only a ~2.8 fold and ~1.3 fold increase in Salmonella burden upon Vps39 and Vps41 depletion, respectively, in macrophages reinforcing that HOPS complex is essential host factor for intracellular Salmonella replication in both epithelial and macrophage cells (Fig 2e and 2f). A similar trend (but overall less replication) was observed in the control and Vps41 lentiviral-mediated shRNA transduced cells (S3o and S3q Fig; knockdown efficiency >70%; control shRNA: ~2 fold and Vps41 shRNA: ~ 0.9 fold change in bacterial burden from 2 hr to 10 hr p.i.) To corroborate the bacterial infection experiments performed under in vitro cell culture conditions, we next assessed whether HOPS subunits are required for in vivo replication of Salmonella in a mouse model. To determine this, we used morpholino-based approach to downregulate Vps41 expression in mice that were further infected with Salmonella by intravenous injection. As a control, standard negative control morpholino was injected in age-matched mice. At day 3 p.i., CFU counts were analyzed from the liver and spleen homogenates of control- and Vps41-morpholino treated mice. The efficiency of Vps41 depletion in both liver and spleen was found to be >80% and >70%, respectively, while no change in the levels of Vps18 (that directly binds to Vps41) was observed (Fig 2g). Similar to our previous findings in cultured cells, striking decrease in in vivo replication of Salmonella was observed upon Vps41 depletion (Fig 2h). Consistent with this, DnaK signal was also strikingly reduced in tissue homogenates from Vps41 morpholino-injected mice (Fig 2g, third panel). Overall, our findings reveal HOPS complex as an essential host factor required for Salmonella proliferation in multiple cell types and in a murine infection model.
To establish its intracellular replicative compartment, Salmonella dynamically interacts with, and acquires both membrane and luminal content from host late endosomes/lysosomes [2]. Since HOPS complex is a crucial factor required for tethering and fusion of incoming cargo with lysosomes, we hypothesized that HOPS function is required for SCV fusion with late endosomes and lysosomes. To this end, we first analyzed SCV maturation upon HOPS depletion by quantifying recruitment of early SCV marker (EEA1) and late SCV marker (LAMP1) at different time points in control-, Vps41- and Vps39-siRNA treated cells. At 10 min p.i., no significant differences in the percentage of EEA1-positive SCVs were evident in HOPS-depleted cells as compared to the control cells (S4a–S4c Fig and quantification shown in Fig 3g and 3h; control siRNA: ~78–84%, Vps41 siRNA:~75%, and Vps39 siRNA: ~71%). LAMP1 acquisition was not observed in either control or HOPS depleted cells at this early time point of infection (S4a–S4c Fig; see intensity profile). At 1 hr p.i., ~70% of SCVs in control siRNA treated cells were now EEA1-negative and had acquired LAMP1 (Fig 3a, see intensity profile; quantification shown in Fig 3g and 3h). In contrast, upon HOPS depletion, we found that EEA1 was still retained around ~40% of SCVs while LAMP1 acquisition was observed only around ~14–25% SCVs at 1 hr p.i. (Fig 3b and 3c; quantification shown in Fig 3g and 3h). These findings suggest a delay in SCV maturation upon depletion of HOPS subunits. Interestingly, by 6 hr p.i., ~62–70% of SCVs had acquired LAMP1 staining in HOPS siRNA treated cells and none were found to be positive for the early endosomal marker, EEA1 (Fig 3d–3f; quantification shown in Fig 3g and 3h). As LAMP1 is distributed on both late endosomes and lysosomes, we also analyzed localization of a specific late endosomal markers-Rab7 and -LBPA on SCVs in control and HOPS depleted cells at 1–6 hr p.i. As previously reported [48], we did not observe acquisition of the late endosomal lipid-lysobisphosphotidic acid (LBPA) to SCV membranes either in control or HOPS depleted cells at 1 hr and 6 hr p.i. (S5 Fig). Interestingly, Rab7 acquisition was unchanged upon HOPS depletion wherein >80–90% SCVs were positive for Rab7 at 1 hr, 3 hr, and 6 hr p.i in both control and HOPS depleted cells (Fig 4a–4f; quantification shown in Fig 4g). Notably, as compared to the control siRNA treated cells, we did observe a modest decrease in Rab7 intensity around the SCVs in HOPS depleted cells at 1 hr p.i. that was recovered by 3 hr p.i. (Fig 4h and 4i). Our findings suggest that especially at 1 hr p.i., several SCVs in HOPS-depleted cells retain characteristics of both early endosomes and late endosomes. (see quantification shown in Figs 3g, 3h and 4g). Taken together, these results signify a delay but not a complete block in SCV maturation upon depletion of HOPS subunits.
Our findings indicate that SCV maturation follows a scheme similar to maturation of early endosomes to multi-vesicular bodies/late endosomes upstream of HOPS-mediated fusion of late endosomes and lysosomes [49,50]. In agreement with this, previous studies have shown that endocytic machinery required for early to late endosome maturation such as Vps34 and Rab7 is also required for SCV maturation [24,51,52]. To confirm that LAMP1 acquisition by SCVs is not inhibited upon fusion with lysosomes, we treated cells with Bafilomycin A1 (Baf A1), a routinely used chemical inhibitor of vesicle fusion with lysosomes [53]. Baf A1 inhibits fusion of lysosomes with other compartments by inactivating the ER Ca2+-ATPase (SERCA) whose activity is required to maintain the lysosomal Ca2+ stores [54,55]. As shown in S4d and S4e Fig, LAMP1 acquisition around SCVs was not impaired in cells pretreated with Baf A1 (see intensity profile graphs in S4f and S4g Fig) although SIF formation was abrogated in the presence of this drug. These findings support our conclusion that LAMP1 acquisition by SCV does not require heterotypic fusion with lysosomes, which in turn is mediated by HOPS complex.
Previous studies have shown that Salmonella colonizes and hyper-replicates within the cytosol of epithelial cells [56,57]. To address whether the cytosolic hyper-replicating Salmonella population is increased upon HOPS depletion, we determined bacterial burdens in control and Vps41 depleted cells using the previously described modified gentamicin protection assay where cells are treated with chloroquine (CHQ) before the end of infection time point. CHQ is a lysosomotrophic agent that accumulates within endosomes/lysosomes and has been shown to degrade vacuolar but not cytosolic bacteria [56,57]. We observed a modest but not a statistically significant increase in the number of cytosolic bacteria at 7 hr p.i (peak time point of cytosolic replication [57]) in Vps41 siRNA treated cells (S4h Fig: control siRNA: 28±3% and Vps41 siRNA: ~36±4%), suggesting that majority of bacteria (~70%) continue to harbor their vacuolar niche upon HOPS depletion. In concordance with these studies, immunogold-EM of ultrathin sections of Salmonella-infected Vps41 depleted cells at 10 hr p.i. showed presence of several vacuolar bacteria surrounded by limiting membrane positive for late endosomal and lysosomal markers-Rab7 and LAMP1 (Fig 4j–4m).
Salmonella survival and replication inside its vacuole strictly correlates with its ability to form SIFs, which begins at 5–6 hr p.i,. and is best visualized at 8–10 hr p.i. by immunostaining for lysosomal glycoproteins in Salmonella-infected cells [10]. Notably, as compared to the control cells, we did not observe SIF formation at later time points of infection (6 hr and 10 hr p.i.) in cells depleted of either of the six HOPS subunits (S6a–S6i Fig). In contrast, SIF formation was observed in TGFBRAP1-depleted cells (S6j Fig), however SIFs were “beaded” and thinner in these cells, which might explain the modest defect in Salmonella replication as shown in Fig 2b. To establish whether formation or stability of SIFs was reduced upon HOPS depletion, we performed live-cell imaging to visualize GFP-LAMP1 (marker for SIFs) dynamics in control-, Vps39- and Vps41-siRNA treated cells that were infected with DsRed-expressing Salmonella. At 9 hr p.i., time-lapse imaging revealed extensive SIF formation in control cells that was completely absent in Vps41- and Vps39-depleted cells (S4–S6 Movies). Moreover, as compared to the control cells, significantly fewer LAMP1-positive vesicles were found to interact with SCVs in Vps41- and Vps39-depleted cells (S5 and S6 Movies).
Previous studies have shown that SCV association with the late endocytic compartments is significantly increased by 6–8 hr p.i., time points that correlate with the onset of SIF formation [10]. However, whether SIF formation is dependent upon SCV fusion with late endosomes/lysosomes and the host machinery that regulates this fusion is not known. Our results demonstrating that HOPS complex localizes to SCV and SIFs, suggest that similar to its role in mediating late endosome-lysosome fusion, this tethering factor could facilitate SCV fusion with lysosomes. To test this, prior to infection we pre-loaded control siRNA- and Vps41 siRNA-treated HeLa cells or control and Vps41 shRNA stably transduced RAW264.7 macrophages with Alexa 647-conjugated dextran (dextran-647) that specifically labels lysosomes, as shown schematically in Fig 5a. Live-cell imaging performed at 10 hr p.i. in control HeLa and RAW264.7 cells showed several dextran-positive endosomes undergoing fusion with the SCVs, resulting in acquisition of dextran by the SCVs (Fig 5b and 5d; S7 and S9 Movies). SIF formation was also observed in both control siRNA/shRNA-treated cells (S7 and S9 Movies). In contrast, little or no interaction of SCVs with the dextran compartment was observed in Vps41 depleted HeLa and RAW264.7 cells (Fig 5c and 5e; S8 and S10 Movies). Quantification of SCVs positive for dextran-647 and its signal intensity, revealed significantly lower dextran acquisition in Vps41 depleted cells compared to control (Fig 5f–5i; percentage of dextran-positive SCVs in HeLa and RAW264.7 cells-control: ~65–70%, Vps41 depletion: 10–15%). These results suggest that HOPS complex is required for acquisition of fluid-phase content by the SCVs from late endosomes and lysosomes. In agreement with these findings, imaging of ultrathin sections of Salmonella-infected control cells by TEM demonstrated several late endosomes (containing numerous MVBs) and lysosomes (containing lamellar membrane sheets) docked at or in close apposition to the SCVs (Fig 6a and 6b; S7b and S7c Fig; see magnified insets). In contrast, late endosomes/lysosomes docking at the SCVs were highly reduced in Vps41 depleted cells (Fig 6c and 6d; S7a, S7d and S7e Fig; see magnified insets). Further, as previously noted in another study [23], we also observed several abnormal “bag-like” SCVs upon Vps41 depletion (Fig 6d; see magnified inset). Additionally, in few TEM sections, SIF formation was also observed in control but not Vps41 depleted cells (S7b Fig, middle panel). Analysis of several TEM images in control cells revealed that of the ~100 SCVs imaged, ~40 SCVs had closely apposed late endosomes, whereas only ~10 of the 100 SCVs in Vps41 siRNA treated cells and none of the ~60 SCVs imaged in Vps41 shRNA transduced cells showed docked late endosomes. As previously reported [58], we also noted that lysosomes (containing lamellar membrane sheets) were reduced in Vps41 siRNA treated cells while several large MVB-containing compartments were observed (Fig 6c and 6d; indicated by white arrowheads). Although docking of late endocytic compartments at the SCVs was reduced upon Vps41 depletion, this did not indicate a general defect in the formation of late endocytic compartments. This was confirmed by LysoTracker Red uptake in control and Vps41 depleted cells, which is a selective probe that labels acidic organelles and routinely used as a specific marker to label late endosomes and endolysosomes. Immunofluorescence analysis and quantification of LysoTracker Red signal intensity by flow cytometry revealed no significant difference in control and Vps41 depleted cells (S8a–S8e Fig). The specificity of this probe was confirmed by treating cells with Baf A1 that neutralizes the pH of late endocytic compartments, and hence the signal intensity was reduced to background fluorescence levels (S8e Fig). We also confirmed that functional endo-lysosomes are formed upon Vps41 depletion by comparing levels of mature cathepsin B and D in control and Vps41 siRNA treated cells (S8f Fig). Taken together, our findings suggest that HOPS complex is a crucial host factor required for SCV fusion with the late endocytic compartments that provide membranes for formation of a replicative vacuolar niche for this pathogen.
Recent studies have shown that content mixing of SCV with the late endocytic compartments and SIF formation not only provides membrane for vacuolar integrity for the growing bacterial population but also provides nutrient access to the vacuolar bacteria for replication [14,15,17]. This was in part established by use of auxotrophic strains of Salmonella that were deficient in biosynthesis of particular amino acids. The mutant strains were able to replicate by obtaining nutrients from the growth medium of the host cells, only if they were proficient in SIF formation [15]. Based on our findings that HOPS complex mediates SIF formation by promoting SCV interaction with the host late endocytic compartments, we investigated role of HOPS subunits in mediating nutrient access from host cell to SCVs. To this end, we infected control- and Vps41-siRNA treated cells with proline auxotrophic strain of Salmonella (proC). This strain lacks the last enzyme required for proline biosynthesis, and is defective in intracellular replication unless proline is provided in the mammalian cell growth media [15]. As previously noted [15], we also found that proC strain was replication-defective as compared to the wild-type (WT) Salmonella strain. This growth defect was completely augmented by addition of proline in the culture media of control siRNA-treated HeLa cells (Fig 6e). In contrast, upon depletion of HOPS subunit Vps41, only a modest increase in the replication of proC strain in presence of extracellular proline was observed, which was significantly less than the control cells under the same experimental conditions (Fig 6e). These results suggest that HOPS complex provides nutrient access from the host late endosomes and lysosomes to the bacteria within the confinements of the vacuole, enabling intravacuolar replication of Salmonella.
Previous studies have revealed that Salmonella mutant strains deficient in SPI2-T3SS effectors sifA, pipB2, sseF and sseG show the most striking changes in SIF formation [12,16]. The most severe phenotype was observed in Salmonella strain lacking sifA where SIF formation was completely abrogated and vacuolar integrity was disrupted, leading to bacterial release in the host cytosol [18]. Our findings thus far indicate that HOPS complex is a crucial host factor required for SCV and SIF fusion with the late endocytic compartments, providing a continuous supply of membranes for SIF formation. To determine whether Salmonella effectors involved in SIF formation promote HOPS recruitment to SCV membranes, we visualized and quantified the recruitment of HOPS subunits Vps41 (both epitope-tagged and endogenous) and Vps18 (endogenous) to LAMP1-positive SCVs in mutant strains deficient in either sifA ssej, pipB2, sseF or sseG effectors (Fig 7 and S9 Fig). We used the sifA sseJ double-mutant strain in these experiments instead of the sifA single mutant strain as the latter loses its vacuolar integrity over time and becomes cytosolic [17]. Surprisingly, as compared to the WT strain of Salmonella, we did not observe recruitment of HOPS subunits-Vps41 and -Vps18 to sifA ssej SCVs, although association of these SCVs with the vacuolar membrane marker-LAMP1 was observed (Fig 7a and 7b; quantification shown in Fig 7g; S9a, S9b, S9g and S9h Fig). Notably, Vps41 and Vps18 continued to localize at the SCVs in cells infected with Salmonella mutant strains pipB2, sseF and sseG (Fig 7c–7f; quantification shown in Fig 7h; S9c–S9f and S9i–S9l Fig). These results suggest that SifA, but not other Salmonella effectors, involved in SIF formation are crucial for recruitment of HOPS subunits.
Intriguingly, we had previously found that SifA interaction partner-SKIP colocalizes and interacts with Vps39 subunit of the HOPS complex [27]. Based on these observations, we hypothesized that SifA in complex with SKIP targets HOPS complex to SCV membranes. Indeed, while little or no colocalization of Vps39 with SifA was observed, Vps39 colocalized with SKIP on peripheral structures shown to be lysosomes, which are transported in an anterograde manner by direct binding between Arl8b-SKIP complex to the plus-end microtubule binding motor-kinesin-1 (Fig 8a and 8b) [20,59,60]. Notably, colocalization between SifA and Vps39 was strikingly enhanced upon co-expression with SKIP and the three proteins were localized on the peripheral pool of lysosomes (compare Fig 8a and 8d). The other subunits-Vps18 and Vps41, also showed a significantly higher colocalization with SifA in presence of SKIP (compare S10a and S10c Fig; compare S10b and S10d Fig). Quantification of Pearson’s Correlation Coefficient (PCC) from 25–30 transfected cells over three independent experiments demonstrated a significant increase in colocalization of HOPS subunits with SifA in presence of SKIP (Fig 8g and 8h).
Recently PLEKHM1, a protein with domain architecture similar to SKIP, was reported to interact with both SifA and HOPS subunits Vps39 and Vps41 [23,29]. While it was speculated that PLEKHM1 acts as a linker between SifA and HOPS complex, no experimental evidence was shown to prove the same. Indeed, we found a strong colocalization of Vps39 with PLEKHM1, which was significantly higher than its colocalization with SKIP (Fig 8b and 8c; quantification of PCC shown in Fig 8f). To determine whether PLEKHM1, similar to SKIP, promotes colocalization of HOPS subunits with SifA, we co-expressed Vps39 and SifA with PLEKHM1. Surprisingly, while Vps39 and PLEKHM1 continued to colocalize on punctate structures, SifA was not recruited to these punctae (Fig 8e; quantification of PCC shown in Fig 8g). These results indicate that PLEKHM1 does not promote HOPS subunit association with SifA. We also noted that colocalization and interaction of SifA with PLEKHM1 was significantly weaker than with SKIP, as revealed by colocalization coefficient quantification and growth curve analysis of yeast two-hybrid assay using SifA as a bait, and SKIP and PLEKHM1 as prey proteins (S10e–S10h Fig). These findings were corroborated by GST pulldown assay where pull down of PLEKHM1 with GST tagged-SifA was found to be much lower as compared to SKIP from transfected cell lysates (S10i and S10j Fig). Additionally, qRT-PCR analysis revealed that SKIP mRNA levels in HeLa cells were ~2.5 fold higher than PLEKHM1 levels (S10k Fig). Taken together, these results imply that at least in this cell line, more amount of the secreted bacterial effector SifA must be bound to SKIP as compared to PLEKHM1.
To conclusively determine whether SKIP is a linker between SifA and HOPS subunit-Vps39, we employed yeast three-hybrid assay to test interaction of SifA and Vps39 in the presence of either SKIP or PLEKHM1 as well as a SifA binding-defective mutant of SKIP (SKIP G828D). In this assay, linker protein is under the control of the Met25 promoter that remains repressed in the presence of methionine in the growth media. As depicted in Fig 9a, under methionine-deficient conditions, SifA showed interaction with Vps39 only in the presence of SKIP, but not SKIP G828D mutant or PLEKHM1. To corroborate these results, we also performed GST pulldown using GST tagged-SifA as bait to pulldown Vps39 in cells with endogenous or overexpressed levels of SKIP. We observed a dramatic increase in the levels of Vps39 pulldown with SifA upon SKIP overexpression (Fig 9b and 9c). This striking increase in pulldown of HOPS subunits was also reflected upon probing for endogenous Vps11, which directly binds to Vps39 during assembly of the HOPS complex (Fig 9b). Vps41 pulldown with SifA was also increased upon SKIP overexpression, although this was less striking as compared to Vps39 and Vps11 (Fig 9b). To establish that endogenous levels of SKIP are sufficient to drive this interaction, we performed co-IP of SifA and Vps39 in control and SKIP depleted cells (Fig 9d; >90% gene silencing efficiency observed). As shown in Fig 9e and 9f, co-IP of Vps39 with SifA was significantly reduced upon SKIP depletion, and was restored upon expression of the siRNA-resistant SKIP construct, suggesting that SKIP acts as a linker to facilitate interaction between SifA and HOPS complex. In line with these observations and in accordance with previous studies [20,61], we found a significant defect in bacterial replication in SKIP-depleted cells as compared to control (S10l Fig; control siRNA: ~3 fold and SKIP siRNA: ~1.3 fold increase in bacterial burden from 2 hr to 10 hr p.i.). Notably, we did not observe any increase in pulldown of HOPS subunit Vps39 with GST tagged-SifA upon PLEKHM1 overexpression (Fig 9b and 9c). Similarly, no effect on the levels of co-IP Myc-tagged SifA with HA-Vps39 was observed upon PLEKHM1 depletion (Fig 9g; >90% silencing efficiency observed; Fig 9h), suggesting that PLEKHM1 does not facilitate interaction between SifA and HOPS complex.
To then determine whether SKIP is required for recruitment of HOPS subunits to SCV membranes, we visualized Vps41 localization to SCVs in control and SKIP siRNA treated cells. As shown in Fig 10a and 10b, while Vps41 was present around the SCVs in control siRNA treated cells, little or no association was observed in SKIP depleted cells at 10 hr p.i. Quantification of Vps41-positive SCVs in control and SKIP siRNA treated cells demonstrated that Vps41 recruitment to SCV membranes was abrogated upon SKIP depletion (Fig 10g). These findings were corroborated by live-cell imaging experiments of GFP-tagged Vps41 either expressed alone or co-expressed with Arl8b in control and SKIP siRNA treated cells (S11–S14 Movies; S11a and S11b Fig). Recruitment of Vps41 to SCVs was rescued by expression of siRNA-resistant SKIP, confirming specificity of SKIP siRNA treatment (Fig 10e–10g). In contrast, Vps41 continued to associate with SCV membranes in PLEKHM1 depleted cells at 10 hr p.i. (S11c and S11d Fig; Fig 10g), which supports our previous results that PLEKHM1 does not regulate SifA interaction with HOPS complex. To corroborate these findings, we used an independent method to disrupt interaction of SifA and SKIP i.e. infection with Salmonella strain expressing a point mutant of SifA (L130D), which is defective in binding to SKIP and formation of SIFs [21,61]. Using co-IP approaches, we confirmed that SKIP does not interact with the previously reported SKIP-binding interface mutants of SifA (S11e Fig) [21]. Also, unlike SifA deletion (sifA), bacteria expressing SifA (L130D) do not escape to the cytosol and continue to be surrounded by LAMP1-positive vacuolar membrane [61], which allowed us to analyze whether HOPS complex was recruited to the SCVs surrounded by an intact vacuolar membrane. Notably, as compared to the cells infected with the sifA strain expressing SifA (WT)-2xHA plasmid, in cells infected with sifA strain expressing point mutant SifA (L130D)-2xHA, little or no association of HOPS subunit-Vps41 with SCVs was observed at 10 hr p.i. (Fig 10c and 10d). Quantification of Vps41-positive SCVs infected with either strain demonstrated that recruitment of HOPS subunit Vps41 to SCV membranes was abrogated in the presence of SKIP-binding defective mutant of SifA (Fig 10h). These findings were corroborated by live-cell imaging experiments of GFP-tagged Vps41 either expressed alone or co-expressed with Arl8b in cells infected with either Salmonella strain (S15–S18 Movies; S11f and S11g Fig).
A previous study has shown that SifA protein expression in host cells results in the extensive clustering/aggregation of specifically the late endocytic compartments marked by LAMP1 and V-ATPase immunostaining [62]. Taking our results presented here into consideration, SifA could promote SCV fusion with late endosomes/lysosomes by virtue of its interaction with the host factors, SKIP and HOPS complex. Indeed, endogenous HOPS subunits-Vps18 and -Vps41, were enriched on the vertices of these clustered LAMP1-positive compartments induced by ectopic expression of SifA (Fig 11a and 11b). To test whether SifA-mediate clustering and aggregation of late endosomes and lysosomes requires presence of SKIP and HOPS subunits, we transfected SifA in control-, Vps39- and SKIP-siRNA treated cells and analyzed particle size of LAMP1-positive compartment. Our results show that SifA-mediated increase in lysosomal particle size depends upon the expression of SKIP and HOPS subunit-Vps39 (Fig 11c–11f). Taken together, our findings indicate that Salmonella virulence factor SifA in complex with the host protein, SKIP, recruits the vesicle fusion machinery of the host including the tethering factor HOPS complex to SCV membranes, thereby, enabling SCV fusion with late endosomes and lysosomes.
Salmonella typhimurium is a successful intracellular pathogen that has developed an array of sophisticated strategies to massively remodel the host endosomal system for its own survival and propagation. Previous studies have shown that SCV biogenesis involves extensive interactions with the host endocytic pathway including late endosomes/lysosomes [2]. However, little is known about how Salmonella mediates these interactions and whether it co-opts the late endosomal-lysosomal vesicle fusion machinery of the host cell for building its replicative niche. Conflicting reports have shown that while Salmonella inhibits activation of the small GTPase Rab7 [24,63], it actively recruits Arl8b on SCV and SIFs [25] wherein both Rab7 and Arl8b are components of protein machinery required for late endosome-lysosome fusion [26]. Intriguingly, Arl8b-positive lysosomes are less acidic and have reduced proteolytic activity than Rab7-positive endosomes [64]. It is interesting to speculate that Arl8b- but not Rab7-positive lysosomes act as source of membrane for SCV biogenesis and SIF formation during later time points of infection. This would ensure membrane and cargo delivery to SCVs without increasing the proteolytic activity within Salmonella’s replicative niche.
In this study, we have investigated the role of HOPS complex, a multisubunit tethering factor required for vesicle fusion with lysosomes, in regulating Salmonella survival and replication inside its vacuole. Our results reveal that HOPS complex is a target for Salmonella effector SifA, which in collaboration with its known binding partner SKIP and the host GTPase, Arl8b, recruits HOPS complex to SCV membranes, thereby enabling SCV fusion with lysosomes (Fig 11g). As late endocytic compartments are a source for both membrane and fluid-phase cargo, including nutrients for Salmonella residing in the vacuole [2,14], silencing of HOPS subunits inhibited Salmonella replication under both in vitro and in vivo conditions.
Unlike the defense strategies used by intracellular pathogens such as M. tuberculosis and C. burnetti [65,66], Salmonella does not block the maturation of its phagosome, which rapidly (~30–60 min p.i.) acquires several (but not all) characteristics of the late endocytic compartments but does not become bactericidal [2]. The acidic pH of the SCV (~<5) is required for the induction of the SPI-2 effectors, which in turn facilitate Salmonella replication inside the host cell [67]. At 1–2 hr p.i., we found weak but consistent localization of HOPS subunits on mature SCVs, which correlated with the recruitment of the lysosomal marker, LAMP1. While HOPS complex localized to the mature SCVs, we did not find an essential role of HOPS subunits, Vps41 and Vps39, in SCV maturation as indicated by a modest delay but not a block in LAMP1 acquisition in HOPS depleted cells. Our results support previous studies suggesting that SCV maturation is akin to an early to late endosome maturation event, regulated by proteins including PI(3) kinase and Rab7 (acquired upon HOPS depletion, Fig 4) that act upstream of the HOPS complex in endo-lysosome fusion [24,51].
Previous live-cell imaging studies of Salmonella-infected HeLa cells and RAW264.7 macrophages have shown that at 6–8 hr p.i., 90% of SCVs interact with dextran-loaded terminal lysosomes, and acquire not only membrane but also fluid-phase cargo from these compartments [9,12]. Besides delivering membranes for SCV biogenesis, fusion with late endosomes/lysosomes provides access to nutrients for bacterial replication [14,15]. Intravacuolar Salmonella can access nutrients from the host endolysosomal compartments by direct fusion of SCV with these membranes or from cytosol by recruitment of nutrient transporters on SCV and SIF membranes. In both of these scenarios, extensive membrane network will be required, which is delivered by host vesicle fusion machinery including HOPS complex. Accordingly, the ability of proline auxotrophic Salmonella strain to acquire proline from the extracellular media was also abrogated in HOPS-depleted cells. Besides their role as a tethering factor, HOPS subunits bind to SNARE proteins, which mediate membrane fusion [32,35]. We found a comparable defect in Salmonella intracellular replication when we depleted other components of the vesicle fusion machinery including small GTPases-Rab7 and -Arl8b, as well as SNARE proteins: Vti1b, Syntaxin 8, Syntaxin 17 and VAMP7 that are known to regulate late endosome-lysosome fusion [68]. These results indicate that Salmonella co-opts the host vesicle fusion machinery for survival and replication within its intravacuolar niche.
One of the hallmarks of Salmonella intracellular lifestyle is presence of striking tubular membranes or SIFs that emanate from the juxtanuclear SCVs [11]. The ability to form SIFs was found to directly correlate with Salmonella’s ability to replicate both under in vitro and in vivo conditions, as supported by the replication defect observed in Salmonella strains defective in SIF formation [19]. Recent studies have now shown that SIF formation allows Salmonella to convert the host cell endosomal system into a continuum with the SCV, not only providing SCVs access to the endocytosed material but the extensive SIF network is proposed to rapidly dilute the antimicrobial activities transferred to the vacuole upon its fusion with the host late endosomes and lysosomes. As a result, SCVs competent to form SIFs have bacteria with significantly higher metabolic activity than one that cannot form SIFs [14]. Using live-cell imaging we found that depletion of HOPS subunits completely inhibited SIF formation by Salmonella, supporting the strong replication defect observed in these cells.
SifA is the most well characterized Salmonella effector named for its essential role in mediating SIF formation [11]. Accordingly, Salmonella strains lacking SifA show a strong replication defect, as they fail to induce SIF formation and escape into the cytosol [19]. SifA has been shown to interact with two host proteins namely SKIP/PLEKHM2 and PLEKHM1 via pleckstrin homology (PH) domains of these proteins [20,23]. We found that SKIP, but not PLEKHM1, acts as a linker to mediate interaction of HOPS complex with SifA by simultaneously binding to HOPS subunit-Vps39. These results were surprising given the fact that previously PLEKHM1 was implicated in recruitment of HOPS complex to mediate SCV fusion with detoxified lysosomes [23]. However, the role of PLEKHM1 as a linker was never directly tested in this study and it was speculated based on the fact that PLEKHM1 binds to both HOPS complex and SifA [23,29]. A direct comparison of PLEKHM1 and SKIP’s linker role and their relative binding affinities for SifA as well as comparison of expression levels of both proteins in HeLa cells led us to conclude that SifA-SKIP promotes recruitment of HOPS subunits to SCV compartment. It will be interesting to determine whether SifA and Vps39 have overlapping binding sites on PLEKHM1, preventing SifA recruitment to PLEKHM1 and Vps39-positive compartment. Our study also suggests a novel role for SKIP in promoting Salmonella intracellular replication, besides its known function in preventing kinesin-1 accumulation on SCVs and regulating vacuolar integrity [20,22,59].
HOPS complex localization to SCVs and SIFs also required small GTPase Arl8b, which is highly enriched on these compartments and regulates lysosomal localization of both of its effectors-SKIP and Vps41 subunit of the HOPS complex [25,27,60]. Recently, we have uncovered that PLEKHM1, like SKIP, binds to Arl8b via its RUN domain and is a shared effector of Rab7 and Arl8b, which simultaneously binds to both GTPases to promote cargo trafficking to lysosomes [26]. Since Salmonella has devised a strategy to inhibit Rab7 activation, on the other hand Arl8b is enriched on SCVs and SIFs, it will be relevant to determine whether PLEKHM1 role in SCV fusion with lysosomes is dependent upon its interaction with Arl8b.
Unlike Salmonella typhimurium, much less is known about the intracellular lifestyle of the human-restricted pathogen-Salmonella typhi, the typhoid-causing strain of the same serovar. Intracellular S. typhi secretes the typhoid toxin inside its SCV, which is then packaged into vesicular carriers that are then transported into the extracellular space to mediate its effect in an autocrine and paracrine manner on the host cells [69,70]. Interaction of S. typhi vacuole with the host endocytic machinery and mechanisms regulating formation and transport of the typhoid toxin-containing vesicular carriers are only beginning to be understood [71,72]. Indeed, like S. typhimurium, intracellular replication of S. typhi was impaired in Rab7-depleted cells, suggesting that S. typhi might also manipulate host late endosomes and lysosomes to regulate biogenesis of its SCV and growth inside the host cells [73]. Future studies are required to address whether the host endocytic machinery regulates S. typhi replication and biogenesis of the typhoid toxin vesicular carriers that will reveal novel targets for development of antimicrobial molecules.
HeLa, HEK293T, and RAW264.7 cells were obtained from the American Type Culture Collection and maintained in DMEM (Lonza) supplemented with 10% heat-inactivated Fetal Bovine Serum (FBS; Life technologies) at 37°C in 5% CO2 humidified incubator. All the cultures were used between passage numbers 5–15.
An Arl8b-KO HeLa cell line was previously described [26]. Arl8b-knockout (KO) HeLa cells were generated using the Arl8b sg/RNA (Target sequence: 5′-GATGGAGCTGACGCTCG-3′) CRISPR/Cas9 All-in-One Lentivector Set (Applied Biological Materials). For stably silencing the expression of Vps41 in RAW264.7 cells, lentivirus mediated shRNA gene silencing approach was used. Briefly, for lentiviral transduction, RAW264.7 cells were seeded in a 35-mm tissue culture dish (Corning) in Polybrene (8 μg/ml; Sigma-Aldrich) and mixed with 500 μl of viral supernatant (day 0). Puromycin (Sigma-Aldrich) was added after 24–48 hr at 5 μg/ml for a minimum of 3 days to select transductants, and experiments were performed on days 5–15 after transduction. shRNA target sequences were as follows: Mission (negative control sequence), CAACAAGATGAAGAGCACCAA and mouse Vps41, GAGTGGCCTGGAGATCTATAT. Development of HeLa-Vps41 shRNA cell line was previously described using Vps41 shRNA, 5′-CCATTGACAAACCACCATTTA-3′ [27].
Primary mouse embryonic fibroblast (MEF) cells were isolated from the embryos of BALB/c mouse. Briefly, embryos were harvested from female mice 15 days after the appearance of the copulation plug. Embryos were placed in 1 ml of 0.05% trypsin/EDTA solution (Life technologies) and finely minced using a sterile razor blade and repeated pipetting was performed to dissociate cells. The trypsin was inactivated by adding DMEM supplemented with 10% FBS and the culture was centrifuged to pellet MEF cells. The pelleted MEF cells were resuspended in culture media, and plated at optimal density in tissue culture dishes at 37°C in 5% CO2 humidified incubator.
The following antibodies were used in this study: mouse anti-FLAG M2 clone (F1804; Sigma-Aldrich), mouse anti-HA (MMS-101P; Covance), rabbit anti-HA (sc-805; Santa Cruz Biotechnology), rat anti-HA clone 3F10 (11867423001; Roche), mouse anti-Myc 9E10 clone (sc-40; Santa Cruz Biotechnology), mouse anti-α-tubulin (T9026; Sigma-Aldrich), mouse anti-GAPDH (sc-166574; Santa Cruz Biotechnology), mouse anti-EEA1 (610457; BD Biosciences), rabbit anti-EEA1 (ab2900; Abcam), mouse anti-LAMP1 (555798; BD Biosciences), rabbit anti-LAMP1 (ab24170; Abcam), rabbit anti-PLEKHM1 (ab171383; Abcam), rabbit anti-SKIP/PLEKHM2 (HPA032304; Sigma-Aldrich), mouse anti-TGFBRAP1 (sc-13134; Santa Cruz Biotechnology), mouse anti-LBPA (Z-PLBPA; Echelon Biosciences), rabbit anti-Catalase (12980; Cell Signaling Technology), rabbit anti-Rab5 (3547; Cell Signaling Technology), rabbit anti-Rab7 (9367; Cell Signaling Technology), rabbit anti-Cathepsin D (K50161R; Meridian Life Sciences), mouse anti-Cathepsin B clone 4B11 (414800; Thermo Fisher Scientific), rabbit anti-Salmonella O-antigen (225341; BD Biosciences), and mouse anti-DnaK (ADI-SPA-880-F; Enzo Life Sciences). Rabbit anti-PLEKHM1 antibody generated against the N-terminal 497 amino acids of human PLEKHM1 protein was a gift from Prof. Paul Odgren (University of Massachusetts Medical School, Worcester, MA) and has been previously used to detect PLEKHM1 by immunofluorescence and Western blotting [26,74]. Rabbit anti-Arl8 antibody used in this study has been described previously [28].
For detection of HOPS subunits, the following antibodies were used: rabbit anti-Vps11 (ab125083; Abcam), rabbit anti-Vps18 (ab178416; Abcam), rabbit anti-Vps33a (16896-1-AP; ProteinTech), rabbit anti-Vps41 (ab181078; Abcam), and mouse anti-Vps41 (sc-377271; Santa Cruz Biotechnology).
All the Alexa fluorophore-conjugated secondary antibodies were purchased from Molecular Probes (Thermo Fisher Scientific). HRP-conjugated goat anti-mouse and goat anti-rabbit were purchased from Jackson ImmunoResearch Laboratories. Alexa Fluor 647-conjugated Dextran, LysoTracker Red DND-99 and DAPI were purchased from Molecular Probes (Thermo Fisher Scientific). L-Proline, Cytochalasin D, Bafilomycin A1, Polybrene, Streptomycin, Gentamicin and Puromycin were purchased from Sigma-Aldrich. Yeast drop-out media were purchased from Clontech.
All the Salmonella typhimurium strains and plasmids used in this study are described in Table 1. For infection of HeLa and MEF cells, late-log S. typhimurium cultures were used and prepared using a method optimized for bacterial invasion [5]. Briefly, wild-type and mutant bacteria were grown for 16 hr at 37°C with shaking and then subcultured (1:33) in LB (Difco) without antibiotics and grown until late exponential phase (O.D. = 3.0). Bacterial inocula were prepared by pelleting at 10,000 x g for 2 min, diluted 1:100 in Phosphate buffer saline (PBS) (pH 7.2), and added to cells (at the specified MOI) for 10 min at 37°C to allow invasion and synchronized infection. After infection, extracellular bacteria were removed by extensive washing using warm PBS and 50 μg/ml gentamicin was added to the medium at 30 min p.i. for incubation at 37°C. After 2 hr p.i., the concentration of gentamicin in the medium was decreased to 5 μg/ml. Following this infection protocol, cells were processed for microscopy and biochemical experiments as described in the individual figure legends.
For infections of RAW264.7 cells, stationary-phase bacterial cultures incubated at 37°C with shaking were diluted (O.D. = 1) and opsonized in PBS supplemented with 20% FBS for 20 min at 37°C. After three washes in PBS, bacteria were resuspended in growth medium without antibiotics, and added to the cells (MOI of 50:1) for 20 min to facilitate phagocytosis. The remaining protocol was similar as in case of infection of HeLa cells.
Cells grown on glass coverslips (VWR) were transfected with desired constructs using X-tremeGENE-HP DNA transfection reagent (Roche) for 16–18 hr. For gene silencing, siRNA duplexes for non-targeting siRNA pool, control siRNA (5′-TGGTTTACATGTCGACTAA-3′), human Arl8b (5′-AGGTAACGTCACAATAAAGAT-3′), human Rab7a (5′-CTAGATAGCTGGAGAGATG-3′) human Vps11 (5′-GAGGCTGAGCTGAGCCTCGTATT-3′), human Vps18 (5′-CTAGATAGCTGGAGAGATG-3′), human Vps33a (5′-CATTGCAGTGTTGCCTCGATATG-3′), human Vps39 (ON-TARGET plus SMART pool), mouse Vps39 (ON-TARGET plus SMART pool), human Vps41 (5′-CCATTGACAAACCACCATTTA-3′), mouse Vps41 (ON-TARGET plus SMART pool), human PLEKHM1 (5′-CCGGTCTCTGCAAGAGGTATTGT-3′), human SKIP (5′-CTTCTGAACTGGACCGATT-3′), human Vti1b (ON-TARGET plus SMART pool), human Stx8 (ON-TARGET plus SMART pool), human Stx17 (ON-TARGET plus SMART pool), human Vamp7 (ON-TARGET plus SMART pool) and human TGFBRAP1 (ON-TARGET plus SMART pool) were purchased from GE Healthcare (Dharmacon), and transfection was performed using Dharmafect 1 as per the manufacturer's instructions.
Cells were fixed in 4% p-formaldehyde (PFA) in PHEM buffer (60 mM PIPES, 10 mM EGTA, 25 mM HEPES, and 2 mM MgCl2, final pH 6.8) for 10 min at room temperature. Post fixation, cells were incubated with blocking solution (0.2% saponin + 5% FBS in PHEM buffer) at room temperature for 30 min, followed by three washes with 1X PBS. After this blocking step, cells were incubated with primary antibodies in staining solution (PHEM buffer + 0.2% saponin) for 1 hr at room temperature, washed thrice with 1X PBS, and further incubated for 30 min with Alexa fluorophore-conjugated secondary antibodies made in staining solution. Cells were washed thrice with 1X PBS and mounted in Fluoromount G (Southern Biotech). Single-plane confocal images were acquired using a 710 Confocal Laser Scanning Microscope (ZEISS) equipped with a Plan Apochromat 63×/1.4 NA oil immersion objective and high-resolution microscopy monochrome cooled camera AxioCamMRm Rev. 3 FireWire (D) (1.4 megapixels, pixel size 6.45 μm × 6.45 μm). For image acquisition, ZEN Pro 2011 (ZEISS) software was used. All images were captured to ensure that little or no pixel saturation is observed. The representative confocal images presented in figures were imported into Adobe Photoshop CS and formatted to 300 dpi resolution. The whole image adjustment of brightness was done using curves function.
For all the colocalization analysis, at least 30 cells for each treatment per experiment were used for three independent experiments. Pearson’s Correlation Coefficient (PCC) was determined using the JACoP plugin of ImageJ where the threshold was set using maximum entropy.
In order to trace the endocytic route, HeLa cells were incubated with Alexa-Fluor 647-conjugated dextran (Molecular Probes) for 16–18 hr. The cells were washed once with 1X PBS and infected with GFP-expressing Salmonella (at an MOI 50:1) and further incubated in a dextran-free medium for the rest of the experiment. Live-cell imaging was initiated at indicated time-points.
For live-cell imaging experiments, cells were seeded on glass-bottom tissue culture treated cell imaging dish (Eppendorf) and infected with either DsRed- or GFP- expressing Salmonella strains (at an MOI 50:1) as described above. Post-infection, imaging dish was loaded into a sealed live-cell imaging chamber (37°C and 5% CO2) for imaging in DMEM. Time-lapse confocal images were acquired at specified time-points using an LSM 710 confocal microscope with a LCI Plan Neofluar objective 63×/1.3 multi-immersion correction and equipped with a high-resolution microscopy monochrome cooled camera AxioCamMRm Rev. 3 FireWire (D). Image acquisition and adjustments to brightness and contrast was performed by using ZEN Pro 2011 software.
Sample processing and TEM was performed at the Harvard Medical School EM Facility (Boston, USA). Briefly, control shRNA or Vps41 shRNA transduced HeLa cells were infected with S. typhimurium SL1344 for 10 hr. Post-infection, cells were fixed in routine fixative (2.5% glutaraldehyde/1.25% PFA in 0.1 M sodium cacodylate buffer, pH 7.4) for at least 1 hr at room temperature and washed in 0.1 M sodium cacodylate buffer (pH 7.4). The cells were then post fixed for 30 min in 1% osmium tetroxide/1.5% potassium ferrocyanide, washed in water three times, and incubated in 1% aqueous uranyl acetate for 30 min, followed by two washes in water and subsequent dehydration in grades of alcohol (5 min each: 50, 70, 95, 2× 100%). Cells were removed from the dish in propylene oxide, pelleted at 3000 rpm for 3 min, and infiltrated overnight in a 1:1 mixture of propylene oxide and TAAB Epon (Marivac Canada). The samples subsequently embedded in TAAB Epon and polymerized at 60°C for 48 hr. Ultrathin sections were cut on a Reichert Ultracut-S microtome, picked up onto copper grids stained with lead citrate, and examined in a JEOL 1200EX transmission electron microscope. Images were recorded with an AMT 2k charge-coupled device camera.
Sample fixation for immunogold EM was carried out as described previously [26], and double immunogold labeling and imaging was performed at the Harvard Medical School EM Facility (Boston, USA). For preparation of cryosections, control siRNA- and Vps41 siRNA-treated HeLa cells were infected with S. typhimurium as described above. After 2 hr p.i., cells were transfected with HA-Rab7 expressing construct and 10 hr p.i. cells were fixed with 4% PFA + 0.1% glutaraldehyde (Glu) prepared in 0.1 M sodium phosphate buffer, pH 7.4. After 2 hr fixation at room temperature, the cell pellet was washed once with PBS and then placed in PBS containing 0.2 M glycine for 15 min to quench free aldehyde groups. Before freezing in liquid nitrogen, the cell pellets were cryoprotected by incubating in three drops of 2.3 M sucrose in PBS for 15 min. Frozen samples were sectioned at -120°C, and the sections were transferred to formvar/carbon-coated copper grids. Grids were floated on PBS until the immunogold labeling was performed.
The double immunogold labeling was performed at room temperature on a piece of parafilm. All the primary antibodies and Protein A immunogold were diluted in 1% Bovine Serum Albumin (BSA) in PBS. In brief, grids were floated on drops of 1% BSA for 10 min to block for unspecific labeling, transferred to 5 μl drops of rat anti-HA, and incubated for 30 min. The grids were then washed in four drops of PBS for a total of 15 min, transferred to 5 μl drops of rabbit anti-rat for 30 min, and washed again in four drops of PBS for 15 min, followed by 15 nm Protein A immunogold for 20 min (5 μl drops). After the 15 nm Protein A immunogold incubation, grids were washed in four drops of PBS, fixed for 2 min with 0.5% Glu followed by four drops of PBS containing 0.2 M glycine for 15 min to quench free aldehyde groups. The labeling process was repeated with rabbit anti-LAMP1 followed by 10 nm Protein A immunogold for 20 min in 5 μl drops. Finally, the grids were washed in four drops of PBS and six drops of double-distilled water. Contrasting/embedding of the labeled grids was performed on ice in 0.3% uranyl acetate in 2% methyl cellulose for 10 min. Grids were picked up with metal loops, and the excess liquids were removed by blotting with a filter paper and were examined in an electron microscope (1200EX; JEOL). Images were recorded with an AMT 2k CCD camera.
For lysates, cells were lysed in ice-cold lysis buffer (25 mM Tris-HCl pH 7.4, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100 and protease inhibitor cocktail). For co-IP experiments, HEK293T cells treated with control or gene-specific siRNA and transfected with indicated plasmids were lysed in ice-cold TAP lysis buffer (20 mM Tris, pH 8.0, 150 mM NaCl, 0.5% NP-40, 1 mM MgCl2, 1 mM Na3VO4, 1 mM NaF, 1 mM PMSF, and protease inhibitor cocktail; Sigma-Aldrich). The lysates were incubated with indicated antibody-conjugated agarose beads at 4°C rotation for 3 hr, followed by four washes in TAP wash buffer (20 mM Tris, pH 8.0, 150 mM NaCl, 0.1% NP-40, 1 mM MgCl2, 1 mM Na3VO4, 1 mM NaF, and 1 mM PMSF). The samples were then loaded on SDS-PAGE for further analysis. Protein samples separated on SDS-PAGE were transferred onto polyvinylidene fluoride (PVDF) membranes (Bio-Rad Laboratories). Membranes were blocked overnight at 4°C in blocking solution (10% skim milk in 0.05% PBS-Tween 20). Indicated primary and secondary antibodies were prepared in 0.05% PBS-Tween 20. The membranes were washed for 10 min thrice with 0.05% PBS-Tween 20 or 0.3% PBS-Tween 20 after 1 hr incubation with primary antibody and 1 hr incubation with secondary antibody, respectively. The blots were developed using a chemiluminescence-based method (Pierce).
To enumerate intracellular Salmonella growth, gentamicin protection assay was performed. Briefly, cells were infected with designated Salmonella strain for different time periods using the protocol described above. At the end of every time point p.i. cells were gently washed with PBS followed by lysis using PBS containing 0.1% Triton X-100 and 1% SDS for 5 min at room temperature. The resulting lysates were serially diluted and plated onto LB agar plates containing streptomycin.
To assess cytosolic Salmonella replication in HeLa cells upon Vps41 depletion, CHQ resistance assay was performed as described previously [56]. Briefly, control siRNA- or Vps41 siRNA-treated HeLa cells were seeded in 24-well plates and infected with S. typhimurium as described above. To evaluate cytosolic replication of Salmonella, 1 hr prior to 7 hr p.i. time point, two wells were treated with CHQ (150 μM) and gentamicin (5 μg/ml) for 1 hr (CHQ-resistant bacteria) and another two wells were incubated with gentamicin (5 μg/ml) only (total bacteria). At the end of 7 hr p.i. time point, duplicate gentamicin treated (total CFU) and duplicate CHQ + gentamicin-treated cells (cytosolic CFU) were solubilized and serial dilutions were plated on LB agar for CFU enumeration.
Six weeks old C57BL/6 male mice were obtained from the CSIR-Institute of Microbial Technology (IMTECH) animal house facility and injected intravenously (i.v.) with 12.5 mg/Kg (of body weight) of either control (CCTCTTACCTCAGTTACAATTTATA) or mouse VPS41-specific (CCATAGCGCAGCCTGAGAGTCAT) vivo-morpholinos (purchased from Gene Tools, LLC) for two consecutive days at an interval of 24 hr, followed by Salmonella infection the third day. For Salmonella infection, stationary phase culture of S. typhimurium strain SL1344 was diluted to a CFU of ~1.3X103 in 100 μl of 1X PBS and injected i.v. The infectious dose was quantified by plating plating dilution series on LB agar plates containing streptomycin. Mice were sacrificed after 3 days and dilution series of spleen and liver lysates (prepared in 0.05% sodium deoxycholate in 1X PBS) were plated on LB agar plates containing streptomycin.
This study was carried out in strict accordance to the guidelines issued by the Committee for the Purpose of Supervision of Experiments on Animals (No. 55/1999/CPCSEA) under the Prevention of Cruelty to Animals Act 1960 and amendments introduced in 1982 by Ministry of Environment and Forest, Government of India. All protocols involving mice experiments were approved by the Institutional Animal Ethics Committee (IAEC) of Council of Scientific and Industrial Research-Institute of Microbial Technology (Approval no. IAEC/16/12 and IAEC/17/13).
Roughly 50 million HeLa cells infected with S. typhimurium SL1344 strain were used for subcellular fractionation of SCVs. At 3 hr and 8 hr p.i., cells were washed thrice with ice-cold PBS and scrapped into a 15 ml centrifuge tube using a rubber cell scrapper. The cells were centrifuged at 1000 rpm for 7 min and the cell pellets were suspended in ice-cold homogenization buffer (250 mM sucrose, 20 mM HEPES (pH 7.2), 0.5 mM EGTA and 5 μg/ml Cytochalasin D) containing protease inhibitor cocktail (Sigma-Aldrich) and transferred to a Dounce Homogenizer with a tight fitting pestle on ice to break the cells. Approximately 30 strokes were applied until almost 90% of the cells were broken without breaking the nuclei. The intact cells and nuclei were pelleted in microcentrifuge tube at 400 x g for 3 min. The resulting supernatant was collected in a fresh microcentrifuge tube to yield the post nuclear supernatant (PNS). The PNS was brought to a final concentration of 39% sucrose and layered on to 2 ml 55% sucrose which was in turn layered onto 65% sucrose cushion in a 13.2 ml open top Beckman ultracentrifuge tube followed by addition of 2 ml 32.5% and 2 ml 10% sucrose solutions. All sucrose solutions (w/v) were prepared in 20 mM HEPES (pH 7.2) and 0.5 mM EGTA. The PNS layered on sucrose gradient was then subjected for ultracentrifugation in a swinging bucket rotor for 1 hr at 100000 x g at 4°C. The fractions of 1 ml each were collected from top to bottom. Pooled fractions 8–10 were adjusted very slowly to a final sucrose concentration of 11% with homogenization buffer without sucrose and layered on 15% Ficoll cushion (5% sucrose, 0.5 mM EGTA and 20 mM HEPES pH 7.2). The samples in open top Beckman ultracentrifuge tube were spun at 18000 x g for 30 min in a Beckman SW 41 Ti rotor at 4°C. The supernatant was discarded and pellet was resuspended in 11 ml homogenization buffer. The samples were spun again at 18000 x g for 20 min in a Beckman SW 41 Ti rotor at 4°C and the resulting pellet was labeled as “SCV” fraction. The pelleted SCV fractions were resuspended in 20 μl of 4X SDS-sample buffer, boiled for 10 min and analyzed by SDS-PAGE and immunoblotting.
A previously published three-step approach, lysis of infected host cells followed by intracellular compartment enrichment and affinity-IP, was used to enrich and determine the presence of HOPS subunits in SMMs [45]. Briefly, 16 hr prior to infection, 5 million HeLa cells were seeded in a 10-cm tissue culture dish and four 10-cm dishes were used per IP. For infection, S. typhimurium SL1344 sseF harboring a low-copy expression vector with a C-terminal HA-tagged SseF and its cognate chaperone sscB (sseF/SseF-HA) was used, and cells infection for a period of 8 hr was carried out as described above. Post-infection, cells were washed thrice with ice-cold PBS and scrapped into a 15 ml centrifuge tube using a rubber cell scrapper, and centrifuged at 1000 x g for 7 min. The resulting cells pellet was suspended in ice-cold homogenization buffer (250 mM sucrose, 20 mM HEPES, 0.5 mM EGTA, pH 7.4), centrifuged at 1000 x g for 10 min, and resuspended in 1 ml of 4°C pre-cooled homogenization buffer containing protease inhibitor cocktail (Sigma-Aldrich). The cells were mechanically disrupted by adding 100 μl of 0.5 mm glass beads (Sigma-Aldrich) using a vortexer (three 1 min strokes) with 5 min of intermediate cooling on ice. The lysate was centrifuged at 100 x g for 10 min at 4°C, and the resulting pellet (labeled as “GEMN pellet”) was washed twice with ice-cold homogenization buffer with protease inhibitor mixture. The final GEMN pellet was resuspended in 500 μl of homogenization buffer supplemented with 1.5 mM MgCl2 and treated with DNase I (50 μg/ml) for 30 min at 37°C. The protein concentration in the GEMN protein fraction was determined using Bradford reagent (Bio-rad).
For affinity-IP, 500 μg of GEMN proteins adjusted to a final volume of 500 μl in solubilization buffer (1.5 mM MgCl2, 10 mM KCl, 0.1% NP-40) were added to 20 μl of pre-blocked (in 1% BSA made in PBS for 30 min) anti-HA antibody-conjugated agarose beads or anti-myc antibody-conjugated agarose beads (Sigma-Aldrich) as a control, and were allowed to mix on rotary shaker at 4°C for 4 hr. At the end of the incubation period, beads were washed five times with 0.1% NP-40 made in PBS to remove non-specific proteins. Finally, the beads were resuspended in 20 μl of 4X SDS-sample buffer, boiled for 10 min and analyzed by SDS-PAGE and immunoblotting.
For protein expression and purification, bacterial expression vectors encoding for GST or GST tagged-SifA were transformed into E. coli BL21 strain. Primary cultures of a transformed single colony were set up for 12 hr at 37°C in LB broth containing plasmid vector antibiotic. Secondary cultures were set up using 1% primary inoculums and subjected to incubation at 37°C to an absorbance of 0.6 at 600 nm and then protein production was induced using 0.5 mM isopropyl β-D-1-thiogalactopyranoside (IPTG) for 5 hr at 30°C. After the incubation period, bacterial cultures were centrifuged at 4,000 rpm for 15 min, washed once with 1X PBS, and resuspended in ice-cold buffer (20 mM HEPES and 150 mM NaCl, pH 7.4) containing protease inhibitor tablet (Roche) and 1 mM PMSF. Cell lysis was performed by sonication, followed by centrifugation at 12,000 rpm for 15 min at 4°C. The supernatants were incubated with glutathione resin (Gbiosciences) on rotation for 2 hr at 4°C to allow binding of GST or GST tagged-SifA, followed by 10 washes with wash buffer (20 mM HEPES, 300 mM NaCl, and 0.5% Triton X-100, pH 7.4). For use in the pulldown assays, protein-bound glutathione resins were blocked with 5% BSA in PBS for 2 hr at 4°C.
For pulldown assays, transfected HEK 293T cells were lysed in ice-cold TAP lysis buffer, and lysates were incubated with protein-bound glutathione resins at 4°C for 3 hr with rotation. Samples were washed four times with TAP wash buffer, and elution was performed by boiling the samples in 1X SDS-PAGE loading buffer and loaded onto SDS-PAGE for analysis.
For the yeast two-hybrid assay, plasmids encoding GAL4-activation domain (AD) and GAL4-DNA binding domain (BD) fusion encoding constructs were co-transformed in S. cerevisiae AH109 strain, streaked on SD plates lacking leucine and tryptophan (SD-L/-W) and allowed to grow at 30°C for 3 days. The co-transformants were replated on non-selective medium (SD-L/-W) and selective medium (SD-leucine/-tryptophan/-histidine; SD-L/-W/-H) to assess interaction.
For measuring yeast growth rate, primary cultures were seeded in SD-L/-W broth (Clontech) from single colonies of S. cerevisiae AH109 strain co-transformed with indicated plasmids, and grown overnight at 30°C to saturation. The resulting cultures were diluted to approximately 0.1 OD (at 600 nm) in SD-L/-W/-H broth (Clontech) and culture growth was monitored every 4 hr for 48 hr.
For performing the yeast three-hybrid assay, the S. cerevisiae Gold strain (Clontech) was made sensitive to methionine (Met) by streaking the yeast on an SD/-Met media at least two times before transforming with the desired plasmids. After co-transformation, yeast cells were replated on SD-L/-W (nonselective; selects only for the presence of plasmid) or SD-L/-W/-H/-M (selective; requires interaction of bait and prey proteins through the linker protein for growth).
The acidotropic dye LysoTracker Red DND-99 (Thermo Fisher Scientific) was diluted in Opti-MEM without phenol red (Thermo Fisher Scientific). To control siRNA- or Vps41 siRNA-treated HeLa cells cultures 100 nM LysoTracker Red was added and uptake for 1 hr was performed. At the end of the internalization period, cells were washed and then resuspended in fresh pre-warmed medium, and red lysosomal fluorescence of 10,000 cells per sample was determined by flow cytometry (BD Accuri). FlowJo v10 software was used to analyze all of the data from flow cytometric experiments. For visualization of LysoTracker Red-uptake signal by confocal microscopy, at the end of the dye uptake cells were fixed in 4% PFA in PHEM buffer at room temperature for 10 min. Post-fixation, cells were washed and mounted on glass slides and analyzed.
Statistical analyses were performed with Prism 6 software (GraphPad). Data are presented as mean ± standard deviation (S.D.) unless otherwise indicated. P-values were calculated using two-tailed unpaired Student’s t test, and differences were considered significant when P < 0.05. The sample sizes are specified in the figure legends for all of the quantitative data.
|
10.1371/journal.ppat.1002276 | Multidrug Resistant 2009 A/H1N1 Influenza Clinical Isolate with a Neuraminidase I223R Mutation Retains Its Virulence and Transmissibility in Ferrets | Only two classes of antiviral drugs, neuraminidase inhibitors and adamantanes, are approved for prophylaxis and therapy against influenza virus infections. A major concern is that influenza virus becomes resistant to these antiviral drugs and spreads in the human population. The 2009 pandemic A/H1N1 influenza virus is naturally resistant to adamantanes. Recently a novel neuraminidase I223R mutation was identified in an A/H1N1 virus showing cross-resistance to the neuraminidase inhibitors oseltamivir, zanamivir and peramivir. However, the ability of this virus to cause disease and spread in the human population is unknown. Therefore, this clinical isolate (NL/2631-R223) was compared with a well-characterized reference virus (NL/602). In vitro experiments showed that NL/2631-I223R replicated as well as NL/602 in MDCK cells. In a ferret pathogenesis model, body weight loss was similar in animals inoculated with NL/2631-R223 or NL/602. In addition, pulmonary lesions were similar at day 4 post inoculation. However, at day 7 post inoculation, NL/2631-R223 caused milder pulmonary lesions and degree of alveolitis than NL/602. This indicated that the mutant virus was less pathogenic. Both NL/2631-R223 and a recombinant virus with a single I223R change (recNL/602-I223R), transmitted among ferrets by aerosols, despite observed attenuation of recNL/602-I223R in vitro. In conclusion, the I223R mutated virus isolate has comparable replicative ability and transmissibility, but lower pathogenicity than the reference virus based on these in vivo studies. This implies that the 2009 pandemic influenza A/H1N1 virus subtype with an isoleucine to arginine change at position 223 in the neuraminidase has the potential to spread in the human population. It is important to be vigilant for this mutation in influenza surveillance and to continue efforts to increase the arsenal of antiviral drugs to combat influenza.
| Recently, a 2009 pandemic A/H1N1 influenza virus was isolated from an immune compromised patient, with antiviral resistance to the neuraminidase inhibitor class of drugs. This virus had an amino acid change in the viral neuraminidase enzyme; an isoleucine at position 223 was substituted for an arginine (I223R). Patients infected with a pandemic virus that is resistant to all neuraminidase inhibitors, would leave physicians without antiviral treatment options, since these viruses are naturally resistant to the other class of antivirals, the adamantanes. To date, it is unknown if this I223R mutant virus is affected in its ability to cause severe disease and to transmit to other humans. Therefore, we have addressed this question by comparing the I223R mutant virus with a wild type reference virus in a ferret pathogenicity and transmission model. We found that the I223R mutant virus was not severely affected in its pathogenicity, although fewer lung lesions and alveolitis scores were found for the I223R mutant virus. In addition, we demonstrated that this virus transmitted efficiently to naïve ferrets. Consequently, we conclude that this I223R mutant virus has the potential to cause disease and may spread among humans. Therefore, influenza surveillance for this resistance pattern is advised.
| Two classes of antiviral drugs are approved for prophylaxis and therapy of influenza virus infected patients [1]. Antiviral therapy against the new (swine-origin) 2009 pandemic A/H1N1 influenza virus relies on the neuraminidase inhibitor (NAI) class of antiviral drugs only, because this subtype is resistant to the adamantane class (amantadine and rimantadine) of drugs [2]. In 2009 pandemic influenza viruses, this resistance pattern is mainly caused by an asparagine at amino acid position 31 (N31) in the viral M2 membrane protein. Fortunately, NAI treatment, both as prophylaxis and therapy, has been shown to be effective against most 2009 pandemic H1N1 virus infections so far [3], [4].
To date, the incidence of NAI resistant 2009 pandemic A/H1N1 viruses is very low. Nevertheless, 565 cases of patients infected with an (H275Y, N1 numbering) oseltamivir (OS) resistant virus have been reported to the World Health Organization [5]. In most of these cases, OS resistance was found in patients receiving prolonged antiviral therapy, in particular patients under immunosuppressive therapy [6]. The H275Y mutant viruses are cross-resistant to peramivir (PER), but remain susceptible to zanamivir (ZA). Successful clearance of a H275Y mutant virus from a patient treated with ZA was reported previously [7].
Within the first years after approval of the NAIs in 1999, antiviral resistance in influenza viruses at a population level was rare (0.4%). In clinical trials, the incidence of resistant viruses was higher, varying from 0.4 to 1% in adults and up to 18% in young children [8], [9]. However, a dramatic increase, up to 100%, of de novo circulating oseltamivir-resistant A/H1N1 viruses characterized the epidemic seasons of 2007-2008 and 2008-2009 [10], [11]. This resistance phenotype was also caused by a H275Y mutation. Remarkably, earlier studies on H275Y mutant H1N1 viruses had characterized these viruses as attenuated and not of clinical importance [12], [13], [14]. The resistant viruses from 2007-2008 did not seem to be affected in replication capacity, transmissibility and their ability to cause severe disease in humans [15], [16], [17]. A compensatory role was assigned to the NA amino acid changes V234M, R222Q and D344N [18], [19]. These substitutions may have restored the initial loss of NA activity due to the NAI resistance mutation and facilitated the appearance of the H275Y change in the epidemic influenza A/H1N1 viruses that circulated before the 2009 outbreak of the new pandemic virus. Recently, several research groups have studied the fitness of H275Y mutant pandemic influenza A/H1N1 viruses using both in vitro and in vivo experiments [20], [21], [22], [23], [24]. Overall, these data indicate that pandemic viruses with the NA H275Y substitution were comparable to their oseltamivir susceptible counterparts in pathogenicity and transmissibility in animal models.
Recently, the identification of a novel multidrug resistant 2009 pandemic A/H1N1 virus was reported, isolated from an immune compromised child with reduced susceptibility to all NAIs [25]. An isoleucine to arginine substitution at position 223 in NA (I223R, N1 numbering) was detected in the patient after antiviral therapy with OS had failed due to the emergence of the H275Y mutation and therapy was switched to ZA. This I223R containing isolate, in which the H275Y mutation had disappeared, showed reduced susceptibility to OS (45-fold), PER (7-fold) and ZA (10-fold). In vitro analysis showed that reversion of the arginine to isoleucine fully restored NAI susceptibility. In another case, an I223R/H275Y double mutant virus was isolated that showed high resistance to the NAIs [26]. In combination with the natural resistance of pandemic A/H1N1 viruses to adamantanes, an infection of such a multi-drug resistant virus leaves physicians without antiviral treatment options. The emergence of this pandemic 2009 A/H1N1 virus prompted us to investigate the properties of this clinical isolate by evaluating its in vitro replication kinetics and its pathogenicity and transmissibility in the ferret model. We here show that this 2009 pandemic influenza A/H1N1 clinical isolate, harboring a neuraminidase I223R substitution retains its virulence and transmissibility, but is less pathogenic than a virus prototype without this mutation. In addition, recombinant NL/602/09 with a single I223R amino acid substitution transmitted as well as its recombinant parental virus, suggesting that no additional mutations are needed to compensate for the presence of this I223R mutation in the 2009 pandemic A/H1N1 virus backbone.
A pandemic 2009 influenza virus with reduced susceptibility to all NAIs that was isolated from a Dutch immune compromised child was studied here. Full genome sequencing of this clinical isolate A/NL/2631_1202/2010 (NL/2631-R223, GenBank accession numbers JF906180-906187) harboring an I223R mutation in the neuraminidase was performed. Since no drug susceptible virus had been isolated from this patient before start of antiviral therapy, the well-characterized NAI-susceptible virus isolate A/NL/602/2009 (NL/602, GenBank accession numbers CY046940-046945 and CY039527-039528) was used as a reference virus in all experiments. This reference virus is a representative of pandemic H1N1 viruses that circulated in 2009, with only amino acid changes I108V and V407I (N1 numbering) in NA being unusual among the deposited sequences in the Influenza Research Database [27], [28]. Pair-wise comparison revealed, in addition to the amino acid change I223R, 5 amino acid differences in NA (V106I, V108I, N248D, N386D and I407V) and 1 in HA (S203T). The NA and HA amino acid positions are given according to the N1 and H1 numbering. Eleven additional amino acid differences were found in gene segments PB2 (3), PB1 (2), PA (2), NP (3) and NS (1) compared to NL/602. None of these mutations have previously been identified as a virulence marker or as a compensatory mutation involved in restoration of NA activity loss, as a result of the presence of resistance mutations. By studying these isolates, a direct comparison could be made between a NAI susceptible and a novel I223R resistant virus, but such comparison does not address the impact of the single I223R mutation directly. Therefore, we introduced the I223R mutation in the recNL/602 backbone, resulting in the drug-resistant recNL602-I223R, to evaluate the impact of the single I223R mutation on virus replication, virus shedding from the upper respiratory tract and transmissibility in the ferret model.
Virus replication was studied in vitro by multi-cycle replication kinetics of the viruses of interest. For this purpose, MDCK or MDCK-SIAT1 cell cultures were inoculated at a multiplicity of infection of 0.001 TCID50 per cell and at fixed time points supernatants were harvested to determine viral titers (Figure 1). Overall, the initial virus replication rates and end point titers were similar for the clinical isolate NL/2631-R223 and recNL/602.
A recombinant derivative of NL/602 with the I223R mutation in NA (recNL/602-I223R) replicated to lower peak titers in both cell lines compared to recNL/602 and NL/2631-R223. In addition, initial virus replication of recNL/602-I223R was delayed by 6 to 12 hours in MDCK-SIAT1 cells.
The pathogenicity of clinical isolate NL/2631-R223 was compared with NL/602 in the ferret model that was previously established to study the ability of influenza viruses to cause pneumonia [29]. Two groups of 6 ferrets were inoculated intratracheally with 106 TCID50 of virus. The animals were weighed daily as an indicator of disease. Over the 7-day period, no significant differences were observed in weight loss between the two groups inoculated with either virus. At day 4 post infection (p.i.), when there were still 6 animals present in each group, the mean percentage of weight loss was 8,2±2,4% and 7,6±6,7% for NL/602 and NL/2631-R223-inoculated animals respectively, not statistically significant (Figure 2A and B). In addition, no marked differences were observed for other clinical parameters, such as lethargy, sneezing and interest in food.
Nose and throat swabs were collected daily from the inoculated animals and virus titers were determined by end-point titration in MDCK cells. Infectious virus shedding from the throat was detected from day 1 p.i. onwards in all ferrets, with similar patterns of virus shedding from the throat of the animals in the two groups (Figure 2C). At day 4 p.i., 5 and 4 animals were shedding virus from the nose in the NL/602 and NL/2631-R223 inoculated group respectively (Figure 2D). Sequence analysis confirmed the presence of the I223R mutation in the respiratory samples collected at day 7 p.i. from the NL/2631-R223 inoculated ferrets.
At day 4 and 7 p.i., three animals of each group were euthanized and lungs were collected for virological and pathological examination. At day 4 p.i., no marked differences were found between the virus titers for both groups of ferrets (Figure 3A). At day 7 p.i., no virus was detected in the lungs of ferrets inoculated with either virus.
Gross pathology of the lungs of all animals revealed pulmonary lesions at day 4 and 7 p.i. (Figure 3B). At day 4 p.i., no marked difference was observed between the groups, but at day 7 p.i., the percentage of affected lung tissue was higher in the group inoculated with NL/602. The mean relative lung weight increased from day 4 to day 7, with no difference between the animals inoculated with NL/602 or NL/2631-R223 (Figure 3C). Histopathological examination of the lungs showed multifocal to coalescing alveolar damage in both groups characterized by the presence of macrophages and neutrophils within the lumina and thickened alveolar walls. At day 4 p.i., the severity of alveolitis did not differ between the two groups (Figure 4D). However, in agreement with the increased percentage of affected lung tissue at day 7 p.i. (Figure 3B), also higher alveolitis scores were determined for the NL/602 inoculated animals at day 7 p.i. (Figure 4D).
The bronchial and bronchiolar epithelium from ferrets in both groups showed slight multifocal necrosis with moderate intra-epithelial infiltrates of neutrophils and multifocal peribronchiolar infiltration of macrophages, lymphocytes, neutrophils and plasma cells. The lumina contained moderate amounts of mucus mixed with cellular debris and few neutrophils. The tracheal epithelium in both groups showed mild neutrophilic infiltrates. The severity of both bronchiolitis and tracheitis increased from day 4 to 7 p.i. in ferrets infected with both viruses, but the differences in scores between groups were minimal (Figure 4E and F).
Individually housed ferrets were inoculated with virus isolate NL/2631-R223 or NL/602 and naïve animals were placed in a cage adjacent to each inoculated ferret at day 1 p.i. to allow aerosol or respiratory droplet transmission. All inoculated ferrets started to shed virus at day 1 p.i. with virus titers up to 106 TCID50/ml in throat and nose swabs (Figure 4A and C).
The naïve ferrets became infected, because of aerosol or respiratory droplet transmission, 1, 2 or 3 days p.e. In the naïve animals, virus was detected in 4 (NL/602), or 3 (NL/2631-R223) out of 4 animals (Figure 4B and D). The exposed animal in the NL/2631-R223 transmission experiment, from which no virus could be isolated, did not seroconvert in the course of the experiment. At day 5 p.e., the presence of the I223R mutation was confirmed by sequencing the NA gene of virus isolated from the throat swabs of the positive animals.
When the multi-cycle replication kinetics were studied of viruses with or without the I223R substitution in MDCK cells, it was noticed that the recombinant virus in which the I223R mutation was introduced, recNL/602-I223R, replicated to lower titers than its parental virus recNL/602 (Figure 1). To address if this difference in in vitro replication capacity could be extrapolated to reduced replication in vivo, the ability of recNL/602-I223R to transmit in the ferret model was studied. It was expected that reduced replication in ferrets would impede the virus to transmit to naïve animals, thereby suggesting that compensatory mutations are needed to balance the fitness loss induced by the I223R mutation. In contrast to the results obtained in MDCK cells, recNL/602-I223R replicated and transmitted as well as recNL/602 when evaluated in the ferret transmission model. Inoculated animals started to shed virus from the upper respiratory tract from day 1 p.i. onwards and transmission was detected in 4 out of 4 (recNL/602), or 2 out of 2 (recNL/602-I223R) naïve animals from day 2 onwards (Figure 5). The presence of the I223R mutation in the recNL/602 backbone was confirmed in throat samples obtained from these animals at day 5 p.e.
Here, a 2009 pandemic influenza A/H1N1 virus isolate, harboring an I223R multidrug resistance mutation, was characterized by studying its replication capacity in MDCK cells and its pathogenicity and transmissibility in the ferret model. This I223R mutant virus is not attenuated for replication in the ferret respiratory tract and transmitted as well as NAI susceptible reference virus NL/602. Furthermore, it was demonstrated here that compensatory mutations for the I223R mutation are not required, since recombinant NL/602 with a single I223R change transmitted as efficiently as its parental virus in ferrets.
To date, 2009 pandemic viruses with an amino acid substitution at position 223 have only sporadically been isolated from patients. A I223V/H275Y double mutant was detected in two closely residing patients who were treated with OS [30]. Besides the I223R single mutant virus studied here, an I223R/H275Y double mutant was detected in an immune suppressed patient treated with OS and ZA [26]. The combination of these mutations resulted in an increased NAI resistance pattern, as compared to the resistance induced by the single mutations. This emphasizes that neuraminidase position 223 is an important marker for antiviral resistance and may be a key residue in the emergence of influenza viruses with resistance to all NAIs, especially in combination with other resistance-associated mutations. So far, the incidence of 2009 pandemic viruses with a 223 change is very low. Notably, 2009 pandemic viruses were reported with a serine to asparagine change at position 247 [31]. In combination with the H275Y change, these viruses demonstrated resistance patterns similar to the I223R/H275Y mutant.
In a pathogenesis experiment, no statistical significant differences were found when weight loss was compared of ferrets inoculated with clinical isolates NL/2631-R223 or NL/602 (Figure 2A and B). In agreement with high viral loads found in respiratory specimens collected from the patient who was infected with NL/2631-R223, high viral loads were detected in the throat of animals inoculated with the same virus. Overall, identical patterns of virus shedding were observed during the course of the experiment in the throats of animals inoculated with either virus. However, virus shedding from the nose could not be detected in all inoculated animals. Although virus shedding from the nose of NL/2631-R223-inoculated animals seem somewhat delayed in comparison with NL/602-inoculated animals, these differences were not significant due to the large variations within groups and small group size after day 4 p.i. (Figure 2C and D).
Both macroscopic and microscopic evaluation of the lungs of the ferrets at day 4 p.i., revealed no major differences in the percentage of affected lung tissue and relative lung weights between NL/2631-R223 and NL/602 (Figure 3B and C). However, at day 7 p.i. the lungs of ferrets inoculated with NL/2631-R223 had not further deteriorated, whereas the percentage of affected lung tissue had increased to 50% in the NL/602 inoculated animals (Figure 3B). This higher score for affected lung tissue in the NL/602-inoculated animals was also reflected by the higher score for the degree of alveolitis at day 7 p.i. compared to day 4 p.i., whereas the alveolitis scores in the NL/2631-R223-inoculated animals at day 4 and 7 p.i. were similar. To recapitulate, both viruses replicated to the same extent in the respiratory tract of ferrets, but the NL/2631-R223 seemed less pathogenic compared to the NL/602 virus.
Despite the moderate pathogenicity of NL/2631-R223, this virus transmitted to 3 out of 4 exposed animals via aerosols or respiratory droplets (Figure 4B). This result is comparable to the data obtained from NL/602, in which 4 out of 4 exposed animals got infected (Figure 4D) [27]. This ferret transmission model was designed as a qualitative model for transmission and with the limited number of animals, quantitative information on virus transmission could not be obtained. Therefore, from these experiments it was concluded that both NL/2631-R223 and NL/602 transmitted via aerosols or respiratory droplets, although a delay in virus shedding by approximately 1 day was observed in the naïve animals exposed to NL/2631-R223 (Figure 4B and D).
When the impact of the single I223R mutation in the recombinant NL/602 backbone on in vitro replication kinetics was evaluated, a reduction in virus replication in MDCK cells was noticed (Figure 1). In addition, the initial virus replication of NL/602-I223R on MDCK-SIAT1 cells started 6 to 12 hours later as compared to its parental virus (Figure 1B).
These results suggested that compensatory mutations may be required to accommodate the isoleucine to arginine substitution at position 223 in NA and emphasizes the importance of the viral backbone used to study resistance-associated mutations. However, when recNL/602-I223R was tested in the ferret transmission model, the virus transmitted to 2 out of 2 exposed animals (Figure 5B). When these results were compared with transmission data of recNL/602 (Figure 5D) [32], no differences were found in the onset of virus shedding and virus titers that were detected in the collected throat and nose swabs from the exposed animals. This observation demonstrates that the transmissibility of recNL/602-I223R is not significantly diminished or can at least not be studied using a ferret transmission model.
Although these results suggest that introduction of the I223R does not attenuate the virus, it cannot be ruled out that other mutations than 223R in NL/2631-R223 may have compensated for the initial loss of fitness due to the I223R mutation. Sequence comparison revealed 5 amino acid differences between NL/2631-R223 and NL/602. The only amino acid substitution that is located near the active site of the neuraminidase is at position 248, where NL/602 harbors an aspartic acid and NL/2631-R223 an asparagine. Interestingly, neighboring residue 247 has been linked to NAI resistance in combination with the H275Y mutation [31]. Further research is needed to study the I223R resistance mechanism in competitive mixture experiments and potential co-mutations on a molecular level [33].
To note, small differences between NL/602 and recNL/602 could be observed in replication capacity and transmission patterns in ferrets (Figure 4 and 5). Previously, differences were also found in pathogenesis experiments, where the wild type NL/602 was detected more abundantly in the lower airways of ferrets than recNL/602 [34]. These observed differences may be a result of the use of a virus isolate rather than a virus generated by reverse genetics and to a different batch of ferrets used in the different studies. A direct comparison between virus isolates and recombinant viruses can, therefore, not be made.
The different inoculation routes and inoculation doses used for influenza research is subject of debate. The intratracheal route of inoculation is often used to study pathogenicity or to study the efficacy of vaccines to prevent lower respiratory tract infection. In contrast, the intranasal route of inoculation is used when transmissibility is studied. Unfortunately, these inoculation routes and inoculation doses do not accurately mimic the natural way of infection and may mask the fitness differences between the drug-resistant and drug sensitive viruses.
However, the recipient animals in the transmission experiment are infected via the natural route; aerosols or respiratory droplets shed by the donor ferret. The virus secretion pattern, which is the combination of the amount of virus secreted and the duration of virus shedding from the upper respiratory tract, of animals exposed to recNL/602 and rec/NL602-I223R are similar. This suggests that no marked differences in viral fitness are introduced by the single I223R mutation.
The present study demonstrates for the first time that a 2009 pandemic A/H1N1 clinical isolate containing a resistance mutation at position 223 in the NA is not attenuated in its replication capacity and transmissibility in a ferret model. Although the pathogenicity of this virus seems less severe compared to a relevant reference virus in the ferret model, it is unclear whether this moderate pathogenicity has implications for infections with multidrug-resistant viruses in humans. Continuous surveillance is needed to monitor the emergence of (novel) influenza viruses with reduced susceptibility to the NAIs or mutations that may facilitate the emergence of circulating multi drug resistant influenza viruses.
Animals were housed and experiments were conducted in strict compliance with European guidelines (EU directive on animal testing 86/609/EEC) and Dutch legislation (Experiments on Animals Act, 1997). All animal experiments were approved by the independent animal experimentation ethical review committee ‘stichting DEC consult’ (Erasmus MC permit number EUR1821) and were performed under animal biosafety level 3+ conditions. Animal welfare was observed on a daily basis, and all animal handling was performed under light anesthesia using ketamine to minimize animal suffering. Influenza virus seronegative 6-month-old female ferrets (Mustella putorius furo), weighing 800–1000 g., were obtained from a commercial breeder.
Madin-Darby Canine Kidney (MDCK) cells were obtained from American Type Culture Collection. MDCK-SIAT1 cells, constitutively expressing the human 2,6-sialyltransferase (SIAT1), were kindly provided by Professor H.D. Klenk, Philipps University Marburg [35]. Both cell lines were cultured in Eagle’s minimal essential medium (EMEM) (Lonza, Breda, The Netherlands) supplemented with 10% fetal calf serum (FCS), 100 IU/ml penicillin, 100 µg/ml streptomycin, 2mM glutamine, 1.5mg/ml sodium bicarbonate (Cambrex), 10 mM HEPES (Lonza) and non-essential amino acids (MP Biomedicals Europe, Illkirch, France). In addition, MDCK-SIAT1 cells were cultured in the presence of 1 mg of antibiotic G418/ml. Influenza virus A/Netherlands/2631_1202/2010 (NL/2631-R223) was isolated from a 5-year-old immune compromised child [25]. Clonal virus of this isolate was obtained by passaging this virus 3 times under limiting diluting conditions in MDCK cells. Full genome sequencing after the last MDCK passage confirmed the absence of mutations. Influenza A/Netherlands/602/2009 (NL/602) was characterized previously [27]. All eight segments of this virus were cloned in a bidirectional reverse genetics plasmid pHW2000 and used to generate recombinant viruses by reverse genetics as described previously [36]. The I223R mutation was introduced in the NA gene of NL/602 using QuickChange multi site-directed mutagenesis kit (Stratagene, Leusden, The Netherlands) resulting in recombinant viruses recNL/602-I223R. The presence of this mutation was confirmed by sequencing.
Virus titers in nasal and throat swabs, homogenized tissue samples, or samples for replication curves were determined by endpoint titration in MDCK cells. MDCK cells were inoculated with 10-fold serial dilutions of each sample, washed 1 hour after inoculation with phosphate-buffered saline (PBS), and grown in 200 µl of infection medium, consisting of EMEM supplemented with 100 U/ml penicillin, 100 µg/ml streptomycin, 2 mM glutamine, 1.5 mg/ml sodium bicarbonate, 10 mM HEPES, nonessential amino acids, and 20 µg/ml trypsin (Lonza). Three days after inoculation, the supernatants of inoculated cell cultures were tested for agglutinating activity using turkey erythrocytes as an indicator of virus replication in the cells. Infectious-virus titers were calculated from 4 replicates by the method of Spearman-Kärber [37].
Multi-cycle replication curves were generated by inoculating MDCK or MDCK-SIAT1 cells at a multiplicity of infection (MOI) of 0.001 50% tissue culture infectious dose (TCID50) per cell. One hour after inoculation, at time point 0, the cells were washed once with PBS, and fresh infection medium was added. The supernatants were sampled at 6, 12, 24, and 48 h post infection and the virus titers in these supernatants were determined by means of endpoint titration in MDCK cells.
For the pathogenesis experiment, statistical analysis was done for each time point, until 4 days after inoculation (when there were still 6 animals present in each group). The Mann-Whitney-U test was used to compare weight losses and virus shedding of the six animals in both groups. P-values less than 0.05 were considered significant.
|
10.1371/journal.pgen.1006511 | Anillin Phosphorylation Controls Timely Membrane Association and Successful Cytokinesis | During cytokinesis, a contractile ring generates the constricting force to divide a cell into two daughters. This ring is composed of filamentous actin and the motor protein myosin, along with additional structural and regulatory proteins, including anillin. Anillin is a required scaffold protein that links the actomyosin ring to membrane and its organizer, RhoA. However, the molecular basis for timely action of anillin at cytokinesis remains obscure. Here, we find that phosphorylation regulates efficient recruitment of human anillin to the equatorial membrane. Anillin is highly phosphorylated in mitosis, and is a substrate for mitotic kinases. We surveyed function of 46 residues on anillin previously found to be phosphorylated in human cells to identify those required for cytokinesis. Among these sites, we identified S635 as a key site mediating cytokinesis. Preventing S635 phosphorylation adjacent to the AH domain disrupts anillin concentration at the equatorial cortex at anaphase, whereas a phosphomimetic mutant, S635D, partially restores this localization. Time-lapse videomicroscopy reveals impaired recruitment of S635A anillin to equatorial membrane and a transient unstable furrow followed by ultimate failure in cytokinesis. A phosphospecific antibody confirms phosphorylation at S635 in late cytokinesis, although it does not detect phosphorylation in early cytokinesis, possibly due to adjacent Y634 phosphorylation. Together, these findings reveal that anillin recruitment to the equatorial cortex at anaphase onset is enhanced by phosphorylation and promotes successful cytokinesis.
| Human diseases such as cancer and congenital trisomies arise from loss of genetic material during cell division. Yet in most divisions, cells preserve their genetic integrity by strict coordination of cell membrane cleavage (cytokinesis), with accurate separation of genetic material (mitosis). Thus, understanding how mitosis and cytokinesis are coordinated can provide insight into human disease. Anillin is one central integrator of cytokinesis that is recruited in a ring-like structure on the membrane of dividing cells. Protein phosphorylation is a common mechanism regulating timing of events in mitosis; although 46 phosphorylation sites have been mapped on anillin, their functional significance is unknown. Here, we evaluated the effect of blocking anillin phosphorylation on human cell division. Surprisingly, most phosphorylation events appeared dispensable for cytokinesis in the assay used. By contrast, phosphorylation of S635 is important for early recruitment of anillin to the midzone membrane, furrow stabilization and efficient cytokinesis. Our findings highlight a central mechanism regulating the timing of human cytokinesis.
| In cytokinesis, cells assemble and stabilize an actomyosin ring between segregated chromosomes to generate daughter cells. The positioning of the cleavage furrow is controlled by negative signals from astral microtubules and positive signals established from the central spindle [1, 2]. At anaphase onset, inactivation of cyclin-dependent kinase 1 (Cdk1) triggers recruitment of centralspindlin to the central spindle and adjacent equatorial cell membrane [3–5]. Centralspindlin recruits the RhoGEF, epithelial cell transforming sequence 2 (Ect2), to locally activate the small GTPase RhoA [2, 4, 6] which specifies localization of the contractile actomyosin ring [7]. Temporal recruitment and activation of the centralspindlin apparatus depends on phosphorylation, including Cdk1-dependent phosphorylation sites on mitotic kinesin-like protein 1 (MKLP1) and Ect2, lost at anaphase onset [3, 6], and anaphase-specific recruitment of Plk1 through phosphorylation of protein regulating cytokinesis 1 (PRC1) [8]. Phosphorylation is likely to regulate additional events in cytokinesis.
Anillin is a key scaffold protein linking the actomyosin ring to the equatorial membrane [9–11]. Anillin binds myosin and F-actin at the N-terminus, and has anillin homology (AH) and pleckstrin homology (PH) domains at the C-terminus (Fig 1A). Anillin’s myosin and F-actin binding domain are required for organization of the actomyosin ring [12, 13]. The AH domain of anillin binds RhoA and shares homology with Rhotekin, a RhoA-GTP binding domain [10]. Drosophila anillin also binds to RacGAP50C, a homologue of MgcRacGAP, suggesting that it may couple the central spindle to a contractile ring [14, 15]. The C-terminal PH domain is important for recruitment of anillin to the equatorial membrane. Recently, a cryptic C2 domain within the AH domain was discovered and, with adjacent PH domain, promotes efficient recruitment to the membrane [16]. Thus, anillin is a hub for midzone membrane regulators and effectors of cytokinesis. Highlighting its required role in these processes, anillin depletion results in cytokinesis failure [13, 17, 18]. Despite its importance, poorly defined mechanisms restrict anillin to operate specifically at cytokinesis and at the equatorial membrane.
Anillin localization varies with the cell cycle. In interphase, anillin primarily localizes to the nucleus [9, 12], but upon entry into mitosis it re-localizes uniformly to the cell cortex. At anaphase onset, anillin is lost from the poles and concentrates at the equatorial zone, prior to onset of cytokinesis [19]. The loss adjacent to the poles is explained in part by Ran-GTP signals which emanate from chromatin [20]. One candidate for anillin recruitment is concentration of its preferred lipid phosphatidylinositol 4,5-bisphosphate (PI(4,5)P2) at the equatorial membrane [21]. Locally activated RhoA promotes anillin recruitment to the equatorial membrane and anillin stabilizes RhoA against extraction with trichloroacetic acid (TCA) [10, 18, 19]. Although these may be sufficient to specify timely recruitment of anillin, additional regulatory mechanisms of temporal-spatial control could reinforce these signals.
Protein phosphorylation is a common mechanism of regulating localization and function in mitosis. Anillin displays phosphorylation-induced band retardation in SDS-PAGE [22] and large-scale phosphoproteomics identified numerous phosphorylation sites [23, 24]. In fission yeast, the Polo-like kinase Plo1 binds to and phosphorylates an anillin-like protein Mid1, facilitating contractile ring assembly [25]. Similarly, phosphorylation sites of S. cerevisiae anillin homolog, Bud4, by Cdk1 promotes cytokinesis [26]. However, human anillin is divergent from its yeast counterparts, and several phosphorylation sites are not conserved, suggesting that the regulation may likewise differ. We hypothesized that phosphorylation of anillin governs its required function in human cytokinesis.
Here we report that human anillin is indeed phosphorylated in mitosis, and that phosphorylation is required for proper cytokinesis. To evaluate functional significance of phosphorylation, we used a stringent assay to comprehensively evaluate non-redundant phosphorylation events on anillin. Surprisingly, most phosphorylation events are dispensable for cytokinesis. However, we identify phosphorylation of a single residue, S635 as a key determinant of anillin localization and function. This phosphorylation site controls its ability to efficiently concentrate on the equatorial plasma membrane, and maintain a stable cytokinetic furrow. This finding suggests that phosphorylation at this site controls timely equatorial accumulation of anillin required for cytokinesis.
To investigate anillin phosphorylation we first characterized how it is modulated in human mitosis. In mitotic HeLa extracts, anillin migrated slowly on SDS-PAGE compared to its S-phase counterpart (Fig 1B). This shift was reversed by Lambda phosphatase, indicating the retardation is attributable to phosphorylation. Consistent with phosphorylation being specific to mitosis, the slow-migrating band diminished upon mitotic exit (Fig 1B, right lanes). These observations were not specific to HeLa; similar mitotic shifts were observed for anillin in other cancer cells and non-transformed human cells (Fig 1C). We conclude that anillin is phosphorylated significantly in human mitosis.
To identify kinases that might be responsible for mitotic phosphorylation of anillin, we purified recombinant GST-tagged fragments and performed in vitro kinase assays with the mitotic kinases polo-like kinase 1 (Plk1), Aurora B, and Cdk1. Under these conditions, all phosphorylated anillin, though on different domains: Plk1 and Cdk1 phosphorylated the 371–607 fragment, whereas Aurora B preferred the PH domain (Fig 1D). To test if one kinase is primarily responsible for the electrophoretic mobility shift, we employed specific inhibitors of Plk1, Aurora B, and Cdk1. Inhibition of Cdk1 with RO-3306 restored fast electrophoretic mobility not seen with other inhibitors (S1 Fig). Thus phosphorylation events that shift anillin mobility are governed directly or indirectly by Cdk1, although these findings do not exclude possible roles by Plk1 or Aurora B, as some phosphorylation events might not significantly alter shift. Additionally, other kinases could operate on anillin, such as Citron or Rho kinases, which have a known role in cytokinesis [27]. We conclude that anillin is phosphorylated in mitosis and is regulated by mitotic kinases, consistent with a possible regulatory role in cytokinesis.
As multiple mitotic kinases can phosphorylate anillin, we focused on residues known to be phosphorylated without kinase bias. From two phosphorylation site databases, we identified 46 sites, the majority discovered by phosphoproteomic mass spectrometry [23, 24]. These 46 sites (Ser, Thr or Tyr) were fragmented into 7 subdomains, and we made constructs for each, replacing all with non-phosphorylatable residues (Ala, Val or Phe, respectively) (Fig 2A). For the mutants containing a large number of phosphorylation sites (A1-A5), we employed the Gibson assembly method to assemble blocks of synthetic double-stranded DNA [28]. To study the functions of these phosphorylation events, we optimized an RNAi knockdown and transgene rescue assay (Fig 2B). This assay was developed under conditions to establish the most robust difference between the GFP and WT controls to allow functional assessment of anillin mutants. In an asynchronous population of cells fixed after 48 h of RNAi treatment, 79% of GFP-positive cells were multinucleate. This phenotype was efficiently suppressed by coexpression of siRNA-resistant wild-type anillin (Fig 2C and 2D). Under these conditions, the GFP-tagged constructs in panel A were assessed for their ability to restore cytokinesis. Strikingly, most non-phosphorylatable constructs suppressed multinucleated cells similar to wild type. However, the A5 mutant had a two-fold increase in multinucleation, despite efficient expression. To evaluate the efficacy of knockdown/addback under these conditions, we performed quantitative immunoblotting with infrared secondary antibodies (Fig 2E), revealing that knockdown was ~80% effective, and that transfection added back anillin below endogenous levels. We conclude that most phosphorylation sites are dispensable in this assay, but that one or more of these sites in the g5 domain play a crucial non-redundant role in successful cytokinesis.
Because anillin predominantly localizes at the cleavage furrow in anaphase, we tested if the A5 mutant impairs localization to the furrow. To do this, HeLa cells were transiently transfected with either GFP-tagged wild-type or A5 anillin, enriched at anaphase and assessed for localization. Endogenous anillin was used as internal control to verify that each cell is evaluated at the time proper for anaphase localization. Indeed, we found that A5 anillin remains diffused throughout the cytoplasm in early and late anaphase, wherein endogenous anillin concentrates at furrow (Fig 3A). This phenotype was scored in two independent ways: a cell population scored by a blinded observer where only cells with strong furrow localization was counted (Fig 3B), and an intensity analysis in which the GFP fluorescence at the equatorial membrane was normalized to the cytoplasmic levels near poles in a single cell (Fig 3C). Both analyses confirmed impaired recruitment of A5 anillin at the furrow. We conclude that the A5 construct, with 11 nonphosphorylatable mutations, is not efficiently recruited to the equatorial midzone on anaphase onset. RhoA is a central regulator of furrow formation [29, 30] and its TCA-fixable cortical pool is dramatically reduced by loss of anillin [10]. To test if A5 was able to stabilize active RhoA, cells were fixed/extracted with TCA [31]. As expected, fixed cortical RhoA is significantly reduced at both early and late anaphase cells expressing A5 anillin (Fig 3D and 3E). These findings suggest that A5 anillin has impaired equatorial recruitment and is unable to fix cortical RhoA at the furrow.
Phosphorylation sites could operate distributively to enhance furrow recruitment of anillin. Alternatively, a single phosphorylation site could be primarily responsible. To test this, we generated single/double non-phosphorylatable mutants for the each of the 11 sites in A5 anillin. Each GFP-tagged mutant was transiently expressed in HeLa cells and analyzed for localization. Most single/double mutants localize properly like their endogenous counterparts (S2 Fig). However, anillin Y634F/S635A behaved like A5, showing a dispersed localization (Fig 4A). To distinguish the relative contributions of Y634 and S635, we tested mutations singly. Both mutants disrupt anillin localization, but S635A had a more marked contribution (Fig 4B–4D). Thus, from 11 sites, we identified S635 as a major residue required for proper anaphase localization of anillin. Moreover, S635 is conserved across metazoans, supporting its role as a critical phosphorylation site for cytokinesis (Fig 4E).
In principle, loss of the hydroxyl group in S635A anillin could disrupt protein function by a mechanism other than preventing phosphorylation. If phosphorylation per se is important, its function could be restored by a phosphomimetic negatively-charged residue at this site. Indeed, anillin S635D concentrated at the equatorial cortex (Fig 4F); this was clearly distinguishable from the one of S635A, as judged by a blind observer assay (Fig 4G) or by quantitative immunofluorescence (Fig 4H), although it did not restore recruitment to the furrow to the extent of wild type. This disrupted localization is specific to anaphase, as anillin S635A localized properly in interphase and metaphase (S3 Fig). Thus, S635 is crucial for anillin localization at the furrow but dispensable for localization at other times. These results support the idea that phosphorylation of S635 is critical for anillin recruitment to equatorial membrane in cytokinesis.
To characterize the cytokinesis defects and membrane recruitment, we performed timelapse videomicroscopy of wild type and mutant anillin constructs after depleting endogenous anillin (Fig 5 and S1–S5 Videos). With depletion of anillin and GFP transfection, we observed marked furrow instability, as expected [10, 13, 19]. This effect was rescued with GFP-wildtype anillin, which was efficiently recruited to membrane and enriched at the furrow by 5–9 minutes after anaphase onset. By contrast, both S635A and double mutant Y634F/S635A were poorly recruited to membrane during mitosis, although some furrow localization is seen. These constructs failed to restore furrow stability and resulted in significant oscillation of furrows. The phosphomimetic, S635D, restores membrane recruitment, enrichment at the cleavage furrow, and stabilization of the mitotic furrow. To assess this in a larger population of cells, we analyzed furrows in timelapse videomicroscopy of HeLa cells expressing mCherry-H2B (S4 Fig). As above, these images reveal transient furrows and then failed cytokinesis to generate a single cell with two nuclei.
To assess frequency of furrow instability with non-phosphorylatable anillin, we developed a fixed-cell assay to identify cells with unstable furrows (Fig 6A and 6B). As expected anillin depletion increased the number of cells with eccentric furrows. Likewise, non-phosphorylatable mutants had increased eccentric furrows compared with wild-type control. Thus, the non-phosphorylatable mutants of anillin are unable to sustain a stable contractile ring, consistent with poor recruitment of these mutants in early anaphase. Together, these data provide high confidence that S635 is important to sustain the mitotic furrow.
We next considered mechanism for the cortex association of phospho-S635 anillin. The AH domain of anillin adjacent to S635 is known to interact with RhoA and Ect2 [10, 32]. Therefore, we reasoned that phopho-S635 anillin might relieve autoinhibition of the AH domain to stabilize its interaction with these interactors. To address this, we performed a pulldown assay from mitotic HeLa extracts with either unphosphorylated (GST WT AH-PH) or the phosphomimetic (GST S635D AH-PH) C-terminal anillin fragment. For RhoA pull down, cells were transfected with constitutively active RhoA (Q63L). However, RhoA and Ect2 were pulled down equally well by the wild-type and phosphomimetic C-terminal anillin fragment (Fig 6C). We conclude that S635 phosphorylation is unlikely to regulate interaction with Ect2 or active RhoA.
A key localization requirement for anillin is phosphatidylinositol phosphate at the plasma membrane, mediated by its C-terminal PH domain [10]. This prompted us to test the effect of phosphorylation at S635 on association with membrane phospholipids. GST-tagged anillin AH-PH variants were incubated with an array of lipids that were immobilized on a support membrane. GST fusions of both wild-type and phosphomimetic AH-PH bound to phosphatidylinositol 4-phosphate PI(4)P, PI(4,5)P2, and PI(3,4,5)P3. However, there was no observed difference in preference for phospholipids (Fig 6D). Thus the specific mechanism by which phosphorylation regulates anillin recruitment and furrow stability appears to be distinct from its roles in binding RhoA, Ect2, and does not control specificity for binding of the C-terminus to specific phospholipids.
To evaluate localization of phosphorylation, we raised a polyclonal antibody against a peptide encompassing phospho-S635, and validated it by several measures. First, dot-blot verified >1000 fold phospho-specificity, as non-phosphorylated peptide was not detected at up to 1250 pmol (Fig 7A). By Western blot of whole cell lysates, pS635 antibody recognized bands from HeLa lysates that matched those detected by anti-anillin antibody and decreased upon anillin depletion (S5 Fig). We were unable to fully validate specificity for pS635 by immunoblot. By contrast, immunofluorescence confirmed reactivity of the phosphospecific antibody at the midbody in late cytokinesis, whereas the non-phosphospecific antibody did not detect an epitope at this site (Fig 7B). Finally, siRNA against anillin/addback revealed that the phosphoepitope is detected with wild-type anillin transfection, but not with S635A (Fig 7C and 7D). We conclude that this antibody specifically detects phosphorylated S635 by immunofluorescence.
Having established the specificity of the pS635 antibody in immunofluorescence, this reagent was used to investigate the dynamics of phosphorylation during cytokinesis. Localized S635 phosphorylation appears in cells after furrowing and persists until late cytokinesis (Fig 7E). Strongly enriched signal is seen at the cortical midzone at late cytokinesis. We performed a more detailed analysis of pS635 signal at different stages of anaphase under different fixation/extraction conditions and confirmed that the signal is only seen in late mitosis (S6 Fig). However, it was puzzling that pS635 is not detected earlier in anaphase when anillin is being recruited to membrane. We considered that the adjacent post-translational modifications could partially interfere with the antibody. Indeed, the antibody fails to detect a doubly phosphorylated peptide at both Y635 and S635 (S7 Fig). We conclude that S635 on anillin is an in vivo phosphorylation site, although our reagent cannot detect early mitotic phosphorylation, possibly due to interference with adjacent post-translational modifications.
Anillin is phosphorylated in mitosis [22], but heretofore, the physiological role of these post-translational modifications has been obscure. We discovered that among 46 phosphorylated residues, S635 is required for efficient anillin recruitment to the furrow and for successful cytokinesis. Several lines of evidence confirm that S635 is important due to phosphorylation at this site rather than to another function of serine hydroxyl. First, it can be detected with a phospho-specific antibody on anillin in late cytokinesis here and previously with mass spectrometry. Second, its localization outside mitosis is unaffected by this mutation. Third, its function is partially restored with a phosphomimetic aspartic acid residue, demonstrating that negative charge at this site is important for its function, but the serine hydroxyl is dispensable. Finally, this residue is phylogenetically conserved in metazoans. Together, these data strongly support a role of phosphorylation at this site to allow anillin to function in cytokinesis.
The kinase responsible for S635 phosphorylation remains obscure. The amino acid sequence surrounding S635 does not match canonical motifs of mitotic kinases. Moreover, in preliminary experiments, we did not observe phosphorylation in the domain containing S635 by Cdk1, Plk1, or Aurora B. However, it is possible that cell contexts or modification of adjacent residues (such as Y634 phosphorylation) could alter kinase specificity. To identify the kinase and understand phospho-regulation of anillin, it will be important to consider a broad host of kinases involved in mitosis, especially those required for cytokinesis, including Rho-associated kinases.
We were unable to identify a specific mechanism by which S635 phosphorylation regulates recruitment of anillin to the equatorial membrane in anaphase. The AH-PH domain is sufficient for membrane recruitment without the region encompassing S635 [10, 16]. However, the fragment encompassing S635 could be an autoinhibitory domain, precluding AH-PH domain binding to membrane until anaphase. Phosphorylation of S635 could relieve this autoinhibitory control of AH-PH to regulate timely membrane association in anaphase.
Mechanistic and structural experiments are needed to determine if this phosphorylation controls affinity of the anillin N-terminus for efficient membrane recruitment. Moreover, it will be important to assess how phosphorylation controls membrane interaction.
One paradoxical finding is that pS635 anillin is observed primarily late in cytokinesis, yet our data suggest phosphorylation at this site is required for early anillin recruitment. One possibility is that phosphorylation of anillin occurs only late in anaphase and controls its roles in abscission [33–35]. This model is consistent with the observation of late cytokinesis failure but appears inconsistent with the impaired recruitment and the furrow instability seen with S635A. Another possible explanation for failure to detect pS635-anillin in early anaphase is that the phospho-antibody has less affinity for pS635-anillin than the total anillin antibody, so the pS635-anillin is only visualized after the ingressing furrow concentrates it in late cytokinesis. A final explanation is that phosphorylation at Y634 masks the pS635 in early cytokinesis, as we observe with phosphopeptide analysis. Thus it is possible that phosphorylation at this site occurs early in mitosis and impairs recognition until late in cytokinesis. In any case, the antibody data provide further confidence that anillin is phosphorylated at this site in human cells.
Although it may seem remarkable that many anillin phosphorylation sites appear dispensable for cytokinesis, our data do not allow for this conclusion. First, our depletion was optimized for detecting/rescuing binucleation and resulted in only ~80% knockdown of endogenous anillin; it is possible that phosphorylation on residual anillin was sufficient to retain some functions of anillin. Second, the transfection assays are heterogeneous, and it is possible that high expression in some cells masked defects of non-phosphorylatable mutants. Third, we evaluated phosphorylation sites by domain-specific mutants. Although many of the domain-specific mutants, A1-4, and A6-7, appeared to function for cytokinesis, it is possible that redundant phosphorylation sites span the boundaries we selected or, perhaps, our assay was not sensitive to subtle functions within cytokinesis. Moreover, we find that mitotic phosphorylation is decreased when Cdk1 is inactivated, possibly removing a negative regulator of anillin function in early mitosis. If so, there are additional layers of regulation as mutations of multiple putative Cdk1 phosphorylation sites did not ultimately preclude successful mitosis and cytokinesis. Additionally, phosphorylation at other sites may be important for differentially regulated functions of anillin, such as meiotic division [33].
Previously known regulatory mechanisms are insufficient to wholly explain the temporal and spatial control of anillin in cytokinesis. For example, astral microtubules [36] and Ran-GTP signals from chromatin [20], can help exclude anillin from polar membrane, although the latter appears to require close apposition of chromatin. Additionally, recruitment of anillin and RhoA to the membrane are mutually dependent [10, 18, 19], suggesting RhoA activation could be partly responsible for timing. Temporal and spatial control of the upstream kinase and anillin phosphorylation can further enhance the specificity of anillin recruitment.
In sum, using a functional screen, we identify an essential phosphorylation at S635, important to reinforce the equatorial zone of anillin. This phosphorylation site provides temporal control of its interaction with the equatorial membrane. This allows anillin to efficiently integrate Rho with its upstream regulators and downstream regulators in a timely fashion to ensure successful cytokinesis. It will be important to identify the kinase responsible for this phosphorylation and to understand how it operates with convergent mechanisms for timely and specific recruitment of anillin for cytokinesis.
Full-length human anillin cDNA isoform 2 (accession number BC070066, Open Biosystems) was cloned into pEGFP-C1 (Clontech). This isoform was used previously to analyze anillin domain functions and encodes a protein with a 37 residue gap between the actin-binding and rho-binding domains, compared with the longest isoform. An RNAi-resistant version was made by polymerase chain reaction (PCR), by engineering of the following changes: nt 798 5’-TGCCTCTTTGAATAAA-3’ 814 to 5’-CGCAAGCTTAAACAAG-3’, creating silent mutations in the cDNA. Various derivatives were made from the RNAi-resistant template for rescue experiments. GST fusions were generated by subcloning of various anillin fragments into the pGEX-6P-1 vector (GE Healthcare).
The phosphodeficient subdomains of anillin (gBlocks: g1-g5) were generated as double-stranded, sequence-verified genomic blocks by Integrated DNA Technologies (IDT). gBlock fragments and PCR amplified anillin fragments were added to Gibson Assembly Master Mix (New England Biolabs) and incubated at 50°C for 1 h. Assembled full-length phosphodeficient mutants of anillin are then cloned into pEGFP-C1.
Cell lines were propagated at 37°C in 5% CO2 in media supplemented with 10% fetal bovine serum and 100 units/ml penicillin-streptomycin. The following media were used: DMEM (HeLa and ACHN), DMEM:F12 (RPE), RPMI (786-O). To enrich cells in anaphase, cells were treated with monastrol for 8 h, and fixed 60 min after release from the monastrol block.
The following siRNA duplexes were used: control (Thermo Scientific siGENOME Non-Targeting siRNA #2 D-001206-14), anillin (40 nM; Thermo Scientific, custom order; CGAUGCCUCUUUGAAUAAA). Lipofectamine 2000 (Invitrogen) was used for siRNA/add back transfection. Cells were analyzed 48 h after transfection. For transient DNA transfection, HeLa cells were transfected using FuGENE HD (Promega) and analyzed 24 h to 48 h post transfection.
GST-tagged anillin fragments were expressed in E. coli (BL21) and expression was induced by the addition of 0.1 mM IPTG at 30°C for 4–5 h. Bacteria were resuspended in PBS containing 250 mM NaCl, 10 mM EGTA, 10 mM EDTA, 0.1% Tween-20, 1 mM DTT, 1 mM phenylmethanesulfonylfluoride (PMSF) and 1 mg/ml lysozyme prior to sonication. Lysates were purified using Glutathione Sepharose 4 Fast Flow (GE Healthcare). For the in vitro kinase assay, a series of GST-tagged anillin fragments (2 μg) were incubated with each kinase in Mg+2-containing kinase buffer with 0.1 mM ATP and 1 μCi [γ-32P] ATP at 30°C for 30 min. The reaction was terminated by addition of SDS sample buffer. Samples were resolved by SDS-PAGE, visualized by Coomassie brilliant blue staining, and finally analyzed using a phosphor imager (Typhoon, GE Healthcare).
HeLa cells were lysed in lysis buffer (50 mM HEPES pH 7.5, 100 mM NaCl, 0.5% NP-40, 10% glycerol, 1 mM DTT, 1 mM PMSF, protease inhibitor cocktail and phosphatase inhibitor cocktail) and incubated with purified GST-tagged anillin AH-PH at 4°C for 4 h. The beads were washed twice with lysis buffer and subjected to SDS-PAGE. Pulled down proteins were detected by immunoblotting using anti-RhoA or anti-Ect2 antibody.
For polyclonal anillin antibodies, GST-tagged anillin amino acids 372–607 were expressed in E. coli, purified by Glutathione Sepharose 4B (GE Healthcare). Tag-cleaved proteins by Prescission protease were then used to immunize rabbits for production of antisera. Raw serum was tested for the specificity of antibodies by Western blot analysis, immunoprecipitation, and immunofluorescence staining (S5 Fig). To make phosphospecific antibodies, phosphopeptides were used to immunize rabbits. Serum was passed through a phosphopeptide affinity column. The eluted antibodies that contain a mixture of phospho- and nonphospho- antibodies were passed through a nonphosphopeptide column (Genemed Synthesis Inc.). The flow through was tested for phosphospecificity.
The following primary antibodies were used: rabbit anti-anillin (1:2000; IF, 1:1000; blot), mouse monoclonal anti-GFP (1:1000; 3E6, Invitrogen), mouse monoclonal anti-cyclin B1 (1:2000; BD Biosciences), mouse monoclonal anti-β actin (1:20,000, Abcam), mouse monoclonal anti-RhoA (1:1000; 26C4, Santa Cruz Biotechnology), rabbit anti-Ect2 (1:1000, Santa Cruz Biotechnology), mouse monoclonal anti-GST (1:2000; B-14, Santa Cruz Biotechnology), mouse monoclonal anti-Flag (1:2000; M2, Sigma-Aldrich), rat monoclonal α-tubulin (1:1000; YL1/2, Millipore).
Reagents used in this study are Lambda protein phosphatase (New England Biolabs), Gibson Assembly Master Mix (New England Biolabs), nocodazole (0.2 μg/ml; EMD Biosciences), monastrol (100 μM; R&D Systems), thymidine (2.5 mM; EMD Biosciences), BI-2536 (200 nM; a gift from P. Jallepalli), ZM-447439 (4 μM; R&D Systems), RO-3306 (10 μM; R&D Systems).
For immunofluorescence (IF), cells were cultured on glass coverslips in 24-well plates and fixed with 4% paraformaldehyde or ice-cold TCA for 15 min. Fixed cells were then blocked for 30 min in 3% bovine serum albumin (BSA) and 0.1% Triton X-100 in PBS (PBSTx + BSA). Primary antibodies were incubated in PBSTx + BSA for 1 h at room temperature and washed three times in PBSTx followed by secondary antibody incubation in PBSTx + BSA for 30 min at room temperature and two washes with PBSTx. Cells were counterstained with DAPI, mounted on glass slides with Prolong Gold antifade medium (Invitrogen), and allowed to cure overnight. Image acquisition was performed on a Nikon Eclipse Ti inverted microscope equipped with CoolSNAP HQ2 charge-coupled device camera (Photometrics).
To visualize GFP-tagged anillin and high resolution imaging of cytokinesis furrow formation, HeLa cells were grown in 12-well plates and transiently transfected with siRNA and anillin cDNA. 24 hr after transfection, cells were transferred to 1.5 coverslip glass-bottom, multi-well plates or 35 mm dishes (MatTek). After an additional 24 hr, cells were imaged on a Leica DMi8 inverted microscope, equipped with a 488 excitation laser, Yokogawa CSU-W1 spinning disk confocal scanner, 63x Plan Apo 1.4 NA oil-immersion objective, and controlled by Metamorph software. Environmental control was maintained using a Tokai Hit stagetop incubator. DIC and GFP images were collected every 30 seconds with a Hammatsu Orca Flash 4.0 CMOS camera. For phase contrast videomicroscopy, stable-expressing mCherry-H2B Hela were cultured and transfected as above prior. Epifluorescence and phase contrast videomicroscopy were captured on the Nikon Eclipse microscope as above with a humidified incubator to maintain cells at 37°C with 5% CO2. Images were captured every 5 minutes with phase contrast and mCherry epifluorescence and mitotic cells were analyzed by visualizing furrows.
Quantitative immunofluorescence was performed to evaluate GFP-Plk1 levels at equatorial membrane versus poles with regions of interest (ROIs) as shown in Figs 3C and 4H. ROIs were chosen internal to the external membrane to minimize the confounding effects of intra-versus extra-cellular levels of GFP-anillin.
Membrane lipid arrays (P-6002, Echelon Biosciences) were incubated with 1 μg/ml GST fusion proteins in PBS containing 0.1% Tween-20 and 3% BSA for 3 h at 25°C. After washing, proteins were detected using anti-GST antibody.
Membranes were imaged with the Odyssey infrared imaging system (LI-COR Biosciences) and quantified using Image Studio Lite v5.2 software (LI-COR Biosciences). To quantify each band a box was drawn around the band to calculate the total pixel intensity within that box. To account for background fluorescence, a second box of equal size was drawn within the same lane and the pixel intensity of that background box was subtracted from the pixel intensity of the box containing the band of interest. Anillin intensities were normalized to β-actin for each sample, then compared to the control knockdown/addback condition to determine the percent anillin knockdown and relative total anillin expression following each transfection. The mean and standard deviation of three independent replicates are reported.
Replicate experiments were performed and analyzed by 2-tailed t-test with the comparisons as shown. There were no corrections for multiple t-tests. For multiple comparisons ANOVA was used as described in legends.
|
10.1371/journal.pgen.1007573 | bric à brac (bab), a central player in the gene regulatory network that mediates thermal plasticity of pigmentation in Drosophila melanogaster | Drosophila body pigmentation has emerged as a major Evo-Devo model. Using two Drosophila melanogaster lines, Dark and Pale, selected from a natural population, we analyse here the interaction between genetic variation and environmental factors to produce this complex trait. Indeed, pigmentation varies with genotype in natural populations and is sensitive to temperature during development. We demonstrate that the bric à brac (bab) genes, that are differentially expressed between the two lines and whose expression levels vary with temperature, participate in the pigmentation difference between the Dark and Pale lines. The two lines differ in a bab regulatory sequence, the dimorphic element (called here bDE). Both bDE alleles are temperature-sensitive, but the activity of the bDE allele from the Dark line is lower than that of the bDE allele from the Pale line. Our results suggest that this difference could partly be due to differential regulation by AbdB. bab has been previously reported to be a repressor of abdominal pigmentation. We show here that one of its targets in this process is the pigmentation gene tan (t), regulated via the tan abdominal enhancer (t_MSE). Furthermore, t expression is strongly modulated by temperature in the two lines. Thus, temperature sensitivity of t expression is at least partly a consequence of bab thermal transcriptional plasticity. We therefore propose that a gene regulatory network integrating both genetic variation and temperature sensitivity modulates female abdominal pigmentation. Interestingly, both bDE and t_MSE were previously shown to have been recurrently involved in abdominal pigmentation evolution in drosophilids. We propose that the environmental sensitivity of these enhancers has turned them into evolutionary hotspots.
| Complex traits such as size or disease susceptibility are typically modulated by both genetic variation and environmental conditions. Model organisms such as fruit flies (Drosophila) are particularly appropriate to analyse the interactions between genetic variation and environmental factors during the development of complex phenotypes. Natural populations carry high genetic variation and can be grown in controlled conditions in the laboratory. Here, we use Drosophila melanogaster female abdominal pigmentation, which is both genetically variable and modulated by the environment (temperature) to dissect this kind of interaction. We show that the pigmentation difference between two inbred fly lines is caused by genetic variation in an enhancer of the bab locus, which encodes two transcription factors controlling abdominal pigmentation. Indeed, this enhancer drives differential expression between the two lines. Interestingly, this enhancer is sensitive to temperature in both lines. We show that the effect of bab on pigmentation is mediated by the pigmentation gene tan (t) that is repressed by bab. Thus, the previously reported temperature-sensitive expression of t is a direct consequence of bab transcriptional plasticity.
| Complex traits such as size or disease susceptibility are typically modulated by both genetic variation and environmental parameters. This has major implications in agronomy, animal husbandry and medicine. Furthermore, as the phenotype (and not the genotype) is the target of natural selection, both genetic and environmental factors are fundamental to understand evolution. Indeed, Waddington showed that a phenotype initially induced by environmental conditions can be selected and become independent of the environment [1,2]. He proposed that genetic variation present in the population was the base of this process that he termed genetic assimilation [1–3]. The importance of standing genetic variation was demonstrated by Bateman who repeated some of Waddington’s experiments and showed that genetic assimilation worked with outbred stocks but not with isogenic stocks [4]. However, in a recent study, it was shown that de novo mutations induced by stressful environments—such as heat-shock—can, in some cases, contribute to genetic assimilation [5]. Based on Waddington’s genetic assimilation, West-Eberhard proposed that divergent lineages could be produced through genetic assimilation of alternative morphs present in a phenotypically plastic ancestral species (“the flexible stem hypothesis”) [6]. A few studies suggest, indeed, that this mode of evolution, also called “plasticity-first evolution“, may not be uncommon in nature. For example, the tadpoles of many spadefoot toad species adjust their development time to the duration of the pond in which they live. Phylogenetic analyses show that this plasticity is ancestral. In contrast, the species Scaphiopus couchii, developing in ephemeral ponds, has evolved a derived and canalized short developmental time independent of the duration of the pond in which it develops (reviewed in [7]). Similarly, some generalist cichlid species show morphological plasticity of their pharyngeal jawbones in response to diet hardness. The plasticity observed in generalist species is ancestral. In contrast, other species of cichlids, which have become specialised on hard or soft diet, show only one type of morphology [8]. It is therefore important to investigate such cases at the genetic and molecular levels to understand the mechanisms of the "flexible stem hypothesis"/"plasticity-first evolution". Model organisms such as Drosophila are particularly appropriate to dissect the interactions between genetic and environmental factors. Indeed, natural populations carry high genetic variation. Furthermore, they can be easily grown in controlled conditions in the laboratory and many genetic tools are available.
In this study, we focus on abdominal pigmentation of Drosophila melanogaster (D. melanogaster) females, a trait highly variable in natural populations and modulated by environmental factors such as nutrition and temperature [9–11]. Female abdominal pigmentation is darker at low temperature, in particular in the most posterior segments A5, A6 and A7 [12]. Abdominal pigmentation is therefore an example of phenotypic plasticity defined as “the property of a given genotype to produce different phenotypes in response to distinct environmental conditions” [13]. Lastly, as abdominal pigmentation in drosophilids is widely used as a model of intra- and inter-specific evolution [14,15], it is particularly appropriate to analyse the contribution of genetic and environmental factors to phenotypic variation and evolution.
Genome-wide association studies have identified several loci linked to female abdominal pigmentation variation in natural populations of D. melanogaster [9,11,16]. Interestingly, although the same loci were identified in different populations, their prevalence depends on the population [9,16]. In European populations, the single nucleotide polymorphisms (SNPs) that are the most strongly associated with pigmentation variation are located in the t_MSE, an abdominal enhancer of the tan (t) gene encoding an enzyme involved in melanin production [9,17]. Functional analysis of these SNPs in transgenic lines confirmed that they affect female abdominal pigmentation [18]. By contrast, in South African populations, the most significant SNPs are located in an intron of bric à brac 1 (bab1) [16]. bab1 and its tandem duplicated paralogue bab2 (named collectively bab) encode transcription factors involved in repression of abdominal pigmentation [19,20]. Independent genome-wide association studies performed on a North American population also identified SNPs located in the t_MSE and the bab1 intron as the most significantly associated with female abdominal pigmentation variation [11]. Furthermore, in another study, genetic variation was detected in a bab1 and bab2 cis-regulatory sequence, the bab dimorphic element (named thereafter bDE), which is located in the first intron of bab1 and controls sex-specific expression of bab1 and bab2 in the posterior abdominal epidermis [21]. This genetic variation was associated with changes in the bDE activity as well as in female abdominal pigmentation [10].
Interestingly, the genes involved in female abdominal pigmentation variation are also involved in female abdominal pigmentation thermal plasticity. We previously showed that female abdominal pigmentation plasticity was caused by temperature sensitivity of a genetic network including the bab locus [22]. More recently, we showed that temperature modulates the expression of t and yellow (y), another pigmentation enzyme gene, thus contributing to female abdominal pigmentation plasticity [23,24]. The effect of temperature on t expression is mediated at least partly by the t_MSE [23]. Thus, t is an essential effector of female abdominal pigmentation plasticity. However, it is not excluded that temperature affects the activity or the expression of upstream regulator(s) of t and consequently its expression.
As the effects of genetic variation and temperature on female abdominal pigmentation were mainly studied independently, our aim was to investigate how they interact in the production of this phenotype. In this study, using two D. melanogaster lines differing in abdominal pigmentation (named thereafter Pale and Dark), we first show that genetic variation at the bab locus affects bab expression and female abdominal pigmentation. We demonstrate that the difference in bab expression is caused at least partly by genetic variation in the bDE. Indeed, a deletion that removes two binding sites for Abdominal B (Abd-B), a direct activator of bab, is present in the bDE of the Dark line. We show that this deletion impacts the activation of the bDE by Abd-B. Furthermore, the expression of bab is modulated by temperature in both lines, and this modulation results at least partly from temperature sensitivity of bDE. Lastly, t, whose expression is modulated by temperature, is also differentially expressed between the two lines and is repressed by bab. Hence, temperature modulation of t expression is at least partly due to bab transcriptional plasticity.
We established the isogenic Pale and Dark lines at 25°C (see the material and methods section) from a natural population sampled in Canada [25] that was polymorphic for female abdominal pigmentation. To compare their abdominal pigmentation and its thermal plasticity, these lines were grown at 18°C, 25°C and 29°C. Abdominal pigmentation of A4, A5, A6 and A7 segments was quantified and analysed in females (Fig 1, S1 Dataset, Table 1 and S1 Fig). Dark females were darker than Pale females at the three temperatures. Although the effect of the genotype ("G" effect) was significant for the four segments, it was particularly pronounced in segments A6 and A7, as revealed by Eta squared (h2) values (A6 and A7: p<0.001, h2 = 0.59; A5: p<0.001, h2 = 0.17; A4: p<0.05, h2 = 0.04) (Table 1). The effect of temperature ("T" effect) was strong for both lines in all four segments (p<0.001, 0.32<h2<0.53). Lastly, the interaction between the genotype and the temperature ("GxT" effect) was significant for segments A6 and A7 (p<0.001). In conclusion, the Dark and Pale lines differed in pigmentation. They were both plastic in response to temperature but their plasticities differed only in segments A6 and A7.
We first aimed at identifying the respective contribution of each chromosome to the difference of abdominal pigmentation between the two lines. We therefore constructed new lines carrying the eight different combinations of the X, the second and the third chromosomes from the Pale and the Dark lines. We compared female abdominal pigmentation of the eight different lines grown at 25°C (Fig 2, S2 Dataset, and S2 Fig). At a glance, an effect of the third chromosome on pigmentation was noticeable for segments A4, A5, A6 and A7 (Fig 2A). Quantification and calculation of Eta squared values (h2) confirmed this effect of the third chromosome, which was moderate in A4 (p = 0.001, h2 = 0.14), strong in A5 (p<0.001, h2 = 0.29) and very strong in A6 and A7 (p<0.001 for both segments, h2 = 0.77 and 0.83, respectively) (Fig 2B and S2 Fig). In addition, the second chromosome also had a significant but much weaker effect in all segments but A4. Weak but significant effects were also observed in A7 for the X chromosome and for the interaction between chromosome II and chromosome III (Fig 2B and S2 Fig). In conclusion, these results show that one or several loci located on the third chromosome should carry genetic variation causing most of the abdominal pigmentation difference between the Dark and Pale lines. A very good candidate on this chromosome was the bab locus, previously identified as a major contributor for variation of female abdominal pigmentation in natural populations of D. melanogaster [10,11,16,26].
To assess a potential effect of genetic variation at the bab locus, we performed an association study on the F2 progeny from a cross between Dark females and Pale males. We focused on the bab regulatory element called bab dimorphic element (bDE [21]), located in bab1 large intron (Fig 3A), as this enhancer was previously shown to carry high natural genetic variation with effect on female abdominal pigmentation [10]. In comparison with the CantonS reference haplotype (bDEC) [21], the Pale line haplotype (bDEP) differed by only one nucleotide (Fig 3B, yellow). In contrast, the Dark line haplotype (bDED) presented a 56 bp deletion (Fig 3B, red), which had not been reported in previously described bab natural alleles. In order to investigate how frequent this deletion is in natural populations of D. melanogaster, we first analysed this region of bDE in the genomic sequences of 30 world-wide populations [27,28] (966 sequences in total). We never identified the 56 bp deletion in these populations. However, as this dataset of world-wide populations is focused mainly on SNPs and has not been evaluated for indels [27], it might not be optimal to identify the 56 bp deletion. Therefore, we analyzed in detail the region of the 56 bp deletion in 205 lines also originating from North America (DGRP lines from Raleigh) [29,30] using the available web-interface (http://dgrp2.gnets.ncsu.edu/). Not only SNPs but also small and large indels have been analyzed in these lines [30]. The 56 bp deletion found in the Dark line bDE was not present in any of these lines. Indeed, a SNP has been identified within the 56 bp (nucleotide 1,084,899, which corresponds to the 18th position in the 56 bp sequence) and characterized in each of the 205 DGRP lines. Thus, the 56 bp deletion is likely to be a recent and rare allele, perhaps deleterious, present at low frequency at least in some Canadian populations. Although this deletion may not be relevant for adaptation, it represents an experimentally tractable system for studying the genetic and environmental interactions that affect a complex phenotype. We used this 56 bp deletion to genotype by PCR 40 F2 females randomly collected (S3 Fig). For each of them, DNA was extracted from the head and thorax whereas the abdomen was kept for pigmentation quantification. Scatter plot of pigmentation in A6 and A7 segments showed a clear segregation of the different genotypes, with bDED/bDED females being darker than bDEP/bDEP females and bDED/bDEP being intermediary between the two former genotypes (Fig 3C and S3 Dataset). Female abdominal pigmentation was strongly associated with the genotype at the bab locus in A6 (p<0.001, h2 = 0,61) and A7 (p<0.001, h2 = 0.74) but not in A5 (p = 0,49) (Fig 3D and S4 Fig). In both A6 and A7, pigmentation differs between all three genotypes (Tukey HSD test, p<0.01 for all pairwise comparisons).
In conclusion, these results strongly suggested that abdominal pigmentation variation in A6 and A7 between the Pale and the Dark lines was mainly linked to genetic variation at the bab locus.
The bDE is directly activated by the Doublesex female-specific isoform (DsxF) and the Hox protein Abdominal B (Abd-B) in the abdominal epidermis of females [27]. Interestingly, the 56 bp deletion present in the bDED allele of the Dark line, removes two Abd-B binding sites (Fig 3B). These sites are the sites Abd-B3 and Abd-B4 that were shown to bind Abd-B in vitro and to contribute, together with twelve other Abd-B binding sites, to bDE activity in vivo [21]. The bDEP allele of the Pale line also differs from bDEC and bDED by a single nucleotide substitution (Fig 3B, yellow). This SNP was previously reported to reduce the activity of the enhancer [10]. To test whether the 56 bp deletion of the bDED allele could also affect the activity of the enhancer, we constructed transgenic lines in which nuclear enhanced green fluorescent protein (nEGFP) was under the control of bDED or bDEP (lines bDED-nEGFP and bDEP-nEGFP). In order to compare the activities of these two variants, the transgenes were inserted at the same genomic location using phiC31-based transgenesis [31], thus avoiding position effects. Then, they were introgressed for six generations in the same genetic background. The resulting lines, homozygous for the transgenes, were grown at 18°C and at 29°C, in order to test the potential effect of temperature on bDE activity. bDE activity was previously shown to peak in the abdominal epidermis from female pupae [21]. We therefore quantified nEGFP at this stage (Fig 4, S4 Dataset and S5 Fig). Both bDED and bDEP drove nEGFP in A6 and A7 but not in A5 (Fig 4A), as already described for bDEC [21]. In A6, the two bDE alleles presented a different activity at 29°C (t-test, p<0.01) but not at 18°C, which led to a non-significant effect of the genotype ("G" effect) (Figs 4B and S5). By contrast, in A7, the effect of the genotype on nEGFP expression was significant (Figs 4C and S5; "G" effect, p<0.001). Interestingly, for the two lines and in both A6 and A7, nEGFP was significantly more expressed at 29°C than at 18°C ("T" effect, p<0.001). The genotype by temperature interaction was marginally significant in both segments ("GxT" effect, p = 0.094 and p = 0.073, respectively) (Figs 4B, 4C and S5).
In conclusion, bDED was less active than bDEP. Moreover, temperature modulated the activity of both alleles, which were less active at 18°C than at 29°C. As bab represses melanin production [20], the lower activity of bDED and the higher activity of both alleles at high temperature correlates with pigmentation intensity. Since the 56 bp deletion of bDED removes two binding sites for Abd-B, which activates this enhancer, loss of these sites might participate in the reduction of activity of bDED as compared to bDEP.
To test the impact of the 56 bp deletion in bDED on its activation by Abd-B, we introduced chromosomes with a deletion or a duplication of Abd-B in the bDED-nEGFP and bDEP-nEGFP transgenic lines. We thus obtained flies heterozygous for bDED-nEGFP or bDEP-nEGFP, and expressing one dose, two doses or three doses of Abd-B. Flies were grown at 18°C or 29°C and nEGFP intensity in A6 and A7 was quantified. The different conditions of genotype, temperature and Abd-B dose were then compared (Fig 5, S5 Dataset, S6 and S7 Figs). As already observed with homozygous nEGFP transgenes (Fig 4), the genotype effect (“G”) as well as the temperature effect (“T”) were significant for A6 and A7 (“G” for A6: p<0.01, h2 = 0.008; “G” for A7: p<0.001, h2 = 0.147; “T” for A6: p<0.001, h2 = 0.235; “T” for A7: p<0.001, h2 = 0.394). A significant effect of the genotype by temperature interaction was also observed for A6 ("GxT": p<0.001, h2 = 0.021). These results confirmed that the activity of bDE depended on the allele and on the temperature, and that the effect of temperature was modulated by the genotype.
The effect of Abd-B dose was strong and significant for A6 and A7 (“D” for A6: p<0.001, h2 = 0.540; “D” for A7: p<0.001, h2 = 0.193). This result was expected as Abd-B was shown to be a direct activator of bDE [21] and as bDED and bDEP share twelve Abd-B sites. Very interestingly, a significant effect of the genotype by Abd-B dose interaction was observed ("GxD" for A6: p<0.001, h2 = 0.021; for A7: p<0.01, h2 = 0.024), demonstrating that the 56 bp deletion in bDED had a direct impact on the activation of this enhancer by Abd-B. This effect could be the consequence of the removal of the two Abd-B sites, the changing of spacing between remaining Abd-B sites, or the removal of other uncharacterized regulatory sites.
Furthermore, there was a significant effect of the Abd-B dose to temperature interaction, independently of the genotype and thus the deletion ("DxT" for A6: p<0.001, h2 = 0.038; for A7: p<0.001, h2 = 0.034,). This means that the effect of Abd-B on bDE activity, for both bDE alleles, was modulated by temperature. However, we did not detect any variation of Abd-B expression in the female abdominal epidermis between 18°C and 29°C, indicating that this modulation was not a consequence of Abd-B differential expression (S8 Fig and S6 Dataset). The effect of temperature on Abd-B thus might be post-transcriptional (activity of the protein itself, binding to the enhancer, interaction with co-factors …).
Lastly, a weak but significant interaction between genotype, temperature and Abd-B dose was observed, ("GxTxD" effect for A6: p<0.05, h2 = 0.010; for A7: p<0.05, h2 = 0.012), indicating that the difference in thermal plasticity between bDED and bDEP was partly due to Abd-B.
As the effect of Abd-B on bDE activity was modulated by temperature independently of the bDE genotype (Fig 5, "DxT" effect), we wondered whether we could detect an effect of temperature and of Abd-B dose on bab1 and bab2 regulation. We thus quantified by RT-qPCR bab1 and bab2 expression in the posterior epidermis of female pupae grown at 18°C or 29°C and expressing one, two or three doses of Abd-B (Fig 6, S7 Dataset and S9 Fig). Unfortunately, we could not interpret the data for three doses of AbdB because the reference genes we used were themselves deregulated. A weak but significant effect of temperature was observed for bab1 and bab2 (“T” effect, bab1: p<0.05, h2 = 0.072; bab2: p<0.05, h2 = 0.23). Thus, temperature modulation of bDE activity may impact bab1 and bab2 expression.
A very strong effect of Abd-B dose was observed for both genes (“D” effect, bab1: p<0.001, h2 = 0,792; bab2: p<0.01, h2 = 0.480). This was expected as Abd-B directly activates bDE [21]. Lastly, a significant interaction between Abd-B dose and temperature was observed for bab1 (“DxT” effect, p<0.05, h2 = 0.070). Thus, temperature modulates the effect of Abd-B on bab1 expression, which suggests that the effect of temperature on bDE activation by Abd-B has functional consequences.
In Fig 4, we showed, using reporter constructs in transgenic flies, that bDED is less active than bDEP but that both alleles are temperature-sensitive. To test whether these effects could have an impact on bab1 and bab2 expression, we quantified by RT-qPCR bab1 and bab2 mRNA in the posterior abdominal epidermis of female pupae and young adults from the Dark and Pale lines (Fig 7, S8 Dataset and S10 Fig). bab1 was less expressed in the Dark line than in the Pale line both in pupae and in adults at the two temperatures ("G" effect; pupae: p<0.01, between 2.6 and 3 times less depending on the temperature; adults: p = 0.055, between 1.3 and 1.9 times less depending on the temperature). Expression of bab2 was also lower in the Dark line but significantly only in adults ("G" effect, p<0.01, between 1.3 and 1.4 times less depending on the temperature). In addition, expression of bab1 and bab2 was modulated by temperature in adults, as both genes were more expressed at 29°C than at 18°C ("T" effect, p<0.05, between 1.2 and 1.9 times more at 29°C than at 18°C depending on the line and the gene). Thus, the lower expression of bab paralogues in the Dark line as compared to the Pale line, and their higher expression at high temperature, correlated with bDE activity and pigmentation intensity.
The two paralogues bab1 and bab2 encode transcription factors with BTB/POZ and Psq domains, which are involved in repression of abdominal pigmentation [19,20]. Among pigmentation genes, it is already known that bab represses y and Dopa-decarboxylase (Ddc) [22,32,33]. Based on previous studies showing the essential role of t in female abdominal pigmentation intensity and plasticity [9,18,23], we wondered whether t was a target of bab in abdominal epidermis. Indeed, this was suggested in several studies, but never demonstrated [10,14,15]. To test this hypothesis, we down-regulated bab2 and/or bab1 in the abdominal epidermis using UAS-RNAi transgenes combined with the drivers pannier-Gal4 (pnr-Gal4) [34] or yellow-Gal4 (y-Gal4) [35]. As expected, bab down-regulation increased abdominal pigmentation (Figs 8A and S11A). Furthermore, t expression, revealed by in situ hybridization, increased when bab1 was down-regulated (Fig 8B). To confirm this result, bab was down-regulated in two reporter lines inducing nEGFP expression under the control of t regulatory sequences: the t_MSE-nEGFP line, that contained only the t abdominal enhancer t_MSE [36], and the 5't-nEGFP line, that contained a genomic fragment about 4 kb long located directly upstream the t transcription start site and including the t_MSE. In these two lines, bab down-regulation induced an increase of nEGFP expression (Figs 8C and S11B).
Taken together, these results demonstrated that t expression was repressed by bab, and that this repression was mediated, at least partly, by the t abdominal enhancer t_MSE.
As shown above, bab expression differs between the Dark and Pale lines and is modulated by temperature. In addition, bab represses t expression. In order to test whether bab differential expression in the Pale and Dark lines impacts t expression, we performed in situ hybridization experiments and RT-qPCR quantification in abdominal epidermes of young females of the two lines raised at 18°C or 29°C (Fig 9, S9 Dataset and S12 Fig). As we showed previously in another D. melanogaster line (w1118 line, [23]), t expression was strongly modulated by temperature in the two lines. Indeed, t was 20 times more expressed at 18°C than at 29°C in the Pale line, and 7.7 times in the Dark line ("T" effect: p<0.001, h2 = 0.87). There was also a significant effect of the genotype, although less strong. Indeed, at 18°C, t was 1.5 times more expressed in the Dark line than in the Pale line, and 3.8 times more at 29°C ("G" effect, p<0.01, h2 = 0.068). Lastly, interaction between the genotype and the temperature was marginally significant, indicating that their effects on t expression were mainly additive ("GxT" effect: p = 0.08, h2 = 0.021).
In conclusion, t was differentially expressed between the Pale and the Dark lines and modulated by temperature in both lines. As t was repressed by bab, this differential expression may result at least partly from the differential expression of bab between the two lines and the temperature sensitive expression of bab in both lines.
Here, we use two Drosophila melanogaster lines, Dark and Pale, established from a natural population, to study the regulatory mechanisms of abdominal pigmentation intensity and plasticity. We demonstrate that bab expression differs between the two lines. This is partly due to the existence of a deletion in the bab dimorphic element (bDE) that modulates the action of Abd-B on bDE activity. bab in turn represses t through the t_MSE. The expression of t differs between the Dark and the Pale lines, and this effect is mainly caused by variation in trans at the bab locus. Indeed, contribution of the X chromosome carrying t is non-significant or extremely weak depending on the segments. Absence of functional variation in cis at the t locus contrasts with previous results on European populations, in which SNPs in t_MSE accounted for most variation in female abdominal pigmentation [9]. Indeed, sequencing of t_MSE of the Dark and Pale lines revealed that they have exactly the same sequence (see the Material and Methods section).
Our results allow also a better understanding of the effect of temperature on female abdominal pigmentation (Fig 10). Abd-B lies at the top of a gene network and plays an essential role in plasticity. Indeed, plasticity of pigmentation increases along the antero-posterior axis in parallel with Abd-B expression [12,37]. Furthermore, ectopic expression of Abd-B in the thorax is sufficient to generate a sex-specific highly plastic pigmentation pattern [22]. Abd-B has opposite effects on pigmentation depending on temperature, as it induces strong melanin production at 18°C, whereas repressing all cuticular pigments at 29°C [22]. Moreover, Abd-B and the female-specific isoform of Doublesex (DsxF) activate bab through binding to bDE [21]. We show here that the impact of Abd-B on bDE activity varies with temperature, increasing expression of bab at high temperature. As Abd-B expression does not vary with temperature, this is likely to be due to post-transcriptional mechanisms. In addition, a part of the modulation by temperature of bDE activity and bab genes expression occurs independently of Abd-B (represented by the “T” effect for bDE activity or bab1 and bab2 expression). The causal mechanism is for the moment unknown. In a previous study, we showed that bab belongs to a chromatin regulator network that mediates the effect of temperature on female abdominal pigmentation [22]. Our new results show that transcriptional modulation of bab by temperature through the bDE is important for temperature-sensitivity of the network. Furthermore, as bab represses t, temperature-sensitive expression of t in female posterior abdominal epidermis, previously reported [23] and confirmed in this study, is, at least partly, a consequence of bab temperature-sensitive expression. Interestingly, expression of other pigmentation enzymes, such as Ddc or y, is also repressed by bab and sensitive to temperature [22–24,33]. This is probably caused, at least partly, by the temperature-sensitive expression of bab. Furthermore, the global repressive role of Abd-B on pigmentation enzymes at high temperature goes at least partly through bab. y was recently shown to be a direct target of bab [33]. Whether t and Ddc are direct targets of bab remains unknown. Lastly, the mechanism underlying the activator role of Abd-B on melanin production at low temperature is unknown but may rely on the activation of t and/or other pigmentation enzyme genes.
Classically, two types of genetic effects are postulated to be involved in phenotypic plasticity [38]. “Allelic sensitivity” applies to a gene whose expression or product activity depends on the environment. “Gene regulation” applies to a regulatory gene that turns on or off its targets depending on the environment. Our study illustrates how these effects may blur when considering regulatory genes and their targets. The temperature sensitivity of bab expression could be classified as “allelic sensitivity”. However, because bab encodes a transcriptional repressor, it corresponds also to a “gene regulation” effect when considering the effect on its targets. Furthermore, bab thermal plasticity leads to temperature-sensitive expression of t, another case of “allelic sensitivity”.
A few other gene networks mediating phenotypic plasticity have been described in other species [39,40]. Similarly to the network we describe, these networks are composed of regulatory genes (involved in hormonal pathways, transcription factors) responding to the environment and structural genes that are their downstream targets. For example, pharyngeal jaw plasticity in cichlids is mediated by a complex gene network in which the transcription factor AP1 plays a major role as it is sensitive to the mechanical strain exerted by the food [39]. AP1 regulates another transcription factor Runx2B, which activates the structural gene osx, a key osteoblast regulator [39]. Such a gene regulatory network perspective is needed to understand how phenotypic plasticity is mediated and how it could be involved in evolution by genetic assimilation [41].
Our results corroborate other studies on fly pigmentation variation and evolution which have identified changes in cis-regulatory sequences involved in evolution of body or wing pigmentation (reviewed in [14,15]). Indeed, genetic variation in a modular enhancer such as bDE affects only one trait (abdominal pigmentation) leaving unaffected other traits controlled by bab via distinct enhancers (legs and ovaries) [19]. Similarly, in vertebrates, it has been shown that morphological variation within or between species very often relies on genetic changes in modular enhancers and not in coding sequences. In several examples, loss of particular structures are associated with mutations reducing the activity of an enhancer: parallel pelvic plate reduction in sticklebacks is caused by recurrent deletions of an enhancer of Pitx1 [42]; limb loss in snakes is caused by nucleotide changes in an enhancer of Sonic hedgehog [43,44]. Conversely, mutations increasing the activity of an enhancer are involved in the enlargement of particular structures. For example, in bats, a limb specific enhancer of Prx1 that has a higher activity than its mouse homologue induces significantly longer forelimbs [45]. Some mutations in cis-regulatory sequences can also modify the timing of gene expression: persistent expression of human lactase expression in adult small intestine is linked to several mutations in the regulatory sequences of the lactase gene that have occurred independently in particular populations, whereas the ancestral allele is not expressed after childhood (reviewed in [46]). These variants correspond to single nucleotide polymorphisms (SNPs) and several of these SNPs were shown to increase the binding of the transcription factor Oct-1 [47,48].
The “flexible stem” hypothesis proposes that phenotypically divergent lineages could result from genetic assimilation of alternative morphs produced by an ancestral plastic species [6]. This is observed in a few species, and extends to the genetic mechanisms generating morphological diversity [49,50]. Indeed, in a few cases, the same genes involved in the development of a particular trait show transcriptional plasticity in a plastic species and divergence of expression within or between species because of differences in their regulatory sequences. Regarding Drosophila pigmentation, this is the case for t, through t_MSE. This enhancer carries genetic variation involved in pigmentation divergence within or between species in the Drosophila melanogaster subgroup [9,11,18,36,51], and its activity is modulated by temperature [23]. We show here that the same applies to bab through the bDE which is also involved in intra- and inter-specific evolution [10]. Similarly in plants, the Reduced Complexity locus (RCO) in some species of Brassicaceae is involved in temperature modulation of leaf dissection as well as in difference of leaf dissection between species [52]. These examples suggest that the environmental sensitivity of particular genes turns them into evolutionary hotspots by facilitating the selection of the functional genetic variation they carry [50]. Indeed a given allele will produce different phenotypes in different environments. It will facilitate the selection of this allele as environmental conditions vary spatially and temporally, which increases the probability for this allele to generate a beneficial phenotype.
The Canadian Drosophila melanogaster population from which the Dark and the Pale lines were established was kindly provided by Sam Yeaman [25]. We established each line by selecting 5–10 mated females grown at 25°C with dark or pale abdominal pigmentation. The phenotypes were fixed in 5 generations by selecting females with similar phenotype in the progeny. Each of the two lines was isogenized by brother-sister crosses for 10 generations. We used standard balancer chromosomes to construct the lines carrying the different combinations of chromosomes from the Dark and Pale lines. The w1118 line was the same as used in our previous studies [23]. The t_MSE-nEGFP transgenic line was previously described [23,36]. The UAS-bab1-RNAi (KK106110) and UAS-bab2-RNAi (GD49042) lines were from the Vienna Drosophila Resource Center [35]. The pnr-Gal4 (BL-3039) [34] and y-Gal4 (BL-44267) were from the Bloomington Stock Center. The line Df(3R)C4,p*/Dp(3;3)P5, Sb1 (BL-3071) carrying an Abd-B deletion and an Abd-B duplication was used to manipulate the dose of Abd-B.
The 873 bp fragments containing the t_MSE of the lines Pale (GenBank accession number: MG755262) and Dark (GenBank accession number: MG755261) were amplified by PCR using the following primers:
The 1.5 kb fragments containing the bab dimorphic element (bDE) of the lines Pale (GenBank accession number: MG755259) and Dark (GenBank accession number: MG755258) were amplified by PCR using the following primers:
The 3.9 kb fragment containing the 5' t regulatory sequence was amplified from the w1118 line (GenBank accession number: MG755260) using the following primers:
The PCR products were cloned by topocloning in pENTR (InVitrogen) according to the manufacturer's instructions and sequenced (GATC Biotech). The bDE from the Pale and Dark lines were aligned with the sequence of Canton S (NCBI, EU835207.1, [21]) using Clustal Omega (ebi.ac.uk).
The bDE and the 5' t regulatory sequence were cloned by LR recombination (Gateway cloning technology) into a derivative of PH-Stinger [53] kindly provided by Dr Nicolas Gompel. The PH-Stinger vector was modified by insertion of a Gateway cloning cassette upstream of the nuclear enhanced green fluorescent protein (nEGFP). In addition, an attB site was inserted to allow genomic integration of transgenes using the PhiC31 integrase system [31]. After sequencing, plasmids were injected into y[1] M{vas-int.Dm}ZH-2A w[*]; M{3xP3-RFP.attP'}ZH-51C embryos (Bloomington Stock Center, BL-24482, insertion of the transgene at position 51C, BestGene Inc.).
Genotyping of the bDE alleles was performed by PCR amplification using the following primers framing the 56 bp deletion identified in the Dark line:
PCR amplification was performed on genomic DNA extracted from head and thorax of single flies. The abdomens were stored in ethanol, then mounted to allow pigmentation quantification. The 475 bp (Dark allele) and 531 bp (Pale allele) PCR products were separated by electrophoresis on a 1% agarose gel.
t in situ hybridization was performed as previously reported [23].
RNA was extracted as previously described [23] from pools of 50 dissected pupae or adult female posterior abdominal epidermes (A5, A6 and A7). Morphological markers were used to collect pupae at a similar developmental stage. Three RNA replicates were analysed per condition for all experiments. After treatment of RNA with Turbo DNase (Ambion), cDNA were synthesized with the SuperScriptII Reverse Transcriptase kit (Invitrogen) using random primers. RT-qPCR experiments were performed in a CFX96 system using SsoFast EvaGreen SuperMix (Biorad). Expression was quantified following the Pfaffl method [54] using the geometric mean of two reference genes for normalization [55]. Reference genes were chosen with an expression level similar to the one of the tested gene: eIF2 and Spt6 for bab1, bab2 and Abd-B quantification, Act5C and RP49 for t quantification. For Abd-B, the primers were specific for the M isoform as the R isoform was not expressed in the pupal abdominal epidermis. Primers for t, Act5C and RP49 were already described [23]. bab1, bab2, Abd-B, eIF2 and Spt6 primers were the followings:
For pigmentation analyses, adult females between 3 and 5 days old were stored in ethanol 70% during ten days. Abdominal cuticles were cut just beyond the dorsal midline and dehydrated in ethanol 100% during 5 minutes. After dehydration, cuticles were mounted in Euparal (Roth).
For nEGFP analyses, pupal and adult abdomens were dissected in PBS, fixed 20 minutes in 3.7% paraformaldehyde in PBS, washed twice 10 minutes in PBS and mounted in Mowiol. As developmental time is sensitive to temperature, morphological markers (wing colour, location of meconium) were used to compare pupae grown at 18°C and 29°C at a similar stage of development [56].
Abdominal cuticles of adult females were imaged and quantified as previously described [23].
GFP intensity was quantified in 10 individuals for each condition. Epidermes of females expressing nEGFP under the control of t_MSE were imaged using a macro-apotome (Zeiss) in order to image the whole abdomen (Fig 7). Epidermes of females expressing nEGFP under the control of the 5' t regulatory sequence (S11 Fig) and pupae expressing nEGFP under the control of the bDE (Figs 4 and S6) were imaged using a micro-apotome (Zeiss). nEGFP intensity of homozygous bDE-nEGFP pupae was measured in hemi-segment A6 and A7 using the ZEN software (Zeiss) to generate maximum intensity projections of 20 z-stacks (Fig 4). For bDE-nEGFP/+ pupae, that present a weak nEGFP signal as compared to cuticle autofluorescence, a macro was developed in ImageJ in order to extract and measure nEGFP intensity of nuclei and generate maximum intensity projections of 20 z-stacks (S6 Fig).
One-way ANOVA and Tukey HSD tests for F2 analyses were performed using the VassarStats website (vassarstats.net). Two-way ANOVA were performed with an Excel sheet from Anastats (Anastats; http://anastats.fr). Three-way ANOVA were made using the OpenStat software (W.G. Miller, http://statprogramsplus.com/OpenStatMain.htm). We checked variance homogeneity with a Levene test and normality of residuals with a Shapiro-Wilk test (http://anastats.fr). For one-way and two-way ANOVA, when variances were not homogeneous, we performed a non-parametric ANOVA (Scheirer-Ray-Hare test) using the OpenStat software. For three-way ANOVA, when variances were not homogeneous, we transformed the data with a BoxCox transformation using R. The Eta squared of the various factors (h2) were calculated as SSfactor/SStotal (SS: sum of squares).
|
10.1371/journal.ppat.1005537 | Targeted Isolation of Antibodies Directed against Major Sites of SIV Env Vulnerability | The simian immunodeficiency virus (SIV) challenge model of lentiviral infection is often used as a model to human immunodeficiency virus type 1 (HIV-1) for studying vaccine mediated and immune correlates of protection. However, knowledge of the structure of the SIV envelope (Env) glycoprotein is limited, as is knowledge of binding specificity, function and potential efficacy of SIV antibody responses. In this study we describe the use of a competitive probe binding sort strategy as well as scaffolded probes for targeted isolation of SIV Env-specific monoclonal antibodies (mAbs). We isolated nearly 70 SIV-specific mAbs directed against major sites of SIV Env vulnerability analogous to broadly neutralizing antibody (bnAb) targets of HIV-1, namely, the CD4 binding site (CD4bs), CD4-induced (CD4i)-site, peptide epitopes in variable loops 1, 2 and 3 (V1, V2, V3) and potentially glycan targets of SIV Env. The range of SIV mAbs isolated includes those exhibiting varying degrees of neutralization breadth and potency as well as others that demonstrated binding but not neutralization. Several SIV mAbs displayed broad and potent neutralization of a diverse panel of 20 SIV viral isolates with some also neutralizing HIV-27312A. This extensive panel of SIV mAbs will facilitate more effective use of the SIV non-human primate (NHP) model for understanding the variables in development of a HIV vaccine or immunotherapy.
| An antibody-based approach targeting human immunodeficiency virus (HIV) envelope (Env) protein may eventually prove to be effective in treating or preventing HIV infection. However, before any candidate HIV treatment or vaccine can be tested in humans, it must first be evaluated in nonhuman primates (NHPs)–the closest living relatives to humans. Simian immunodeficiency virus (SIV) is the closest available non-chimeric virus—NHP model for studying and testing HIV vaccines or therapies. The SIV model complements the simian-human immunodeficiency virus (SHIV) model in distinctive ways, although less is known about SIV Env-specific antibody responses in NHPs. There are several sites on HIV Env that are vulnerable to antibody-mediated protection, and here we isolated and analyzed monoclonal antibodies (mAbs) from NHPs targeting analogous sites on SIV Env. In particular, we studied mAbs for their ability to bind the viral Env protein and to block infection of cells by widely divergent strains of SIV. These well-characterized SIV Env-specific antibodies will allow for more thorough NHP pre-clinical testing of various antibody-based SIV/HIV vaccine and immunotherapeutic strategies before proceeding to human clinical trials and may yield unanticipated findings relating to molecular mechanisms underlying the unusual breadth of neutralization observed in HIV-2 infection.
| Generating protective antibody responses by vaccination is the ultimate goal of an effective HIV vaccine [1–4]. As such, a number of highly potent bnAbs targeting major sites of HIV-1 Env vulnerability such as the CD4bs [5–8], peptido-glycans of variable loops V1, V2 and V3 [9–12], the membrane-proximal external region (MPER) [13–15] and the gp41-gp120 interface [16, 17] have been isolated and examined for their potential impact on HIV vaccine design [18–20]. The specificity and effector functions of protective, non-neutralizing antibodies (pnnAbs) are likewise being scrutinized for their potential complementary role toward protection against HIV infection [21–24]. However, recent studies highlight the challenges to developing an effective HIV-1 vaccine [25–34] and suggest that a better understanding of SIV Env-specific antibody responses might complement and inform HIV vaccine design. This possibility is underscored by the protective effects of Env targeted antibodies elicited by adenovirus-vectored immunogens in SIV protection trials [35–38] and the surprising discovery that HIV-2, a derivative of SIVsmm, commonly elicits bNabs in natural human infection [39–41]. A better understanding of protective SIV Env-specific antibody responses may thus facilitate more effective use of the SIV challenge model to evaluate candidate vaccines and immunotherapies before proceeding to costly, time consuming and resource intensive human clinical trials.
Design of a HIV immunogen that can i) focus the antibody response to protective yet subdominant or sterically hindered epitopes, ii) engage Abs encoded by germline B cell receptors (BCRs) and iii) drive sufficient antibody affinity maturation to generate protective antibody responses will likely require iterative immunogen design [42]. Additional work will be required to optimize the antibody specificities and functions, alone or in combination, which are necessary and sufficient to protect against HIV infection. Finally, it will be necessary to assess which vaccine regimens and adjuvant combinations can achieve the desired germline BCR engagement, affinity maturation, antibody persistence and ultimately, protective efficacy against HIV challenge. All of these unanswered questions necessitate a relevant NHP model for HIV vaccine research
The SHIV—NHP model of HIV infection has been used extensively to study antibody-mediated correlates of protection [43–46]. Chimeric SHIVs are often constructed by replacing the envelope gene and additional accessory proteins of the pathogenic molecular clone of SIVmac239 with corresponding genes from selected HIV-1 subtypes followed by in vivo passaging for enhanced virus replication [47]. Such constructs have proven invaluable for screening candidate HIV immunogens and the development of pathogenic SHIV chimeras has allowed for testing of antibody-mediated protection [48–53]. However, SHIVs have limited genetic diversity [54, 55] compared with SIV challenge stocks that reflect the diversity present in primary circulating isolates of HIV-1 [56]. Thus, protection against SIV may better estimate the protective efficacy of a HIV vaccine and may complement the SHIV model used with clinically relevant reagents. Indeed, vaccine protection against acquisition of neutralization resistant SIV challenges in rhesus macaques suggests a role for antibody-mediated protection [35–37, 57]. However, the epitope specificities and effector functions of SIV-specific antibodies mediating protection have yet to be fully characterized. Thus, developing reagents to study SIV-specific antibody responses in NHP can provide an informative model for defining antibody-mediated correlates of protection.
Our overall goal was to identify SIV-specific antibodies from macaques that may inform the development of effective HIV antibody-based interventions. Some of the most potent HIV-1 bnAbs target the CD4bs, variable regions V1/V2 and the glycan/V3 loop of gp120 [1]. Given the paucity of SIV-specific probes, we designed scaffolded probes to isolate SIV V1V2-specific mAbs and developed a novel competitive probe binding procedure for isolation of SIV mAbs targeting the CD4bs as well as high-mannose glycans on gp120. Both the scaffolded probes and competitive probe binding technique were highly efficient for the targeted isolation of SIV-specific B cells. Subsequent cloning, expression and characterization of individual mAbs identified many novel, potent mAbs targeting multiple sites of SIV Env vulnerability, including the first reported SIV CD4bs-specific neutralizing mAbs isolated from SIV-infected rhesus macaques.
SIV-positive plasma and peripheral blood mononuclear cells (PBMC) were obtained from previously completed animal study protocols [35, 38, 58] (S1 Table).
A protein scaffold (1JO8) [59] that provides an appropriate hairpin was identified to suitably incorporate the SIV Env V1V2 region based on stable expression, clash score and solvent accessibility. This scaffold allows V1V2 to be expressed at high yield in a context that maintains proper conformation of a native V1V2 protomer [60]. A soluble trimeric SIVmac239 gp140 foldon protein expression vector was generated by encoding SIVmac239 from residues 1 thru 722, followed by the foldon trimerization motif as previously described [61]. The following mammalian expression vectors were used for synthesis of SIV proteins: pcDNA3.1(-) encoding SIVmac239 gp140 foldon trimer (FT) and pVRC8400 [62] encoding either 1JO8-scaffolded SIVsmE660.CP3C or SIVsmE660.CR54 V1V2 loop sequences (GenScript). All constructs contained C-terminus 6X His-tag for protein purification followed by an Avi-tag motif for biotinylation. The SIVmac239 ΔV1V2V3 gp120 plasmid encoding gp120 residues 44 to 492 (HXBc2 numbering) with truncations in the V1V2 and V3 regions as in previously reported HIV-1 CoreE gp120 proteins [63], in which residues 124 to 198 in the V1V2 loop and residues 302 to 323 in the V3 loop of SIV gp120 were replaced with GG and GGSGSG linkers, respectively and kindly provided by Andrés Finzi. Construction of a synthetic gene encoding full-length cyanovirin-N (CVN) inserted into a pET-26(+) vector (Novagen) has been previously described [64]. The CD4-Ig plasmid encoding the first two N-terminal domains of the CD4 molecule which are sufficient for high-affinity gp120 binding fused with the Fc region of human IgG1 was kindly provided by Joseph Sodroski [65].
All SIV proteins and CD4-Ig were expressed by transient transfection of 293Freestyle (293F) cells in serum-free medium using 293fectin transfection reagent (Invitrogen) according to manufacturer’s instructions. Cell culture supernatants were harvested 6 days post-transfection, passed through a 0.22 μm filter to remove any cell debris and supplemented with protease inhibitor tablets (Roche). All SIV proteins were purified using Ni Sepharose excel affinity media (GE Healthcare) followed by size exclusion chromatography (SEC) on a HiLoad 16/600 200 pg Superdex column (GE Healthcare). CD4-Ig was purified using a recombinant protein A affinity column (GE Healthcare) as previously described [66]. Recombinant CVN was produced as previously reported [67]. Briefly, CVN was expressed in the BL21-DE3 E. coli strain (New England Biolabs), followed by purification using reversed-phase chromatograpy (Sep-Pak Vac 35cc (10g) tC18 cartridges, Waters) and gel-filtration (Superdex 75, GE Healthcare) to ensure separation of monomeric and domain-swapped dimeric CVN.
Cryopreserved PBMC were thawed and stained with LIVE/DEAD Fixable Violet Dead Cell Stain (Life Technologies) as previously described [68, 69]. Cells were washed and stained with an antibody cocktail of CD3 (clone SP34-2, BD Biosciences), CD4 (clone OKT4, BioLegend), CD8 (clone RPA-T8, BioLegend), CD14 (clone M5E2, BioLegend), CD20 (clone 2H7, BioLegend), IgG (G18-145, BD Biosciences) and IgM (clone G20-127, BD Biosciences) at room temperature in the dark for 20 mins. The cells were washed twice with PBS and subsequently stained with fluorescently labeled SIV probes to stain for CD4bs-, cyanovirin binding site (CVNbs)- or V1V2-specific B cells.
For staining of CD4bs-and CVNbs-specific B cells, SIVmac239 gp140 FT was used in combination with 4-fold or 5-fold molar excess of CD4-Ig fusion protein or CVN protein, respectively. Cells were first re-suspended in 200 μl PBS with CD4-Ig:SIVmac239 gp140-PE or CVN:SIVmac239 gp140-PE, respectively, incubated at room temperature in the dark for 20 mins followed by 3 washes with PBS and then re-suspended in 200 μl PBS containing SIVgp140-APC and incubated further for 20 mins at room temperature in the dark. For staining of V1V2-specific B cells, cells were re-suspended in 200 μl PBS containing PE-labeled 1JO8 SIVsmE660.CP3C V1V2 and/or 1JO8 SIVsmE660.CR54 V1V2 and incubated for 20 mins at room temperature in the dark. The stained cells were washed 3 times and re-suspended in 1 ml of PBS, passed through a 70 μm cell mesh (BD Biosciences) then analyzed and sorted with a modified 3-laser FACSAria cell sorter using the FACSDiva software (BD Biosciences). Probe-positive B cells were sorted as single cells into wells of a 96-well plate containing lysis solution as previously described [5]. Flow cytometric data was subsequently analyzed using FlowJo (v9.7.5).
Single B cell RNA was reverse transcribed as previously described [5], diluted 2-fold by addition of 26 μl nuclease-free water and the cDNA plates were stored at -20°C. Individual rhesus immunoglobulin (Ig) heavy (H), light kappa (Lκ) and light lambda (Lλ) chain genes were amplified by nested PCR using 5 μl cDNA as template. All PCR reactions were performed in 96-well PCR plates in a total volume of 50 μl. For first round amplification, first-round rhesus-specific PCR primers (S2–S4 Tables) were used to amplify gene transcripts containing 2 U of HotStar Taq Plus DNA Polymerase (QIAGEN), 1 μl dNTP-Mix (10 mM each nucleotide) (QIAGEN), 0.5 μg carrier RNA, 1 mM MgCl2, 1 μl forward primer mix (50 μM each primer), 1 μl reverse primer (25 μM each primer), using the following PCR program: 5 min at 94°C; 50 cycles of 30 sec at 94°C, 45 sec at 50°C, 45 sec at 72°C: followed by 10 min at 72°C. One-twentieth the volume of first-round PCR product was amplified by nested PCR with second-round rhesus-specific PCR primers (S2–S4 Tables) under the same conditions used for first round PCR. The second round of PCR was performed for 5 min at 94°C followed by 30 cycles of 30 sec at 94°C, 45 sec at 60°C, 45 sec at 72°C and a final 10 min extension at 72°C. Amplified PCR products were analyzed on 2% agarose gels (Embi-Tec) and positive reactions sequenced directly. PCR products with productive Igγ and IgLκ or IgLλ sequence were re-amplified with 3 μl of unpurified first round PCR product as template and combinations of single gene-specific V and J gene primers incorporating unique restriction digest sites. Resulting PCR products were run on a 1% agarose gel and purified with QIAGEN Gel Extraction Kit (QIAGEN) and eluted with 25 μl nuclease-free water (Quality Biolgical). Purified PCR products were digested with appropriate restriction digest enzymes AgeI, Nhel, BsiWI and ScaI (all from ThermoScientific) before ligation into rhesus Igγ, IgLκ and IgLλ expression vectors containing a murine Ig gene signal peptide sequence (GenBank accession number DQ407610) and a multiple cloning site upstream of the rhesus Igγ, Igκ or Igλ constant regions (all 3 expression vectors were kindly provided by Kevin Saunders). Transcription of these expression vectors is under the influence of human cytomegalovirus (HCMV) promoter allowing clones to be selected based on resistance to kanamycin. Full-length IgG was expressed as previously described [5] by co-transfecting 293F cells with equal amounts of paired heavy and light chain plasmids then purified using Protein A Sepharose beads (GE Healthcare) according to manufacturer’s instructions.
Binding of SIV-specific mAbs to purified proteins or synthetic peptides was measured by enzyme-linked immunosorbent assay (ELISA) as previously described [5]. For CD4bs competition ELISA, plates were coated with 2 μg/ml SIVmac239 gp140 FT in PBS at 4°C overnight. After blocking with 200 μl B3T buffer (150mM NaCl, 50mM Tris-HCl, 1mM EDTA, 3.3% fetal bovine serum, 2% bovine albumin, 0.07% Tween-20), serial dilutions of unlabeled competitor mAbs were added to captured SIVmac239 gp140 FT in 100 μl B3T buffer for 15 mins prior to addition of biotinylated CD4-Ig (at a concentration determined to yield O.D. of roughly 1.0–2.0). Alternatively, competition with sCD4 was performed by addition titrating amounts of sCD4 to SIVmac239 gp140 FT-coated plates for 15 mins prior to addition of individual mAbs (at a concentration determined to yield O.D. of roughly 1.0–2.0) and binding was detected by HRP-conjugated anti-monkey IgG (Rockland Immunochemicals) at a 1:5,000 dilution for 1 hour. Antibody cross-competition ELISA was performed by adding titrations of unlabeled competitor mAbs to SIVmac239 gp140 FT-coated plates for 15 mins prior to addition of individual biotinylated mAbs (at a concentration determined to yield O.D. of roughly 1.0–2.0). Plates were incubated for 1 hr at 37°C, washed 3 times with B3T buffer followed by incubation with streptavidin-horseradish peroxidase (HRP) for 1 hr at 37°C. For peptide competition ELISA, titrations of peptides were added to SIVmac239 gp140 FT-coated plates for 15 mins prior to addition of individual mAbs (at a concentration determined to yield O.D. of roughly 1.0–2.0) and binding detected by HRP-conjugated anti-monkey IgG (Rockland Immunochemicals) as above. The signal was developed by addition of 3,3′,5,5′-tetramethylbenzidine (TMB) substrate (SureBlue; KPL) for 10 min. Reactions were terminated with 1 N sulfuric acid, and the optical density (OD) was read at 450 nm. The following reagent was obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH: SIVmac239 Env Peptide Set. The following reagent was obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH: Soluble Human CD4 from Progenics and sCD4-183 from Pharmacia, Inc. [73]
Plasmid DNA encoding SIV gp160 was used in combination with a luciferase reporter plasmid containing the essential HIV structural genes to produce SIV Env pseudoviruses as described previously [74]. Plasmids encoding SIV gp160 for clones SIVsmE660.11 [75], SIVmac239.cs.23 [76], SIVmac251.6 [76] and SIVmac251.cs.41 [77] were generously provided by David Montefiori. Plasmids encoding SIV gp160 for clones SIVsm (FFv 18Nov04 ENVPL2.1, FJv 15Nov06 ENVPL2.1, FWk 12Aug04 ENVPL4.1 and RSo8 17Jan06 ENVPL1.1) and SIVmac251 (RSo8 17Jan06 ENVPL1.1 and RZj5 9Apr09 ENVPL2.1) were kindly provided by Cynthia Derdeyn [78]. Full-length infectious molecular clones of transmitted/founder viruses corresponding to SIVsmm lineage 1 (RM174.V1,V2,V3.tf), 5 (FTq) and “outlier” (SL92b) were derived by methods previously described [79]. In brief, naïve Indian rhesus macaques were inoculated intravenously with plasma from sooty mangabey monkeys naturally infected with the SIVsmm lineage 1 or 5 viruses or with a primary lymphocyte culture of SIVsmm SL92b obtained from a naturally infected sooty mangabey. All three rhesus macaques became productively and chronically infected. Chronic plasma from these three animals was then inoculated intravenously into three naïve Indian rhesus macaques. Twelve days later, acute infection plasma was collected, plasma viral RNA isolated, viral cDNA generated, full-length T/F SIVsmm sequences single genome amplified by limiting dilution PCR, and the products molecularly cloned, as described [79]. IMC sequences (GenBank accession numbers KU182919-23) were identical to the respective inferred T/F viral genomes, which represent examples of highly diverse naturally-occurring strains of SIVsmm [80]. Virus neutralization was measured using single round infection of TZM-bl target cells by SIV Env-pseudovirus or replication-competent viruses i.e. infectious molecular clones (IMC) in the presence of the protease inhibitor indinavir as previously described [81]. The 50% inhibitory concentration (IC50) was defined as the antibody concentration that caused a 50% reduction in relative light units (RLU) compared to virus control wells after subtraction of background RLU. Half-maximal inhibitory concentration (HalfMax) was defined as the antibody concentration that caused a 50% reduction in maximum neutralization for a given mAb while the maximum neutralization (VMax) was defined as the maximum % neutralization observed over the range of mAb concentrations tested.
To identify SIV CD4bs-specific B cells, we prepared 2 probes exhibiting differential binding capacity for CD4bs-specific B cells (Fig 1). A CD4bs-occluded SIVmac239 gp140 FT probe was prepared by mixing PE-conjugated SIVmac239 gp140 FT with a 4-fold molar excess of CD4-Ig fusion protein [66]. An APC-conjugated SIVmac239 gp140 FT served as a CD4bs-accessible SIVmac239 gp140 probe. SIV CD4bs-specific B cells were identified by first staining cells with CD4-Ig:SIVmac239 gp140-PE to label all SIV-specific B cells except those blocked by excess CD4-Ig (Fig 1). After extensive washing to remove unbound CD4-Ig:SIVmac239 gp140-PE and excess CD4-Ig, cells were stained with SIVmac239 gp140-APC in order to label all SIV-specific (including CD4bs-specific) B cells (Fig 1). Thus, cells stained negative for CD4-Ig:SIVmac239 gp140-PE but positive for SIVmac239 gp140-APC identified putative CD4bs-specific B cells. B cells binding to both CD4-Ig:SIVmac239 gp140-PE and SIVmac239 gp140-APC would be expected to bind to gp140 at epitopes outside of the CD4bs.
In combination with a rhesus B cell staining panel (Fig 2A) we used this competitive probe binding staining procedure to sort 160 putative CD4bs-specific B cells from 4.4 million PBMC (0.02% of total B cells) from a SIVmac251-infected rhesus macaque (DBM5) [58] (S1 Table; Fig 2B). Amplification of immunoglobulin heavy and light chain variable regions yielded 37 matched heavy and light chain pairs belonging to 13 clonal families. We cloned and expressed 16 mAbs to characterize their binding and neutralization activity (Table 1). Among these 16 mAbs were 3 distinct clonal families including ITS07 and ITS16 mAbs. Rhesus heavy chain V gene usage of cloned mAbs was mostly restricted to IGHV4 alleles although ITS07 mAbs were IGHV3. Mutation frequency of heavy and light chain genes within the V region (based on nt sequence divergence from nearest assigned germline sequence) ranged from 2–11%.
All putative CD4bs-specific mAbs were tested by ELISA for binding to SIVmac239 gp140 FT used for cell sorting. While 4 mAbs showed no detectable binding, 12 mAbs bound to SIVmac239 gp140 FT with varying affinities (Fig 3). These 12 mAbs also bound to monomeric SIVmac251.30 gp140 and in most cases to SIVsmE660.CP3C and/or smE660.CR54 gp120s as well. To evaluate whether mAbs cloned from B cells isolated using our 2-step competitive probe binding staining protocol were indeed specific for the CD4bs, we re-evaluated binding of the 12 SIV-specific mAbs to SIVmac239 gp140 FT by competition ELISA with CD4-Ig. Binding of CD4-Ig to SIVmac239 gp140 FT was effectively competed by 11 out of 12 mAbs (Fig 2C) confirming their specificity for the CD4bs.
For those mAbs that failed to bind to SIVmac239 gp140 FT, we hypothesized that residual CD4-Ig in the staining protocol may have facilitated SIVgp140-APC labeling of CD4-induced (CD4i)-specific B cells, i.e. binding to an epitope at or near the host cell co-receptor binding site which is exposed following binding of the primary receptor CD4. We tested whether the presence of soluble CD4 (sCD4) could facilitate binding of mAbs DBM5-2E10, 2E11, 2B3 and 1A11 to SIVmac239 gp140 FT. The addition of sCD4 had no effect on binding of these mAbs to SIVmac239 gp140 FT indicating these mAbs bound neither CD4bs- nor CD4i-specific B cells (Fig 2D). Nonetheless, our sort strategy was highly efficient for isolating CD4bs-specific B cells, as 11 out of 12 SIV-binding mAbs were CD4bs-specific.
We next assessed neutralizing activity of the 11 CD4bs-specific mAbs by TZM-bl assay against 4 SIV Env pseudoviruses and the IMC HIV-27312A. All but one CD4bs-specific mAbs neutralized the highly neutralization-sensitive (tier 1) isolates SIVsmE660.CP3C and SIVmac251.H9.15, the moderately neutralization-resistant (tier 2) isolate SIVsmE660.CR54 as well as the primary isolate HIV-27312A (Fig 4). ITS02 was unique among the CD4bs-specific mAbs for its strain-specific neutralization of SIVmac251.H9 but not SIVsmE660, SIVmac251.30 or HIV-27312A. None of the CD4bs-specific mAbs cross-neutralized the highly neutralization-resistant (tier 3) SIVmac239 (S1 Fig). As previously reported by other groups we observed that neutralization curves of tier 2 isolates of SIVsmE660 and SIVmac251 plateaued below 100%, and in some instances, below 50% neutralization despite using clonal, pseudo-typed viruses [39, 56, 82]. In order to compare the potency of individual mAbs, irrespective of neutralization plateau levels, we also calculated half-maximal (HalfMax) concentrations (i.e., the concentration required to achieve half-maximal neutralization) as well as maximum percent neutralization (VMax) values (i.e. the maximum % neutralization over the range of mAb concentrations tested) for individual mAbs (Fig 4). Based on these values, we determined that ITS01 and ITS20 were also weakly neutralizing against tier 2 SIVmac251.30 and that the potency of individual CD4bs mAbs was similar irrespective of the VMax levels.
Given the efficiency of our CD4bs competitive binding sort technique, we used the same strategy to isolate antibodies specific for the SIV Env glycan targets of cyanovirin (CVN), a potent inhibitor of primary and lab-adapted isolates of HIV and SIV [83]. Cyanovirin selectively binds to Man8 D1D3 and Man9 residues on N-linked glycans present on gp120 [67]. Binding of CVN also occludes the unique 2G12 neutralization epitope of HIV-1 Env gp120 [84]. We sorted putative cyanovirin binding site (CVNbs)-specific B cells from 4 SIVsmE660-infected rhesus macaques (05D247, A4V014, ZB08, ZB42) [35] (S1 Table; Fig 5A) and cloned and expressed a total of 32 mAbs from these 4 macaques (Table 2). Although 29 out of 32 mAbs cloned from the CVNbs competition sort bound to SIVmac239 gp140 FT (Fig 3), only 11 of these mAbs competed with CVN for binding to gp140 (Fig 5B). Intriguingly, the presence of CVN rescued the binding of 2 mAbs (ITS56 and ITS57) that failed to bind to SIVmac239 gp140 FT alone suggesting recognition of an epitope on both gp140 and CVN that is only present when they are bound together or conformational change(s) induced upon CVN binding that facilitated binding by these mAbs.
The anti-HIV-1 activity of CVN is reportedly mediated through high affinity interactions with oligomannose residues suggesting multiple potential binding sites [67]. However, previous studies have reported that binding of CVN occludes subsequent binding of the bnAb 2G12 but not mAbs targeting the V3 and V4 loops, C4 region, CD4bs or CD4i epitopes of HIV-1 indicating a more defined binding epitope for CVN [84]. To map the SIV epitope(s) targeted by mAbs isolated by CVNbs competition we also assessed competition between these mAbs and sCD4. The presence of competing sCD4 did not block binding of any mAbs isolated by CVNbs competition (Fig 5B). However, one mAb (ITS51) showed enhanced binding to SIVmac239 gp140 FT in the presence of sCD4 (Fig 6A) indicating this mAb likely targets a CD4i site. Additionally, we tested mAbs from this sort for binding to overlapping SIVmac239 Env 15-mer peptides. Of 31 mAbs that exhibited SIV gp140 FT binding, 30 were negative for binding to SIVmac239 Env peptides. Only ITS52 was mapped to a linear peptide sequence near the V3 loop tip (Fig 6B). To further evaluate the epitope binding specificities of mAbs isolated by CVNbs competition, we tested binding of individual mAbs to a SIVmac239 ΔV1V2V3 gp120 core protein generated by deletion of the V1V2 loops and truncation of the V3 loop [63]. Of 31 mAbs that bound to SIVmac239 gp140 FT, 7 were also positive for binding to SIVmac239 ΔV1V2V3 gp120 core protein (Fig 6C). As expected, ITS01, a CD4bs-specific mAb also bound to ΔV1V2V3 gp120 core protein while V2-specific (ITS03) and V3-specific (ITS52) mAbs did not.
To assess overlapping SIV Env epitope-binding specificities of ITS52 and other mAbs isolated by CVNbs competition, we performed a matrix cross-competition ELISA of CVNbs mAbs. For 11 CVNbs mAbs tested there were multiple patterns of cross-competition although most were competed efficiently by multiple mAbs suggesting overlapping specificities (Fig 7). The ability of ITS01 to compete with ITS51 for binding to SIVmac239 gp140 FT, together with the sCD4 and CD4-Ig competition data (Fig 6A), indicate that ITS51 targets the CD4i-site of SIV. Overall, our CVNbs competition sort yielded mAbs targeting the V3 loop, CD4i-site and ΔV1V2V3 SIVmac239 gp120 core protein.
Cyanovirin has been reported to neutralize HIV-1, HIV-2 and SIV primary isolates at low nanomolar concentrations [67, 83]. To determine whether mAbs isolated by CVNbs competition could mediate similar virus neutralization breadth and potency, we assessed neutralization activity of all 32 mAbs isolated by the CVNbs competition sort against a small panel of SIV Env pseudoviruses and HIV-27312A. Nineteen mAbs were non-neutralizing against all 5 viruses tested while 13 mAbs neutralized SIVsmE660.CP3C (tier 1), SIVsmE660.CR54 (tier 2) and SIVmac251.H9 (tier 1) (Fig 4). Of these, 8 mAbs also neutralized SIVmac251.30 (tier 2) and 3 cross-neutralized HIV-27312A. Thus, CVNbs mAbs showed similar virus neutralization cross-reactivity as CD4bs mAbs; however, the neutralization potency of several CVNbs mAbs was significantly higher than that of CD4bs mAbs. In some instances 10,000-fold lower IC50 (1,000-fold lower HalfMax) values were obtained for CVNbs mAbs compared with the most potent neutralizing CD4bs mAbs (Fig 4). Of note, neutralizing activity correlated with specificity for the CVNbs. All 11 mAbs which efficiently competed with CVN for binding to SIVmac239 gp140 FT were neutralizing while of the remaining non-CVNbs-specific mAbs tested, only ITS56 and ITS57, which required the presence of CVN for binding to SIVmac239 gp140 FT, were neutralizing (Figs 4 and 5B).
Both human and non-human primate studies have shown that V1V2 serum IgG binding levels correlate with protection against HIV/SIV infection [36–38, 85]; therefore, we were also interested in isolating SIV V1V2-specific B cells. We generated 1JO8-scaffolded SIV V1V2 probes [60] from tier 1 (SIVsmE660.CP3C) and tier 2 (SIVsmE660.CR54) isolates of the SIVsmE660 challenge swarm used in a recently completed SIV challenge study [38]. Individual fluorescently labeled 1JO8-scaffolded SIV V1V2 probes were used to stain and isolate V1V2-specific B cells from a SIVmac239-vaccinated and SIVsmE660-infected rhesus macaque (S1 Table). We sorted 36 (0.13% of memory B cells) and 110 (0.16% of memory B cells) SIV V1V2-specific B cells using 1JO8-scaffolded SIVsmE660.CP3C and SIVsmE660.CR54 V1V2 probes, respectively (Fig 8). A total of 26 mAbs belonging to 9 distinct clonal families were cloned from 146 individually sorted cells and some of these represented identical clones (Table 3). There were 20 unique mAbs: 3 identical clones of ITS03 and 2 identical clones for each of ITS09.01, ITS10.04, ITS12.01 and ITS30.
The 1JO8 SIVsmE660 probes were highly specific since only one of 20 mAbs expressed did not bind to SIVsmE660 gp120 (Fig 3). The remaining 19 mAbs bound to both SIVsmE660.CP3C and SIVsmE660.CR54 gp120s with several also binding to SIVmac251.30 gp120 and SIVmac239 gp140 FT. The epitope binding specificities of SIV V1V2 mAbs were assessed by ELISA binding using overlapping 15-mer SIVmac239 Env peptides spanning the V1V2 region. Peptide mapping revealed that ITS06.01 and ITS06.02 bound to several peptides from V1 corresponding to Env 109–147 (SIVmac239 numbering) (Fig 9A) suggesting a potentially discontinuous epitope within this region while ITS12.01 and ITS12.02 targeted Env 185–195 (ETWYSADLVCE) (Fig 9B), an epitope at the C-terminus of the V2 loop. Interestingly, although both of these mAbs bound to this SIVmac239 peptide epitope, neither mAb bound to either SIVmac239 gp140 FT or SIVmac251.30 gp120. Another linear B cell epitope in V2 was identified by ITS03 and ITS09.01–04 mAbs which bound within Env 173–183 (TGLKRDKKKEY) (Fig 9B). While Env 173–179 was sufficient for binding by ITS03, ITS09.02 and ITS09.04, additional residues were necessary for binding by ITS09.01 and ITS09.03. All mAbs which did not bind to either SIVmac239 gp140 FT or 15-mer peptides were screened for binding to potentially protective SIVsmE660 15-mer peptide sequences [38]. These mAbs all bound to a 15-mer peptide corresponding to Env 142–156 (ENVINESNPCIKNNS), an epitope that is present in SIVsmE660 but not SIVmac239 (S2 Fig).
We also assessed neutralizing activity of SIV V1V2-specific mAbs against tier 1 and 2 clonal isolates of SIVsmE660 and SIVmac251 and HIV-27312A. Despite strong binding to peptide ENVINESNPCIKNNS, which is present in SIVsmE660.CP3C and SIVsmE660.CR54, only 4 of 9 mAbs specific for this epitope showed weak neutralization against SIVsmE660.CP3C (Fig 4). Likewise, ETWYSADLVCE-specific mAbs ITS12.01 and ITS12.02 showed strain-specific neutralization of SIVsmE660 isolates but not SIVmac251 or SIVmac239 despite strong binding to SIVmac239 gp140 FT and linear peptides. In contrast, ETDRWGLTKSI-specific mAbs ITS06.01, ITS06.02 and ITS13 were cross-neutralizing for SIVsmE660 and SIVmac251 isolates albeit with varying degrees of breadth and potency. ITS06.02 and ITS13 neutralized tier 1 and tier 2 isolates of SIVsmE660 and the tier 1 isolate SIVmac251.H9 but not SIVmac251.30 (tier 2). ITS06.01 was the only V1-specific mAb which neutralized both tier 1 and 2 isolates of SIVsmE660 and SIVmac251. Three V2-specific mAbs targeting the TGLKRDKKKEY epitope (ITS03, ITS09.03 and ITS09.04) were also neutralizing against the same 4 isolates. None of the V1V2-specific mAbs neutralized SIVmac239 or HIV-27312A.
The 1JO8-scaffolded V1V2 probes efficiently labeled V1V2-specific B cells from the chronic phase of SIV infection; however, we wanted to determine whether these probes could also be used to isolate low-frequency vaccine-elicited B cells that might be cross-reactive for heterologous challenge virus. We used the 1JO8 SIVsmE660 V1V2 probes to sort B cells from a pre-challenge, SIVmac239-vaccinated macaque (ZG12) [38] (S1 Table). In order to maximize probe binding to heterologous, low frequency pre-challenge memory B cells, we used both 1JO8 SIVsmE660.CP3C and SIVsmE660.CR54 V1V2 probes in combination to sort 74 (0.8% of memory B cells) SIV V1V2-specific B cells (Fig 10A). We cloned 7 unique mAbs belonging to 6 clonal families (Table 4) and characterized their binding specificities and neutralization activity. Six out of 7 mAbs tested bound to the 1JO8 SIV probes used for cell sorting and to SIVmac239 gp140 FT (Fig 3). Peptide mapping revealed that ZG12-2H10, which failed to bind SIVmac239 gp140 FT, did not bind to any SIVmac239 Env 15-mer overlapping peptides and was non-neutralizing against all viruses tested (Fig 4). Five mAbs were mapped to 1 of 3 epitopes including two new epitopes not identified by infection-related V1V2 mAbs (Fig 3). Both ITS40 and ITS41 which targeted the V2 epitope EQEQMISCKFNMTGL (Fig 10B), only neutralized tier 1 SIVsmE660.CP3C (Fig 4) while ITS45 targeted Env 101–115 (CVKLSPLCITMRCNK) (Fig 10C) but was non-neutralizing against all isolates tested. Similar to ITS06.01 and ITS06.02 mAbs isolated from chronic SIV infection, ITS42 and ITS44, bound to several peptides from V1 corresponding to Env 109–147 (Fig 10C) and neutralized tier 1 isolates of SIVsmE660 and SIVmac251 as well as the tier 2 isolate SIVsmE660.CR54 (Fig 4).
While several SIV mAbs were cross-neutralizing for SIVmac251 and SIVsmE660 isolates tested, the small panel of closely related SIV Env pseudoviruses used to test neutralization activity limits our ability to assess neutralization breadth of individual SIV mAbs. Therefore, we tested selected SIV mAbs for neutralization against an additional 15 SIV isolates including 10 SIV Env pseudoviruses and 5 transmitted/founder IMCs (Fig 11). Combined with the SIVmac251, SIVsmE660 and HIV-27312A viral isolates initially tested (Fig 4), this expanded 21-virus panel (Fig 11A) more closely reflects inter-clade genetic diversity of HIV-1 (Fig 11B) [80]. Most SIV mAbs tested showed neutralization of multiple SIV strains including neutralization of genetically diverse tier 2 and tier 3 SIV isolates. (Fig 12). In general, SIV CD4bs and CVNbs mAbs displayed greater neutralization breadth than SIV V1V2 mAbs while CVNbs mAbs were among the most potent (Fig 13, S6–S8 Figs). Neutralization breadth, as measured by the percentage of SIV/HIV-2 isolates neutralized, was greatest for CD4bs mAbs ITS01 and ITS20, which neutralized up to 81% and 85% of viruses tested, respectively (Fig 13).
Unsurprisingly, strain-specific mAbs ITS02 (CD4bs) and ITS10.01 (V1) had the lowest neutralization breadth, 20% and 14%, respectively, while the neutralization breadth of CVNbs mAbs ranged from 62–76%.
Broadly neutralizing mAbs isolated from HIV-1 infected individuals exhibit some unique features such as high diversity in the variable heavy chain region (VH) genes due to extensive somatic hypermutation (SHM) [86] and long, protruding CDRH3 sequences [87]. The level of SHM for HIV-1 bnAbs ranges from 11–32% divergence from putative VH germline nucleotide sequence [88] while HIV-specific antibodies with low or no neutralizing activity display approximately 9–12% VH sequence divergence from germline [89, 90]. Compared with HIV-1 bnAbs or even rhesus memory B cells, which are approximately 5% divergent from VH germline nucleotide sequence [91], the SIV-specific bnAbs isolated here have relatively low levels of SHM. With regard to CDRH3 length, HIV-1 CD4bs bnAbs have relatively short CDRH3 sequences while those targeting quaternary bnAb epitopes of the V1/V2 and V3 loops have long CDRH3s likely to facilitate penetration of the glycan shield and access to the V1/V2 and V3 loops [12, 87]. Based on the distribution of CDRH3 length in rhesus naïve B cells [91], we did not observe unusually long CDRH3 sequences among SIV V1V2-specific mAbs and this was not altogether surprising since the V1V2 mAbs isolated in this study all targeted linear peptide epitopes and displayed limited neutralization breadth against our 21 virus panel. By comparison, some of the SIV CD4bs and CVNbs bnAbs had longer CDRH3 sequences; however, there is no structural data as yet to support the requirement for long CDRH3s to access recessed epitopes by SIV bnAbs as is the case for HIV-1 quaternary-preferring bnAbs with long CDRH3s. In general, neutralizing SIV mAbs did not display some of the unique features frequently observed in HIV-1 bnAbs.
The SIV NHP model for HIV-1 is useful for studying vaccine mediated and immune correlates of protection but little is known about binding or neutralizing epitopes on SIV Env. Our goal was to isolate and characterize SIV Env-specific mAbs that might facilitate effective use of this NHP model for understanding the variables in development of a HIV vaccine or immunotherapy. We demonstrate the use of a novel competitive probe binding strategy for the targeted isolation of SIV Env-specific mAbs from rhesus macaques and present a detailed assessment of nearly 70 SIV mAbs targeting the CD4bs, CD4i-site, CVNbs and V1, V2 and V3 loops of SIV Env. We characterized individual SIV mAbs with regard to immunoglobulin genetics, epitope specificity, peptide and protein binding as well as virus neutralization breadth and potency.
Various studies have characterized neutralization epitopes of SIV using murine, guinea pig, rabbit and goat antisera [92, 93] as well as murine- [94–98] or rhesus-derived [99–102] SIV-specific mAbs. The range of epitopes described includes SIV mAbs targeting linear epitopes in variable loops 1–4 [38, 92, 93, 95–97, 100–102] and conformational epitopes involving the V3-V4 region [100, 101] as well as those overlapping or proximal to the CD4bs [95, 98] and co-receptor binding site [95]. However, SIV-specific mAbs isolated to date have been produced exclusively from hybridomas [94–98], EBV-transformed B cells [99–101] or phage display [102]. Ours is the first study to describe the targeted isolation of SIV epitope-specific B cells from rhesus macaques using direct and indirect binding to novel SIV probes.
Given the paucity of reagents for the targeted isolation of SIV-specific B cells we developed a simple competitive probe binding strategy to sort CD4bs-directed B cells from which we cloned multiple SIV CD4bs mAbs. Nearly 70% of mAbs isolated by this method were confirmed to target the CD4bs—significantly improved efficiency as compared to the isolation of HIV CD4bs mAbs using HIV-1 resurfaced stabilized core 3 (RSC3) protein [5]. By substituting CVN in place of CD4-Ig as competitor ligand, we were able to modify the target cell population to sort CVNbs-specific B cells. We subsequently cloned several mAbs from sorted B cells and confirmed their specificity for the CVNbs, thereby validating this competitive probe binding strategy as a powerful technique for the targeted isolation of SIV-specific B cells. Based on the simplicity and efficiency of our competitive probe binding sort strategy we propose that this method may be preferable to the use of engineered probes for targeted isolation of epitope-specific B cells—at least epitopes for which probe binding ligands are available. With this method, even antibodies could serve as competitive ligands for Env trimer probes to facilitate isolation of additional antibodies targeting a given epitope without the need for time-consuming probe development. Competitive ligands cross-reactive for divergent HIV/SIV Env probes could extend the applicability of this competitive sort strategy to diverse strains of HIV/SIV. This simple and effective competitive probe sort technique may also prove useful for the isolation of virus-specific B cells in general. The 1JO8-scaffolded SIV V1V2 probes we designed and tested were also remarkably efficient at labeling both high frequency infection-related and low frequency vaccine-elicited SIV V1V2-specific B cells. Several of the SIV V1V2 epitopes targeted by mAbs isolated with the 1JO8-scaffolded SIV V1V2 probes have previously been identified following isolation of B cells using other strategies [103] thereby validating the use of these probes for targeted isolation of V1V2-specific B cells. Overall, the ability to efficiently target SIV mAbs of defined specificities will increase the usefulness and relevance of the SIV model for studying the induction and maturation of virus-specific B-cells.
Many SIV-specific B cell epitopes previously reported have been identified using murine derived or HIV-2-specific mAbs exhibiting cross-reactivity with SIV [103]. Here, we provide the most extensive study of SIV Env-specific mAbs isolated from rhesus macaques including the first reported rhesus SIV CD4bs-specific mAbs. Of nearly 70 SIV Env-specific mAbs isolated the most broadly neutralizing SIV mAbs were CD4bs-specific, likely due to the conserved nature of the CD4bs for maintaining functional contact with its primary receptor CD4. Indeed, cross-reactivity of HIV-2 CD4bs mAbs for SIV has previously been reported [39]. Among the CD4bs mAbs, ITS02 was notable for its strain-specific neutralization of SIVmac251 (tier 1) only, suggesting that despite strong competition with CD4-Ig for binding to SIV Env, its epitope is likely proximal to rather than directly at the CD4bs.
As with the CD4bs mAbs, CVNbs mAbs also displayed considerable neutralization breadth against our 21-virus panel; however, only a fraction of CVNbs mAbs cross-neutralized HIV-2. Given the lack of information regarding epitope specificity of most of the CVNbs mAbs it is unclear whether the lack of cross-reactivity for HIV-2 is due to sequence, glycosylation or other conformational differences between SIV and HIV-2 or some combination thereof. The single CVNbs mAb for which we determined peptide-binding specificity within the V3 loop, neutralized 62% of isolates tested despite minimal sequence variation in this epitope among SIVs and HIV-27312A. Of note, several CVNbs mAbs were significantly more potent than either the SIV CD4bs- or V1V2-directed mAbs. Among these were 2 clonally related CVNbs mAbs (ITS61.01 and ITS61.02) with extraordinarily high potency and unusually long (25 residues) heavy chain complementarity determining region 3 (CDRH3) loops similar to V2 and V3 glycan reactive mAbs that are among the most potent HIV-1 bnAbs [9, 12, 104–106]; however, these high potency mAbs do not target V2 or V3 glycans since they bind to V1,V2,V3-deleted SIV gp120 core protein.
Compared with the relatively minor sequence variation between the V3 loops of SIV/HIV-2, there is considerable sequence diversity between SIV and HIV-2 within the V1 and V2 loops and this was reflected by the fact that all SIV V1V2 mAbs isolated were non-neutralizing against HIV-2. Interestingly, all SIV V1V2 mAbs were mapped to linear peptide epitopes although it is unclear whether this was due primarily to the 1JO8-scaffolded probe used for isolating B cells and/or the immunization history or immune response of animals used for cell sorting. Among the major sites of HIV-1 Env vulnerability, the V1V2 loops are of particular interest for an HIV vaccine based on results of both human and NHP vaccine efficacy trials showing that levels of V1V2-specific serum binding, but not neutralizing activity, directly correlate with resistance to HIV/SIV infection [38, 85]. Among the SIV V1V2 mAbs isolated we identified both neutralizing as well as binding, non-neutralizing mAbs which may serve as useful reagents for delineating the role of V1V2-binding mAbs towards protection against infection. An important caveat is the presence of two conserved cysteine residues in the V2 region of most SIV and HIV-2 strains, which are absent in all HIV-1 strains [107]. These twin-cysteine residues may form a disulfide bond that contributes to Env trimer stabilization since twin-cysteine mutants exhibit decreased gp120 association with the Env trimer cell-cell fusion and virus infectivity. Future studies will need to address whether the conserved twin-cysteine motif may contribute to structural or functional differences between SIV/HIV-2 and HIV-1 and any potential impact on V1V2-directed mAb responses in SIV/HIV-2.
Overall, we isolated multiple SIV mAbs directed against major targets of SIV Env vulnerability analogous to bnAb targets of HIV-1, namely the CD4bs, peptide epitopes of V1/V2 and V3 loops and potentially glycan targets of SIV Env (Fig 14). We did not isolate SIV mAbs targeting the V4 loop although it is possible that some of the SIV mAbs from the CVNbs sort may map to this region. Targeted isolation of V4-specific SIV mAbs could prove useful as the V4 region of SIV contains immunodominant epitopes and represents an early target for neutralizing mAbs [108, 109]. Compared with HIV-1 bnAb targets, we were unable to isolate quaternary-structure-preferring SIV-specific mAbs. While a pre-fusion SIV trimer structure has yet to be determined, it is likely that the SIV foldon trimer and 1JO8-scaffolded probes used for B cell sorting adopt an open quaternary conformation, analogous to HIV-1 soluble, uncleaved trimers [32, 110], which likely precluded us from isolating SIV quaternary-structure-preferring neutralizing mAbs. In fact, structural analysis of vaccine-induced HIV CD4bs-directed mAbs has revealed that despite high affinity binding to soluble Env ligands such as foldon trimers, such mAbs display a suboptimal angle of approach resulting in non-bnAbs with limited breadth and lack of neutralization activity against neutralization-resistant isolates such as JRFL [111]. Additional methods will be needed to isolate SIV bnAbs targeting quaternary epitopes, V1, V2 or V3 glycans, the immunodominant V4 region, the gp120-gp41 interface and MPER region.
The range of SIV mAbs isolated includes binding, non-neutralizing mAbs as well as strain-specific and cross-neutralizing mAbs exhibiting varying degrees of neutralization breadth and potency. We isolated SIV mAbs from SIV-vaccinated, pre-challenge as well as SIV-vaccinated and infected macaques. All but one of the SIV mAb epitopes identified from the latter showed high sequence similarity among SIV isolates tested making it difficult to determine whether individual mAb responses were elicited by the immunogen or the challenge virus Env. Only one of the SIV mAb epitopes identified (Env 142–15) was present in the SIVsmE660 challenge virus but not the SIVmac239 immunogen indicating that mAbs targeting this region were elicited following infection.
Comparison of the neutralization profiles for individual mAbs and corresponding serum samples for most animals revealed that the isolated mAbs recapitulated the breadth of serum neutralization in most cases (S9 Fig). In addition, the neutralization plateau effect reported for some HIV-1-specific mAbs [16] was also evident for neutralization curves of SIV mAbs against tier 2 isolates of both SIVmac251 and SIVsmE660, irrespective of mAb specificity. While differences in Env trimer glycosylation may explain incomplete neutralization by glycan-dependent mAbs [112], emerging data suggests conformational heterogeneity of Env trimers even within a clonal pseudovirus population may account for neutralization curve plateaus for glycan-independent mAbs [38, 113]. Other groups have also observed striking heterogeneity in neutralization sensitivities between SIV isolates [78, 82]. While the majority of clones within the well-characterized SIVsmE660 vaccine challenge stock are highly neutralization sensitive, approximately 10–25% exhibit an intermediate neutralization sensitivity phenotype and 10% are outright neutralization resistant [82]. Despite the broad range of epitopes targeted by SIV mAbs isolated and their capacity to bind SIVmac239 gp140 FT protein, none were able to neutralize the highly neutralization resistant SIVmac239. This was not wholly unexpected since sera from these animals also failed to neutralize SIVmac239 (Fig 12). This discrepancy between binding and neutralizing activity against a particular Env protein/virus has also been observed for HIV-1 CD4bs-directed mAbs and is thought to be result from inefficient recognition of cognate epitope due to quaternary packing conformational constraints in the context of functional, membrane-bound trimer despite high affinity binding to soluble Env ligands [111, 114, 115].
With the development of our novel competitive probe binding sort strategy and subsequent isolation and detailed characterization of nearly 70 SIV Env-specific mAbs we now have the necessary reagents with which to study immune and vaccine mediated correlates of protection in the SIV NHP challenge model of HIV-1. This includes testing of SIV bnAb passive immunization alone or as an adjunct to antiretroviral therapy (ART) by direct injection or gene therapy. As well, the binding, non-neutralizing SIV Env-specific mAbs identified here will serve as useful reagents for delineating the contribution of antibody-dependent cellular cytotoxicity (ADCC), antibody-dependent cell-mediated viral inhibition (ADCVI) and additional FcR-mediated activities toward control and/or prevention of HIV/SIV infection. Finally, use of additional probes and methods to isolate SIV-specific B cells will facilitate more thorough and rigorous pre-clinical evaluation of mAb-based immunotherapies for treatment and/or prevention of SIV infection in NHPs.
|
10.1371/journal.pcbi.1006129 | Effects of spatiotemporal HSV-2 lesion dynamics and antiviral treatment on the risk of HIV-1 acquisition | Patients with Herpes Simplex Virus-2 (HSV-2) infection face a significantly higher risk of contracting HIV-1. This is thought to be due to herpetic lesions serving as entry points for HIV-1 and tissue-resident CD4+ T cell counts increasing during HSV-2 lesional events. We have created a stochastic and spatial mathematical model describing the dynamics of HSV-2 infection and immune response in the genital mucosa. Using our model, we first study the dynamics of a developing HSV-2 lesion. We then use our model to quantify the risk of infection with HIV-1 following sexual exposure in HSV-2 positive women. Untreated, we find that HSV-2 infected women are up to 8.6 times more likely to acquire HIV-1 than healthy patients. However, when including the effects of the HSV-2 antiviral drug, pritelivir, the risk of HIV-1 infection is predicted to decrease by up to 35%, depending on drug dosage. We estimate the relative importance of decreased tissue damage versus decreased CD4+ cell presence in determining the effectiveness of pritelivir in reducing HIV-1 infection. Our results suggest that clinical trials should be performed to evaluate the effectiveness of pritelivir or similar agents in preventing HIV-1 infection in HSV-2 positive women.
| The risk of contracting HIV-1 is significantly higher in people who have genital HSV-2 infections. Here, we put forward a new mathematical model to describe HSV-2 infection and the process of HIV-1 infection in the genital mucosa surrounding HSV-2 lesions. We determine how the characteristics of HSV-2 infection affect the risk of HIV-1 infection, and determine whether reducing the severity of HSV-2 symptoms with antiviral drugs can be expected to decrease the risk of HIV-1 infection. We find that the risk of HIV-1 infection is dependent on three factors: the amount of HIV-1 the patient is exposed to, the severity of HSV-2 lesions, and the number of CD4+ T immune cells in the genital mucosa. Our model predicts that antiviral drugs targeting HSV-2 can cause a therapeutic decrease in lesion severity and CD4+ T cell count in the genital mucosa. This furthermore causes a significant decrease in the risk of HIV-1 infection but the dose of HSV-2 antiviral drug must be sufficiently high. Our results support further development and testing of new HSV-2 antiviral drugs to help decrease the world-wide burden of HIV-1.
| Herpes simplex virus-2 (HSV-2) is one of the most common sexually transmitted infections (STIs). Estimates from 2012 indicate that around 20 million people are newly infected by HSV-2 every year, with 11.3% of the human population infected [1]. While STIs often coincide, the establishment of human immunodeficiency virus 1 (HIV-1) in HSV-2 infected individuals is shockingly common. An estimated 38-60% of new HIV-1 infections in women and 8-49% of new HIV-1 infections in men may be attributable to HSV-2 infection due to the enhanced conditions a herpetic genital lesion presents for the entry and establishment of HIV-1 [2–4].
Genital HSV-2 lesions compromise the natural barrier of the skin and facilitate entry of HIV-1. In addition, the tissue surrounding a herpetic lesion is often rich in CD4+ T cells, the main target cell for HIV-1. These conditions may cause a 2 to 3-fold increase in the probability of HIV-1 infection establishment and drastically increase the spread of the HIV-1 epidemic [2, 5]. For this reason, it is important to understand the relationship between HSV-2 and HIV-1 infections and to find ways to decrease the risk of HSV-2 positive patients acquiring an HIV-1 infection.
The spread of HSV-2 usually occurs through skin-to-skin genital contact where the virus infects epithelial cells and replicates within them. Following this initial infection, the virus spreads to nearby neurons where it establishes latency in the dorsal roots of the neural ganglions [6]. This reservoir of HSV-2 in nerve tissue is protected from the immune system, leading to life-long infection. Viruses are slowly shed from the neurons and released back into the genital tract where they may spark a new productive epithelial infection, viral shedding, and ultimately transmission during sexual contact [7].
The development of a herpetic lesion in the epithelial tissue is largely dependent on the immune presence at that site [7]. Despite the high number of shedding episodes that occur in HSV-2 positive patients, the immune system rapidly responds, clearing small plaques of infection in two to twelve hours [8]. The two types of immune cells thought to be most important in HSV-2 infection control are the CD4+ and CD8+ T cells. Once the lesion is resolved, these immune cells also prevent re-infection, remaining at previous sites of infection for up to twenty weeks [9].
Cytotoxic CD8+ T cells are often thought of as the main effector cell population responsible for the control of HSV-2 in the infected epithelium [10]. As such, previous mathematical models of HSV-2 have included CD8+ T cells as the only immune cell present during HSV-2 infections [7, 9, 11–13]. However, CD4+ T cells have more recently been shown to be important in the control of HSV. In experiments where CD8+ T cell deficient mice were infected with HSV-1, the virus and lesions could still be cleared at genital and neural sites. However, in the alternative situation where CD4+ T cell deficient mice were infected with HSV-1, the infection could not be cleared [14]. CD4+ T cells are among the first immune cells to arrive at the site of infection, appearing within the first 48 hours of HSV-2 infection in the epithelium [15]. Once CD4+ T cells arrive, they release interferon gamma (IFN-γ) and other cytokines required for CD8+ T cell recruitment to the infection site [16]. Activated CD8+ T cells then kill infected cells by delivering perforin and activating apoptotic pathways [17]. Thus, CD4+ T cell dynamics should be included in mathematical models of HSV-2 infection to fully represent the system.
The influx of immune cells to the lesion site creates a favourable environment for HIV-1. Not only does HIV-1 have a greater probability of successfully traversing the damaged epithelial layer at the lesion site, but the immune response creates an environment dense in CD4+ T cells, the primary target cell of HIV-1 [2, 18]. Additionally, CD4+ T cells at lesions express high levels of chemokine receptor type 5 (CCR5), the co-receptor which HIV-1 most commonly uses to establish initial infection [4]. Ex-vivo studies have also shown a strong effect of HSV-2 infection on HIV-1 dynamics at herpetic lesions. In cervical tissue cultures infected with HSV-2, HIV-1 virions attached more frequently to sites containing HSV-2 infected cells than to sites containing uninfected epithelial cells, indicating that HIV-1 preferentially establishes infection at areas of HSV-2 infection [3].
HSV-2 and HIV-1 coinfection can also lead to higher transmission of HIV-1, with genital ulcers or microlesions shedding both HIV-1 and HSV-2 [19]. With 80% of HSV-2 viral shedding events occurring without visible lesions, this creates a potentially significant transmission scenario for both viruses [8]. Without proper knowledge about the state of their infection, infected individuals may remain asymptomatic and unknowingly infect sexual partners with either virus. Reciprocally, an asymptomatic HSV-2 positive individual may unknowingly be at an increased risk of HIV-1.
While no drug or vaccine has been developed that is capable of completely clearing or preventing HSV-2 infection, antivirals designed to decrease infection severity and outbreak frequency have long been available. Acyclovir and other related compounds work by inhibiting HSV-specific DNA polymerases and helicases, and result in less viral replication [20]. Acyclovir and its ester prodrug, valacyclovir, have been shown to reduce the occurrence of genital lesions by 47-75% and the rate of viral shedding by 80% [21, 22]. Reducing HSV-2 infection severity (number and severity of lesions, and immune cell presence) should be expected to reduce the risk of HIV-1 infection; however, clinical studies of HSV-2 positive individuals found that the incidence of HIV-1 infection was not reduced by (val)acyclovir treatment [21, 23, 24]. The reason for this discrepancy remains unknown. Some studies have argued that the doses given in these studies may simply have not been high enough [24]. Others point to the short half-lives of these drugs, which may be insufficient to completely suppress the extremely rapid kinetics of HSV-2 replication [13]. Determining whether other antivirals such as pritelivir, which has a much longer half-life and suppresses HSV-2 replication better than (val)acyclovir, may in fact decrease HIV-1 infection rates would have important global health implications. The use of mathematical models can provide insight into HSV-2 and HIV-1 infection dynamics in patients receiving antivirals, and may help to determine correct dosage amounts.
While a considerable amount of mathematical modelling has focussed on HIV-1 or HSV-2 infections independently [7, 9, 11–13, 25–28], none have analyzed the establishment of HIV-1 coinfection in individuals with chronic HSV-2 from a mechanistic and immunological perspective. Further, mathematical models have yet to be utilized to understand how antiviral drugs may decrease the risk of HIV-1 infection in persons infected with HSV-2. Here, we apply a spatial stochastic model to describe the dynamics of a genital herpetic lesion caused by HSV-2, and quantify the risk of HIV-1 acquisition based on an individual’s current state of HSV-2 infection. Using our model, we study the effects of HSV-2 antiviral drugs on HSV-2 outbreaks and the likelihood of HIV-1 infection following exposure. We also predict the dosage of antiviral drugs needed to achieve significant reductions in the HIV-1 infection probability. Our modelling predictions are consistent with previous work and can be tested by future studies on HIV-1 acquisition in treated and untreated HSV-2 patients.
We begin by creating a mathematical model to describe the general dynamics of a chronic HSV-2 infection in a small region of the genital mucosa. As herpetic lesions rarely reach diameters exceeding 6mm [7], we model a 2cm × 2cm region of the mucosa. We model the tissue to a depth of 74 μm, representing the average thickness of infectible epithelial tissue as measured by previous histological studies [7]. We further divide the model region into n × n equally sized cuboidal grid sites to spatially resolve lesion development. A schematic of the model is shown in Fig 1 and full details are given in the Methods.
We developed two versions of the model. In both, the important dynamics and interactions between HSV-2 free virus (V), infected and healthy epithelial cells (I and H respectively), and CD4+ and CD8+ immune cells (T and E respectively) are included. The two models differ based on assumptions around how the mucosal immune response is induced. The first, and simpler, version of the model assumes that immune cell proliferation is proportional to the number of infected epithelial cells (Fig 1B), following previous models of HSV-2 lesion dynamics [7, 9, 11–13]. The second version incorporates the effects of cytokines (C) produced by epithelial cells in response to infection (Fig 1C).
We modified our model to examine how the characteristics of an HSV-2 infection may affect the likelihood of HIV-1 acquisition in women exposed to HIV-1. Briefly, we added cell and viral populations and interactions that capture the basic processes: infection of CD4+ T cells by HIV-1, conversion of these cells into those which actively produce HIV-1, and the production and decay of extracellular virus. Our modified model is shown schematically in Fig 5 and described in full detail in the Methods.
HSV-2 infections are thought to increase HIV-1 infectivity in two ways: by creating an epithelial lesion that serves as a viral entry point, and by providing a simultaneous increase in the local number of CD4+ target cells. To determine how these factors affect the probability of HIV-1 infection following exposure, we chose to examine the scenario where the genital tissue of HSV-2 females is exposed to HIV-1 through sexual encounters with HIV-1 infected males. HSV-2 simulations were “paused” at different points during lesion development, creating spatially explicit simulated tissue samples (Fig 6A). We then introduced HIV-1 to each simulated tissue sample in an amount proportional to that found within the semen of HIV-1 infected males (3 × 103 – 3 × 105 virions/mL) and to the current amount of tissue damage within the sample. Using this initial condition, we simulated dynamics of HIV-1 infection and clearance, as described in the Methods. By running many HIV-1 simulations for each HSV-2 condition, we were able to determine the probability of HIV-1 infection given the current state of the female’s genital tissue (Fig 6B). We implicitly assumed that the dynamics of HIV-1 infection (or failure to infect) occur quickly on the timescale of the HSV-2 model. To test this assumption, we examined the time it takes for the infecting HIV-1 to either take hold, or go extinct. We found that in the vast majority of simulations, the HIV-1 dynamics were indeed very rapid (see S2 Fig), justifying our assumption.
We used this process to examine different simulated tissue samples. In particular, we focused on HIV-1 exposure to tissue that was healthy with no HSV-2 infection (1 sample), tissue exposed one week before the peak severity of a lesion (10 samples), tissue exposed at the time of peak severity (10 samples), tissue exposed during peak CD4+ T cell levels (10 samples), and tissue exposed one, two, or four weeks after the peak severity of a lesion (10 samples each). In total, this led to an analysis of 61 tissue samples. Properties of these simulated tissue samples are summarized in Table 1.
We plot the probability of HIV-1 infection for each of the 61 samples against the number of CD4+ T cells and the amount of tissue damage at the time of HIV-1 introduction as shown in Fig 6C. We do this for three possible seminal concentrations of HIV-1. By fitting linear planes to the data, the probability of HIV-1 infection can be expressed as a simple function of the CD4+ T cell concentration and the amount of tissue damage within the simulation region at the time of exposure.
When a simulation region is exposed to seminal HIV-1 concentrations of 3 × 103 virions/mL, characteristic of chronic HIV-1 infection, HIV-1 infection probability within the simulation region can be expressed as
Prob(infection in sim region) ≈ 4 . 8 × 10 - 4 T + 0 . 033 L - 0 . 014 (1)
where T represents the CD4+ T cell concentration (cells/mm2) and L represents the percentage of damaged tissue within the simulation region. The corresponding linear approximations for the probability of infection following exposure to seminal HIV-1 concentrations of 3 × 104 virions/mL and 3 × 105 virions/mL were found to be
Prob(infection in sim region) ≈ 5 . 6 × 10 - 3 T + 0 . 29 L - 0 . 21 (2)
and
Prob(infection in sim region) ≈ 4 . 4 × 10 - 2 T + 1 . 65 L - 1 . 41 (3)
respectively. Note that the numerical values in these equations are dimensional, with appropriate units so that each overall probability is dimensionless. In each case, we constrain the fit so that the function passes through the point representing the HIV-1 infection probability in healthy tissue predicted by the model (0.005%, 0.02% and 0.34% respectively). We note that in all scenarios, increases in either CD4+ T cell count or tissue damage increase the risk of HIV-1 acquisition.
A strong reason to use a spatial model when estimating the risk of HIV-1 infection in HSV-2-affected tissue is that there is likely to be a correlation between the location of tissue damage and the location of CD4+ T cells. HIV-1 is expected to penetrate into regions of greater tissue damage with higher probability, where it can encounter high CD4+ T cell concentrations, thus increasing the probability of successful infection above what might be expected if this spatial correlation is not considered. An artificial example of how the distribution of tissue damage and CD4+ T cell count within a simulation region can change the probability of HIV-1 infection is shown in S3 Fig.
To further investigate the importance of spatial effects, we estimated the risk of HIV-1 acquisition using a non-spatial version of the model. When we remove the spatial aspect, we effectively assume that tissue damage and CD4+ T cell counts are uniform across the simulation region, removing possible hot spots for infection. To perform these simulations, we introduced HIV-1 into the same simulated tissue samples as examined above, but after pooling all sub-volume data. We then fit the risk of HIV-1 infection as a function of CD4+ cell count and the percentage of tissue damage, to produce equations similar to those given above (Eqs 1–3). In the first two scenarios, corresponding to exposure to semen concentrations of 3 × 103 and 3 × 104 virions/mL, the fits did not change substantially (changes in linear fit coefficients of less than 20%). However, in the third scenario, (semen HIV-1 concentration of 3 × 105 virions/mL), the loss of spatial effects caused a large decrease in the estimate of HIV-1 risk. This was due to the (imperfect) spatial correlation between the location of CD4+ T cells and damaged tissue, driven by the delay in CD4+ T cell response as noted in Fig 2.
As the results so far only consider a single (4cm2) patch of genital epithelium, we next calculated the risk of HIV-1 over the entire genital tract. The female genital tract has a surface area of approximately 88 cm2 [30], corresponding to 22 of our simulation regions. Results taken from 50 simulations each of length 12 months were combined to obtain the distributions of CD4+ counts and levels of tissue damage at sites surrounding HSV-2 infected neurons (Fig 7A). We sampled from these distributions, to create a set of genital tract profiles (each composed of 22 simulation regions) for HSV-2 infected patients, and then calculated the probability of infection at each site in the genital tract profile using Eqs 1–3. These risks were then combined to calculate the risk of HIV-1 infection per sexual act when exposed to different per-ejaculate HIV-1 viral loads. As expected, high seminal HIV-1 viral loads lead to higher probability of infection. Since an HSV-2 infected individual will have some regions of healthy tissue and others with infected neurons releasing HSV-2, not all 22 regions were considered lesioned, but instead some were assumed to simply possess the characteristics of healthy tissue. Similarly, some individuals will have more infected regions than others, thus we varied the number of simulation regions assumed to have neurons dripping HSV-2 into the system. We defined three severities of HSV-2 infection as follows: Mild, moderate and severe infections were respectively defined to have 1, 2, and 3 simulation regions (out of the 22 regions) where HSV-2 is actively being introduced into the system. The severity of HSV-2 infection is found to significantly alter the risk of HIV-1 infection after exposure (Fig 7B). When weighing how tissue damage and CD4+ cell counts each contribute to this risk, CD4+ cells are the main contributor. Lesions are sporadic and their durations are relatively short; this means they only contribute to HIV-1 risk a fraction of the time. CD4+ T cell counts remain elevated for much longer, indicating that for the majority of time the density of CD4+ T cells in the mucosa is an important driver of HIV-1 acquisition.
Pritelivir is a recently developed antiviral for HSV-2 treatment. It acts as a viral helicase-primase inhibitor, preventing viral replication, and is currently in clinical trials. We examined how doses of 10, 30, 55, and 80 mg/day would be predicted to affect the characteristics of HSV-2 infection and the corresponding risk of contracting HIV-1. These doses all fall within the range of doses from recent drug trials [31]. In Fig 8A–8C we show distributions of HSV-2, CD4+ T cell, and total tissue damage within one simulation region, obtained from fifty one-year-long simulations at each drug dose. As the pritelivir dose increases, the number of HSV-2 virions and amount of tissue damage associated with a lesion decreases accordingly. The median CD4+ T cell count, however, is relatively unaffected by these particular doses of pritelivir, remaining at more than double the amount seen in healthy tissue even at the highest examined pritelivir dose [4]. As the number of tissue-resident CD4+ T cells is thought to affect the risk of HIV-1 acquisition, this result indicates a deficiency of HSV-2 antiviral strategies in preventing HIV-1 infection.
To further investigate this hypothesis, we used our simulated data on CD4+ T cell densities and epithelial tissue damage to calculate the overall risk of HIV-1 infection when the entire female genital tract is exposed to HIV-1 in semen (Fig 8). We again examined scenarios where patients have different severities of HSV-2 infection determined by the number of sites in the genital epithelium (1, 2, or 3) where HSV-2 is actively being introduced by infected neurons.
For a female with completely healthy tissue, the per-sexual-act probability of HIV-1 infection when exposed to per-ejaculate HIV-1 concentrations of 3 × 103, 3 × 104, or 3 × 105 virions/mL was calculated to be 0.11%, 0.44% and 7.41% respectively. These percentages are indicated by the red horizontal lines in Fig 8B and match values reported in the literature [5, 32–36]. In patients not receiving antivirals, studies have recorded HIV-1 infection risk to be 2-3 times higher in asymptomatic HSV-2 infection, compared to healthy patients, and 7 times higher if lesions are present [2, 37]. Our simulations were in general agreement with these observations. The smallest median increase in HIV-1 risk due to HSV-2 infection was 1.2 fold (1.0-9.1 fold from 5th to 95th percentile), seen in simulations where HSV-2 infection was considered mild and tissue was exposed to the lowest HIV-1 semen concentration. The highest median increase in HIV-1 risk due to HSV-2 infection was 8.6 fold (1.4-39.6 fold), seen in simulations where HSV-2 infection was considered severe and tissue was exposed to an intermediate HIV-1 semen concentration.
When simulations included pritelivir treatment, the risk of HIV-1 infection was significantly decreased (p<0.01) at some pritelivir doses (Fig 8B). In some cases, 30 mg/day was enough to cause a significant decrease in HIV-1 infection risk, and pritelivir doses of 80 mg/day showed significant decrease in risk for all scenarios examined. Despite this, none of the pritelivir doses were able to return the risk of HIV-1 infection to the baseline values seen in healthy tissue. Even at the highest dose of pritelivir, the median risk of HIV-1 contraction at best remained 1.2-fold higher (1.0-6.1 fold from 5th to 95th percentile) than that seen in healthy patient simulations (occurring when HSV-2 infection was mild and HIV-1 semen concentrations were low) and at worst remained a median of 5.6-fold higher (1.4-34.2 fold) than that seen in healthy patient simulations (occurring when HSV-2 infection was severe and HIV-1 semen concentrations were moderate).
By developing a spatial stochastic model to describe the dynamics of chronic HSV-2 infections in the genital mucosa, we determined the characteristics of infection essential for describing the growth and control of HSV-2 lesions, and how these traits can be directly used to predict the risk of HIV-1 infection.
We found that immune cells alone were not able to control the spatial spread of lesions. This was due to their slow mobility within the tissue and their inability to keep up with the spread of infection. As we included in our preferred model, rapidly diffusing cytokines were needed to stimulate the recruitment and proliferation of immune cells to control the spread of infection. Other responses to cytokines, such as the antiviral state uninfected cells can enter, may also play an important role in controlling lesion growth, as suggested in previous models of viral dynamics [38]. We did not include these effects in the present work, but they may prove important. Future work may include examining the roles of each of these factors in controlling simulated lesion development.
Using our model, we also determined how the risk of HIV-1 infection in women through vaginal intercourse depends on the extent of genital HSV-2 infection. While it has long been accepted that HSV-2 infection increases the risk of HIV-1 acquisition, the relationship has not been quantified in detail. We have argued that the risk of HIV-1 infection is controlled by the local concentration of CD4+ T cells and the degree of damage caused by HSV-2 within the genital tissue. The severity of HSV-2 infection is often measured by the viral load in genital swabs; however, our model indicates that HSV-2 viral load is only weakly correlated with HIV-1 risk. When HSV-2 viral loads peak due to a new lesion, there is a delay before tissue damage and CD4+ T cell counts peak due to the nature of the tissue’s response to infection. This means that HSV-2 viral load is not an effective metric of HIV-1 susceptibility.
As HSV-2 antivirals decrease the severity of HSV-2 infection, we wanted to quantify how these effects might decrease the risk of HIV-1 acquisition. With a particular focus on the newly developed HSV-2 antiviral drug pritelivir, we showed that certain doses may decrease the risk factors linked with HIV-1 infection susceptibility, namely CD4+ T cell count and HSV-2 lesion severity. These decreases led to a significant overall decrease in HIV-1 acquisition risk, as predicted by our simulations. While such a decrease in risk from HSV-2 antiviral drugs has been predicted, clinical studies have failed to detect the effect [23, 24].
There are several possible reasons for this disconnect between theory and data. In clinical studies observing the effect of HSV-2 antivirals on HIV-1 infection probability, many participants were co-infected with other STIs and this effect was not controlled for [23, 24]. Non-ulcerative STIs can increase CD4+ T cell levels to twice those seen in healthy tissue [39]. Our results indicate that CD4+ T cell counts in the genital mucosa are the greatest determinant of HIV-1 infection risk, meaning higher CD4+ T cell counts due to other STIs may have kept CD4+ T cell counts too high for significant reductions in HIV-1 infection risk.
Another reason for this disconnect may be that clinical studies aimed at detecting the effects of HSV-2 antivirals on HIV-1 infection used acyclovir as their study drug. Acyclovir has rapid pharmacokinetics with a short half-life of 3-4 hours [23, 24]. Previous mathematical models of HSV-2 infection that included the effects of antiviral drug decay show that when drug concentrations reach sub-therapeutic levels, rapid HSV-2 breakouts can occur, preventing full control of the infection [13]. Lack of HIV-1 risk reduction observed in these acyclovir studies may thus be explained by sub-therapeutic drug levels. Pritelivir has a half-life estimated to be approximately 80 hours [13] which should reduce the effects of drug decay and allow for better control of HSV-2 infection breakouts and HIV-1 infection probability.
In future work, it may be also interesting to examine how the effects of HIV-1 antibodies or HIV pre-exposure prophylaxis (PrEP) could change the risk of HIV-1 contraction in HSV-2 patients. A recent study has examined how antibody presence and efficacy may affect the transmission of HIV-1 to healthy individuals [35]. Calibrating this model to include the effects of HSV-2 infection may provide interesting insights as further preventative measures against HIV-1 are developed.
As with all mathematical modelling work, our study has some limitations. A primary limitation is related to the role of CD4+ T cells in controlling HSV-2. We incorporated CD4+ T cells in our model as an essential part of studying HIV-1 infection. However, the direct effect of CD4+ T cells on the control of HSV-2 in the mucosa remains unclear. As more experimental information on the role of CD4+ T cells becomes available, it will be important to revise and improve our model. New developments also may support further mathematical modelling of the synergy between HSV-2 and HIV-1 from the immunological perspective.
In conclusion, our work provides insight to herpes lesion development in the genital mucosa and forms a logical framework for studying the synergy between HSV-2 and HIV-1 infections. Our results support the use of pritelivir or other HSV-2 antivirals as an effective means of preventing HIV-1 infection in patients infected with HSV-2, and highlight the essential mechanisms by which pritelivir reduces the risk of HIV-1. While clinical studies have been performed to examine the effects of pritelivir on HSV-2 infection alone, none have yet examined its effect on the risk of HIV-1 infection. Furthermore, the modelling framework laid out here to examine the spatial development of lesions in the genital epithelium could potentially be expanded to mathematical studies of other skin abnormalities, building on experimental knowledge of tissue-resident immune cells and their interactions with invading pathogens or (pre-)cancerous cells.
We focus on a 2cm × 2cm × 74μm volume of genital epithelium. This is modelled as a single-layer n × n array of cuboidal subvolumes. The assumptions of the model (described in turn below and summarized in Fig 1) are used to form a chemical master equation that describes the dynamics of all possible reactions in the system. Following Gillespie’s direct simulation algorithm, modified to allow for spatially separated sites, the model is executed to produce one simulation [40, 41]. We assume that reactions only occur between cells, virions, and cytokines that exist within the same grid site, while motion of an object from one location to another is modelled as a distinct reaction (with fixed rate).
In the absence of infection, suppose each subvolume contains H0(n) healthy epithelial cells. Taking the diameter of a healthy epithelial cell to be ∼17 μm and assuming a roughly cuboidal cell shape gives approximately 6 × 106 cells inside the entire model region [7]. We therefore set H0(n) = 6 × 106/n2. Defining the actual number of epithelial cells in a given subvolume (indexed by i and j) as Hi, j, we model tissue repair as a simple exponential return to the healthy state with rate g. This is described as the state-dependent reaction
H i , j → H i , j + 1 with rate g ( H 0 ( n ) - H i , j ) . (4)
HSV-2 infections are initiated by the release of virus from neurons innervating the genital mucosa, which occurs at a constant rate ϕi, j. The model region is assumed to be centered around an HSV-2 infected neuron meaning ϕi, j = 0 for all grid sites except for the center site (ic, jc). Healthy epithelial cells become infected by HSV-2 virus following the law of mass action at a rate proportional to β(n), and new virus is produced by these infected cells at a per-capita rate p. Free HSV-2 virus decays at a per-capita rate c. These events are described by the following set of reactions:
V i c , j c → V i c , j c + 1 with rate ϕ i c , j c (5)
{ V i , j → V i , j - 1 , H i , j → H i , j - 1 , I i , j → I i , j + 1 } with rate β ( n ) H i , j V i , j (6)
V i , j → V i , j + 1 with rate p I i , j (7)
V i , j → V i , j - 1 with rate c V i , j (8)
Modelled CD8+ effector T cells (Ei, j) are taken to be HSV-2 specific and interact and kill infected epithelial cells at a rate f(n), again following the law of mass action. Infected cells also die at a per-capita rate a as they succumb to the infection. CD8+ T cells exit the system at a small rate δ in keeping with the long period these cells remain at and protect previous sites of infection [4, 14]. In the infection-free state, a population of CD4+ T cells remain at the site at a per-site equilibrium number of λ(n)/d. While CD4+ T cells have not been included in previous models of HSV-2 infection [7, 9, 11–13], we include their dynamics due to their importance in the establishment of HIV-1 infections and in order to get a better representation of the full dynamics occurring at herpes lesion sites. These events are described by the following set of reactions:
I i , j → I i , j - 1 with rate ( f ( n ) E i , j + a ) I i , j (9)
E i , j → E i , j - 1 with rate δ E i , j (10)
T i , j → T i , j + 1 with rate λ ( n ) (11)
T i , j → T i , j - 1 with rate d T i , j (12)
The presence of infection stimulates the proliferation of CD8+ T cells and CD4+ T cells in the genital mucosa. We analyze two descriptions of this stimulation. Letting X1 and X2 represent the per-capita production rate of CD8+ T cells and CD4+ T cells in response to infection respectively, these descriptions can be written as
Model I
X 1 = I i , j I i , j + r 1 ( n ) θ 1 , (13)
X 2 = I i , j I i , j + r 2 ( n ) θ 2 . (14) Model II
X 1 = C i , j I i , j + r 3 ( n ) θ 1 , (15)
X 2 = C i , j I i , j + r 4 ( n ) θ 2 . (16)
In Model I, immune cell proliferation is dependent on the number of infected cells (Ii, j) while in Model II immune cell proliferation is dependent on cytokine concentrations (Ci, j) produced by infected cells. In both cases, immune cell proliferation saturates to a maximum per-capita proliferation rate (θ1 and θ2 for CD8+ and CD4+ T cells respectively) as the number of infected cells or cytokines increases. ri(n) determines when proliferation reaches half its maximum rate where i ∈ {1, 2, 3, 4}. These proliferation rates are then encoded as reaction rates for the reactions Ei, j → Ei, j + 1 and Ti, j → Ti, j + 1 respectively.
With cytokines included in Model II, we must also include their dynamics. Cytokines are assumed to be produced at a rate b, dependent on the number of infected epithelial cells in the same grid site, and decay from the system at a per-capita rate m. This corresponds to the following reactions:
C i , j → C i , j + 1 with rate b I i , j (17)
C i , j → C i , j - 1 with rate m C i , j (18)
We now specify the model of mobility. Since immune cells actively move around the epithelium in search of infection, and viruses and cytokines can passively diffuse through their environment, we allow HSV-2 virus, cytokines, CD4+ T cells, and CD8+ T cells to move horizontally through the tissue into neighbouring grid sites. We define the rates of motion for a specific diffusing body N ∈ {E, T, V, C} at grid site (i, j), where i, j ∈ {1, 2, …, n}, as
D N i , j = ω N N i , j - 1 + N i - 1 , j - 4 N i , j + N i + 1 , j + N i , j + 1 h 2 . (19)
Here, ωN represents the motion coefficient specific to the moving body and h2 represents the horizontal cross sectional area of each grid site. These motion rates are converted into reaction rates and considered alongside all other reactions in the Gillespie simulation scheme.
At the boundaries of the simulation region, virus, cytokines, CD8+ T cells, and CD4+ T cells can exit the model region. However, as immune cells are also present in surrounding tissue, we suppose that on average there is a constant flux of CD8+ and CD4+ T cells entering the model region through its boundaries. We take Eave and Tave to represent the specific CD8+ and CD4+ T cell numbers expected to exist in chronic HSV-2 infected tissue, per simulation region. Then Eave/n2 and Tave/n2 represent the average CD8+ and CD4+ T cell concentrations per grid site on an n × n grid. Influx of CD8+ T cells and CD4+ T cells into boundary grid sites therefore occurs at rates ωEEave/n2 and ωTTave/n2 respectively. Virus and cytokines are not allowed to enter from the boundary but can exit outwards, reflecting our assumption that no other lesions are in close proximity to the modelled patch.
In Model II, we suppose that immune cells preferentially migrate up cytokine gradients [16]. To reflect this assumption, we make the direction of immune cell motion dependent on the quantity of cytokine present in neighbouring grid. The probability of immune cell motion into each of the neighbouring regions is given by
Prob ( motion into ( i , j ) ) = C i , j C t o t , (20)
with (i, j) taking on the indices of the four surrounding grid sites and Ctot being the total quantity of cytokines in these four sites. When cytokines are not present, immune cells diffuse into their four neighbouring sites with equal probability.
We evaluate the effective tissue damage at site (i, j) as the fraction of epithelial cells removed from a particular subregion, relative to healthy uninfected tissue. This value is defined as
L i , j = H 0 ( n ) - H i , j - I i , j H 0 ( n ) , (21)
This formula can also be summed to find the fraction of tissue damage within the entire simulation region:
L tot = H 0 ( 1 ) - ∑ i ∑ j H i , j - ∑ i ∑ j I i , j H 0 ( 1 ) . (22)
This measure of tissue damage is used to assess the probability of HIV ingress to the simulation region, as described below.
A schematic summary of the model extended to study HIV-1 infection dynamics is shown in Fig 5. HIV-1 (Pi, j) propagates by infecting CD4+ T cells. This is assumed to occur through mass action at a rate k(n). Once a CD4+ cell becomes infected, it moves into the T1 class, representing latently infected cells known to be in the eclipse (not yet producing HIV) phase. These cells mature into actively infected cells (T2) at a per-capita rate η. Actively infected CD4+ cells produce HIV-1 virus at a per-capita rate ψ; this virus decays at per-capita rate ℓ. T1 and T2 cells may be cleared from the system at a per-capita rates g1 and g2 respectively. The model includes no specific immune response against HIV-1-infected CD4+ cells as we assume no immunity against HIV-1 infection yet exists at the earliest stage of infection. These assumptions regarding HIV infection are encoded as reactions as follows:
{ Ti,j→Ti,j−1,Pi,j→Pi,j−1,T1i,j→T1i,j+1 }withratek(n)Ti,jPi,j (23)
{ T 1 i , j → T 1 i , j - 1 , T 2 i , j → T 2 i , j + 1 } with rate η T 1 i , j (24)
T 1 i , j → T 1 i , j - 1 with rate g 1 T 1 i , j (25)
P i , j → P i , j + 1 with rate ψ T 2 i , j (26)
P i , j → P i , j - 1 with rate ℓ P i , j (27)
T 2 i , j → T 2 i , j - 1 with rate g 2 T 2 i , j (28)
HIV-1 motion is assumed to occur at rate D P i , j with diffusion coefficient ωP between neighbouring subvolumes (Eq 19 with N = P).
To represent exposure to the HIV-1 virus, simulations were stopped at various points and HIV-1 was introduced into the system. To minimize computational time but still gather the data necessary to calculate HIV-1 infection probability, we tracked only the dynamics related to HIV-1 infection establishment once HIV-1 virus was introduced into the simulation region. By making this simplification, we assume that the changes in HSV-2 infection dynamics have a minimal effect on CD4+ T cell count during the small time window during which HIV-1 infection establishment occurs. This assumption was examined thoroughly by tracking the length of time it took for the infection process to reach completion (viral extinction, or the establishment of eight HIV-1 infected cells). See S2 Fig.
We chose to model the scenario where HSV-2 infected females are exposed to HIV-1 during vaginal sex with HIV-1 infected male partners. Recent ex-vivo experiments indicate that 0.24% of HIV-1 virions in semen penetrate the female’s genital mucosa [30]. We examined exposure to various seminal HIV-1 concentrations representative of chronic (3.0 × 103 HIV-1 virions/mL), moderate (3.0 × 104 HIV-1 virions/mL), and acute (3.0 × 105 HIV-1 virions/mL) infections [30, 35]. Assuming that the average semen volume per ejacuate is 3 mL [36], and the vaginal surface area is approximately 88 cm2 [30], approximately 1, 10, or 100 HIV-1 viruses would enter our simulation region of 4 cm2 after a sexual encounter with a chronically, moderately, or acutely HIV-1 infected partner (respectively) when the female’s genital tissue is healthy. Carias et al. also found that 10 times as many viruses were able to penetrate the epithelium if weak cell junctions were present compared to tissue without weak cell junctions [30]. Assuming the effects of tissue damage are similar, we used this estimate as a conservative representation of the increased viral entry at the site of lesions. We suppose that if no lesion is present in the model region (Li, j = 0 ∀(i, j)), then 1, 10, or 100 HIV-1 virions per simulation region enter at the time of exposure depending on the viral load in semen. These counts increase linearly with tissue damage until 10 times more virus enters the tissue if all tissue in the simulation region is damaged (Li, j = 1 ∀(i, j)). We chose a linear relationship to minimize the complexity of the model as the exact relationship between tissue damage and virion entry remains unknown. The entry points of the virus are also made dependent on site-specific tissue damage with the virus being randomly distributed among the grid sites following a multinomial distribution with the probability of a virus being distributed to site (i, j) given by the fraction of tissue damage found at that site.
Once HIV-1 virions have been placed within the simulation region, we use Gillespie’s algorithm to calculate the infection probability for each subvolume in our computational domain. The simulation is stopped once the infection has gone extinct or propagated enough to imply infection establishment. Following previous work of Pearson et al., we assume that infection is effectively established once the simulation region has at least 8 infected cells [28]. Simulations for each subvolume were repeated 10,000 times. From these simulations we define the probability of HIV-1 infection as the fraction of successful infections. Finally, after running simulations for each subvolume, we combine the probabilities of infection to achieve an overall probability of HIV-1 infection for the entire simulation region:
Prob ( infection in sim . region ) = 1 − ∏ i ∏ j [ 1 − Prob ( infection at site ( i , j ) ) ] . (29)
To account for the effects of the HSV-2 antiviral drug pritelivir in the mathematical model designed to describe chronic HSV-2 infection in the genital mucosa, we include a new parameter ζ to describe the effectiveness of the antiviral in suppressing HSV-2 replication. Here, ζ can take values in [0, 1] with 0 representing an antiviral dosage that has no effect on HSV-2 replication and 1 representing complete suppression of replication. As pritelivir is a helicase-primase inhibitor that suppresses viral replication, we assume this suppression decreases both the amount of virus that enters the simulation region from the neurons, and the amount of virus an infected epithelial cell produces; this replaces rates ϕ and p in the model with ϕ(1 − ζ) and p(1 − ζ) respectively.
We analyzed four potential values for ζ (ζ = 0.15, ζ = 0.5, ζ = 0.7, ζ = 0.85) and studied the resulting effects on lesion dynamics. Previous analysis predicts these ζ values correspond to doses of 10, 30, 55, and 80 mg of pritelivir a day [13]. These amounts all fall within the range of pritelivir doses given to patients in recent drug trials [31]. While other models have included the pharmacokinetics and pharmacodynamics of the drug [13], we more simply assume patients have a constant dosage within their genital mucosal tissue. This assumption can be considered acceptable due to pritelivir’s long 80-hour half-life which keeps drug conditions within the body relatively constant [13].
To determine HIV-1 infection probability for the entire genital region, we assume that 22 of our 4cm2 simulation regions can be used to represent the entire area of the genital region, estimated to be 88 cm2 [30]. Within the genital tract of an HSV-2 infected individual, there may be multiple sites where different neurons are releasing HSV-2 into the genital tissue. Depending on the severity of HSV-2 infection, some of the 22 sites composing the entire genital tract will have dynamics given by our simulations while others can be considered “healthy”. We examined a range of different HSV-2 infection severities, varying the average number of sites centered around a single HSV-2 infected neuron. We chose to look at cases where patients have an average of 1, 2, or 3 sites of infection. This corresponds to HSV-2 neuronal release rates of 50, 100 and 150 virions/day/genital tract which have previously been found plausible [42].
We also incorporated the idea that sites of neuronal drip can move to different spots in the genital tract; at any given time, each of the 22 patches has a 1/22, 2/22, or 3/22 probability of being the site of neuronal drip, depending on the HSV-2 infection severity being examined. If acting as a site of neuronal drip, the state of that site was assumed to be described by a randomly chosen time point from one of the 50 full simulations of our model. Using the CD4+ T cell and tissue damage characteristics of each site, we determined the probability of HIV-1 infection at each of the 22 sites and then combined these using the equation
Prob(infection in vagina) = 1 - ∏ i = 1 22 [ 1 - Prob(infection in sim. region i ) ] (30)
to get the overall probability of HIV-1 infection per sexual act. This process was repeated 1000 times for each drug dose to produce distributions in infection probabilities.
Due to recent interest in mathematically analyzing HSV-2 infections in the genital mucosa, many of the dynamics describing HSV-2 infection are well parameterized [7, 9, 11–13]. However, CD4+ T cell dynamics at the lesion site have not previously been examined from a mathematical standpoint. We therefore determined new values for the parameters governing CD4+ T cell behaviour in the genital epithelium. One challenge surrounding this task was the limited data regarding CD4+ T cell numbers in the genital mucosa. However, as CD4+ T cells are the main target of the HIV-1 virus, a representation of their dynamics is essential if we want to address questions related to HIV-1 infection. Fortunately, recent studies on the immune presence in the genital mucosa during different stages of herpetic lesion development reported both CD4+ and CD8+ T cell numbers at the lesion sites [4, 43].
A healthy individual without HSV-2 infection has approximately 68 CD4+ T cells per mm2 circulating around the epidermal layer of the genital epithelium [4]. Scaling this number to the 4 cm2 of our computational region indicates there should be 27200 CD4+ cells in the model when the patient is infection-free, directly corresponding to the fraction λ/d. Assuming the death rate, d, of CD4+ T cells is similar to that of CD8+ T cells, we set d = 0.07/day and λ = 1900/region-day to achieve the correct infection-free equilibrium value. With these two parameter values chosen, we can choose the parameters found in the expression for X2 (above), the rate of CD4+ T cell response to infection. A striking feature appearing in the experimental reports is the relatively unchanging ratio between CD4+ and CD8+ T cells [4, 43]. During the progression of an HSV-2 genital lesion, the CD4+ to CD8+ T cell ratio remains fairly constant, ranging from approximately 0.6 to 2.0 with a mean value of 1.06 in healthy tissue and 1.24 in HSV-infected tissue [4, 43]. As parameters describing CD8+ T cell dynamics are already known, we simply varied the parameters describing X2 until we reached a state where the CD4+ to CD8+ T cell ratio consistently fell within an appropriate range. In running fifty, one-year simulations of the full model including cytokines, with parameters of r3 = 42/day, r4 = 38/day, θ1 = 1.70/day, and θ2 = 1.40/day, the CD4+ to CD8+ ratio had an average of 1.4, ranging from 0.7-3.1. Since these ratios are similar to those observed experimentally, we used them for all simulations.
All parameter values of the model are recorded in Table 2.
Many mathematical models have examined initial HIV-1 infection. However, few explicitly examined the dynamics occurring within the genital tissue, where HIV-1 infection is most often acquired. As HIV-1 and immune cell counts usually come from blood or plasma samples, the parameters of most models are fit to these numbers [25, 27, 28]. While dynamics occurring within the blood and epithelium may be similar, they may not be occurring at the same rates. While we were able to choose some parameter values based on those used in previous models, others were estimated based on our limited knowledge of infection behaviour in the female genital tract. This is with specific reference to the estimates for parameters ψ and ℓ. As only within-blood estimates have been recorded in the literature for these parameters, we chose values that corresponded with those fit for HSV-2 dynamics. Here, ψ matches with the rate of HSV-2 production and ℓ matches with the value for HSV-2 clearance. All HIV-1 parameters used in the model are listed in Table 3.
|
10.1371/journal.ppat.1005319 | Extracellular Histones Induce Chemokine Production in Whole Blood Ex Vivo and Leukocyte Recruitment In Vivo | The innate immune system relies to a great deal on the interaction of pattern recognition receptors with pathogen- or damage-associated molecular pattern molecules. Extracellular histones belong to the latter group and their release has been described to contribute to the induction of systemic inflammatory reactions. However, little is known about their functions in the early immune response to an invading pathogen. Here we show that extracellular histones specifically target monocytes in human blood and this evokes the mobilization of the chemotactic chemokines CXCL9 and CXCL10 from these cells. The chemokine induction involves the toll-like receptor 4/myeloid differentiation factor 2 complex on monocytes, and is under the control of interferon-γ. Consequently, subcutaneous challenge with extracellular histones results in elevated levels of CXCL10 in a murine air pouch model and an influx of leukocytes to the site of injection in a TLR4 dependent manner. When analyzing tissue biopsies from patients with necrotizing fasciitis caused by Streptococcus pyogenes, extracellular histone H4 and CXCL10 are immunostained in necrotic, but not healthy tissue. Collectively, these results show for the first time that extracellular histones have an important function as chemoattractants as their local release triggers the recruitment of immune cells to the site of infection.
| The detrimental effects of extracellular histones under pathological conditions have lately attracted considerable attention. However, little is known about their functions as damage-associated molecular pattern molecules. Our study shows for the first time that extracellular histones trigger the induction of chemotactic chemokines from monocytes. As this interaction is dependent on a pattern recognition receptor, namely toll-like receptor 4, our data indeed point to an important role of extracellular histones in danger signaling. Notably, CXCL9 and CXCL10 are chemoattractants, and the recruitment of immune cells to the site of histone injection in a subcutaneous mouse model supports the concept that low levels of extracellular histones constitute a part of the host response.
| The rapid response to infections or tissue injury is one of the main features of the innate immune system. To this end pattern recognition receptors (PRRs) play an important role in these processes. PRRs can be divided into different groups or classes such as toll-like receptors (TLRs), C-type lectin receptors, scavenger receptors, and complement receptors [1–3]. Many PRRs are exposed on the host immune cells and they are evolutionary developed to target conserved non-self patterns (pathogen-associated molecular patterns, PAMPs). However, apart from sensing microbial signatures, PRRs can also bind and respond to host derived danger signals [1–3], also referred to as damage-associated molecular patterns (DAMPs). Upon binding to PRRs, DAMPs can trigger several signaling pathways which in turn can lead to a production of pro-inflammatory mediators, including cytokines, chemokines, and vasoactive peptides [4–6]. DAMPs thereby act as adjuvants that further increase the host response in addition to the initial cause of inflammation. One important group of DAMPs consists of the extracellular histones. Histones are ubiquitous proteins that are mainly involved in organizing DNA into nucleosomes and chromatin. Therefore they normally have only an intracellular function and are not actively released into the extracellular environment. However, during inflammatory and/or necrotic conditions histones can be mobilized from stressed, damaged, or dying cells. Such complications are for instance seen in patients with malignant tumors, trauma associated lung injuries, malaria, or severe infectious diseases [7–10].
Several studies have reported that intravenous injection of high histone doses (10–50 mg/kg) causes severe thrombocytopenia and tissue damage in mice and at even higher concentrations (75 mg/kg) death within minutes [11, 12]. Subsequent in vitro and in vivo experiments have shown that at such concentrations, extracellular histones can evoke an aggregation of platelets, a formation of thrombi, exposure of phosphatidylserine on erythrocytes, and cell necrosis [11–18]. (For a review [19]). By employing knockout animals (TLR2, TLR4, TLR9, and MyD88), extracellular histones have also been found to induce release of pro-inflammatory cytokines in mice (interleukin-6 (IL-6), IL-8, and tumor necrosis factor-α (TNF-α)). Notably, extracellular histones can be found in complex with DNA which has been reported to enhance their immunostimulatory and immunogenic properties [20, 21]. Though these results clearly point to the involvement of toll-like receptors, a characterization of the interaction between histones and TLRs at protein chemical level has not been described. Further, it has not been reported whether histones are able to induce the release of other mediators (for instance substances with chemotactic activities).
The present study was undertaken to analyze a potential role of extracellular histones as sentinels in innate immunity. We show that histone H4 binds directly to the TLR4/myeloid differentiation factor 2 (MD-2) complex and that extracellular histones specifically target monocytes in human peripheral blood. As a consequence, monocytes release the non-ELR CXC chemokines CXCL9 and CXCL10, respectively. In addition, in vivo experiments show that this leads to a recruitment of leukocytes. Together our findings implicate an important role of extracellular histones in evoking the innate immune system by sensing danger and damage signals without causing harmful effects for the host.
In the first series of experiments, we wished to study the role of extracellular histones as potential DAMPs and their ability to induce inflammatory reactions. To this end, calf thymus histones (CTHs) were incubated with human heparinized blood and cytokine levels were determined semi-quantitatively with a multi-cytokine membrane array. A densitometric evaluation of the secreted cytokine pattern revealed that CTH stimulation triggered an increase in the levels of IL-6, IL-8, TNF-α, and IFN-γ when compared to blood incubated with buffer (PBS) alone (Fig 1A). These findings are in line with reports from Xu et al. who studied the release of these mediators in a murine model of inflammation [17]. We also found increased levels of the IFN-γ inducible chemokines, CXCL9 and CXCL10, but not CXCL11, and noted an up-regulation of the chemokines CCL2, CCL3, CCL7, and CCL20, respectively (Fig 1A, boxes).
CXCL9, CXCL10, and CXCL11 belong to the family of non-ELR CXC chemokines that exert their chemotactic activities by binding to CXCR3, a G protein-coupled receptor expressed on monocytes, macrophages, neutrophils, eosinophils, activated T-lymphocytes and NK-cells [22–26]. As the induction of these chemokines by extracellular histones has not been described, the focus was put on these three proteins and in particular on CXCL10 throughout this study. Thus, we investigated the production of the three chemokines over time. Blood from healthy volunteers was treated with CTHs for 12h and the chemokine response was recorded. Plasma levels of CXCL9 and CXCL10 increased significantly (Fig 2A and 2B) while no protein elevation for CXCL11 was observed (Fig 2C). We also found elevated levels of the other three chemokines, CCL3, CCL20, and CCL7 in these samples, suggesting that extracellular histones induce a broad immune response (Fig 2D–2F).
Further analysis of the non-ELR CXC chemokines revealed that the CXCL10 concentration reached a plateau after an incubation time of 10h, while the CXCL9 levels continued to increase even up to 12h. This was in contrast to the levels of CXCL11 which remained constant low over time (S1A Fig). The release of CXCL10 in blood was dose-dependent (S1B Fig) and did not cause cell damage as measured by the release of LDH (S1C Fig). However, at histone concentrations exceeding 50 μg/ml LDH release from PBMCs was recorded (S1D Fig). Together the data show that extracellular histones can trigger the mobilization of chemokines in human blood.
Cytokine (IL-6, IL-8, and TNF-α) induction evoked by stimulation with extracellular histones has been reported to involve a preceding activation of TLRs [17]. We therefore sought to investigate if the histone-dependent release of CXCL10 was mediated via TLR signaling. To address this, we used inhibitory antibodies against human TLR2, TLR4, and an isotype antibody as control. The inhibitory activity of the antibodies was first verified by their ability to block the release of interleukin-6 from blood treated with Pam3CSK3 (TLR2 agonist) or LPS (TLR4 agonist), respectively (S2 Fig). In the next series of experiments, CTHs in combination with the anti-TLR2, anti-TLR4, or isotype control antibodies were incubated with heparinized blood for 12h and the release of CXCL10 into plasma was measured. Fig 3A shows that anti-TLR4 significantly inhibited CXCL10 production, whereas anti-TLR2 and the isotype control had no effect. To further confirm that histones induce CXCL10 production through TLR4, the TLR4 antagonist CLI-095 was employed and the CTH-induced CXCL10 production was measured. Also these experiments revealed that the CTH-induced chemokine production is TLR4 dependent (Fig 3B).
The mobilization of CXCL10 is regulated by interferons [27] and incubation of whole blood with extracellular histones can induce the release of IFN-γ (Fig 1). We therefore tested the role of IFN-γ in our experimental model. To this end, human blood was incubated with CTHs in the absence or presence of an inhibitory monoclonal antibody against human IFN-γ or an isotype control. As seen in Fig 3C, CXCL10 production was down-regulated in the presence of the IFN-γ antibody, but not when the control antibody was applied. Together our data suggest that the release of CXCL10 is induced by a chain of events involving an activation of TLR4 that subsequently leads to activation of the IFN-γ receptor, resulting in the expression of chemokines.
Having characterized the mechanisms that lead to an up-regulation of histone-induced CXCL10 in human blood, we next wished to identify the cell populations that are involved in these processes. We decided to focus on histone H4, since we and others have shown that histone H4 interacts with many cell types including endothelial cell, platelets, and erythrocytes [12, 18]. To study the binding to human blood cells, histone H4 was added to blood and after fixation, cell-bound protein was detected with an antibody against histone H4 followed by a FITC-conjugated secondary antibody. Antibody binding was then recorded by flow cytometry analysis. Histone H4 interacts with monocytes and platelets, while no binding to neutrophils or the lymphocyte population was observed (Fig 4A). To investigate whether the interaction of histones and TLR4 takes place at the plasma membrane of monocytes, histone H4 was incubated with THP-1 monocytes. Cells were double immunostained with gold-labeled antibodies against histone H4 and human TLR4 respectively, and analyzed by transmission electron microscopy. TLR4 and exogenous histone H4 were found in close proximity (Fig 4B), suggesting that this interaction is responsible for the activation of monocytes. Additional surface plasmon resonance experiments were conducted to confirm the binding of TLR4 to histone H4. A sensorchip was immobilized with histone H4 and a flow of TLR4 in complex with its co-factor MD-2 was applied (Fig 4C). An antibody against human TLR4 was then used to confirm that the receptor and not only MD-2 had bound to histone H4 (Fig 4D). Together our results show that histone H4 binds TLR4/MD-2 on monocytes.
To exclude involvement of other cell types than monocytes for CXCL10 production, human PBMCs were separated from neutrophils using Polymorphprep. Isolated PBMCs were then stimulated with CTHs for 6 and 10h and analyzed for intracellular CXCL10. A gating strategy for cellular subsets was defined (Fig 5A), where CXCL10 production was found in CD14++CD16+ monocytes, while little or no CXCL10 was produced in response to CTHs in the other cell populations (Fig 5B). The production of CXCL10 started to appear after 6h and was clearly induced at 10h (Fig 5C and 5D) which is consistent with other results (S1A Fig). Beyond this, no CXCL10 was produced in neutrophils (S1E Fig). Thus, these findings clearly demonstrate that monocytes are crucial for histone-induced CXCL10 production.
Next, we wished to map the cell-binding site in histone H4. Therefore, a panel of overlapping histone H4-derived peptides was incubated with heparinized blood and the induction of CXCL10 was monitored. These results revealed that the cell-binding site is located at amino acids 35–54 (IRR20) towards the amino terminal portion of the protein (Fig 6A). To measure the affinity of the interaction between peptide IRR20 and TLR4/MD-2, surface plasmon resonance experiments were performed. IRR20 was immobilized to a CM5 sensorchip and TLR4/MD-2 was applied in a flow over the surface. Steady-state affinity analysis showed that IRR20 binds to TLR4/MD-2 with high affinity (KD = 2.726 nM) (Fig 6B). Additional modeling suggests that the binding segment forms a hairpin loop surrounded by α-helices (Fig 6C).
To elucidate the patho-physiological relevance of our findings, an in vivo model was used. CTHs were intravenously injected into 10 female Balb/c mice and blood samples were recovered from the animals by cardiac puncture at two time points (4h and 24h respectively). The blood was then centrifuged and samples from 5 animals in each group were pooled. The release of various inflammatory mediators into the plasma was measured in a multi-cytokine array. As shown in Fig 7A, CXCL10 levels increased up to five fold and CXCL9 levels up to 3 fold, respectively, when plasma samples were analyzed after a 4-hour incubation with CTHs. While the CXCL10 concentration decreased to background levels after the 24-hour treatment, CXCL9 increased further and reached levels that were about five times higher than in plasma samples from non-treated mice (Fig 7B). These findings suggest that CXCL9 and CXCL10 production occurs in vivo as a response to intravenous injections of CTHs.
CXCL9 and CXCL10 have been reported to act as chemoattractants for cells expressing the CXCR3 receptor, such as monocytes, macrophages, neutrophils, eosinophils, activated T-lymphocytes and NK-cells [22–26]. To test whether the local injection of histones into Balb/c mice leads to a recruitment of such immune cells, we employed a CTH air pouch model (see Material and Methods). Mice receiving a PBS injection instead of CTHs served as control. Sixteen hours after injection, fluids from the air pouch were collected and analyzed in the presence of counting beads by flow cytometry. Fig 8A depicts the increased cellular influx when animals were treated with CTHs. Further flow cytometry analysis revealed that different populations of leukocytes recruited into the air pouch are monocytes, eosinophils, NK-cells, and neutrophils, respectively (Fig 8B). As these cells have been reported to express the CXCR3 receptor [22–26] we next tested whether the recruitment of these cells correlates with the CXCL10 levels in the air pouch. For this purpose C57BL/6 wild-type mice received a subcutaneous injection of CTHs and the CXCL10 levels in the air pouch were measured after 16h of incubation. TLR4 and MyD88 knockout animals were used to confirm that the histone-induced release of CXCL10 is dependent on TLR4 receptor signaling. As shown in Fig 8C, increased levels of the chemokine were recovered from the air pouch of wild-type, but not of TLR4 and MyD88 knockout mice. Consequently, also the influx of leukocytes into the air pouch was significantly reduced in the knockout animals when compared to wild-type animals (Fig 8D). To exclude a LPS contamination in the CTH preparation, the synthetic histone H4-derived peptide IRR20 was injected in wild-type and TLR4 knockout mice. As seen for CTHs, IRR20 was able to trigger the induction of CXCL10 (Fig 8E) and recruitment of leukocytes (Fig 8F) in wild-type but not TLR4 knockout mice. Together these experiments show that subcutaneous injection of extracellular histones leads to the local release of CXCL10 and recruitment of leukocytes to the side of injection.
To confirm that extracellular histones are important for chemokine production during S. pyogenes infection, we inoculated C57BL/6 wild-type mice with S. pyogenes by subcutaneous injection. Accumulation of extracellular histone H4 in the air pouch was detected by Western blot analysis and Fig 9A depicts that increasing concentrations were measured 4h, 8h, and 24h after infection, respectively. We recently reported that p33, also known as gC1q receptor is a specific inhibitor of extracellular histones [12]. To test whether p33 can interfere with the chemotactic activity of the extracellular histones, the protein was injected into the air pouch 8h after infection. Subsequent Western blot analysis revealed that this treatment let to a reduction of histones in the air pouch which was in a similar low range as seen in animals that were treated with p33 for 24h in the absence of bacteria (Fig 9A). No histones were detected in air pouches of mice that were injected with PBS only. Purified recombinantly produced histone H4 was used to show the specificity of the antibody.
Having shown that extracellular histones are generated during infection, we next investigated whether this leads to an induction of CXCL10 in the air pouch. Fig 9B shows that elevated concentrations of the chemokine are detected in mice infected for 24h with S. pyogenes. However, when the mice were treated with p33, the levels of CXCL10 were significantly decreased and only minor CXCL10 concentrations were found in non-infected animals that received a p33 injection (Fig 9B). Together these findings suggest that endogenous extracellular histones were responsible for the CXCL10 production in the infected mice.
Our previous results show that histone H4 is released into the tissue during severe deep tissue infections such as necrotizing fasciitis and cellulitis [12]. Thus, we wished to study whether the release of extracellular histones co-localizes with the induction of CXCL10 in patients suffering from S. pyogenes tissue infections. Immunofluorescence microscopy was used to analyze tissue biopsies from patients with severe deep tissue infections, i.e. severe cellulitis or necrotizing fasciitis, evoked by S. pyogenes. Fig 10A (upper panel) illustrates that the infection caused necrotic tissue leading to severe cell damage. Further immunostaining analysis revealed the mobilization of extracellular histone H4 and CXCL10 and at some locations co-localization of the two factors was noted (Fig 10A, arrows, yellow). In the healthy tissue, only intact cells and consequently also only intracellular DNA was seen. We were therefore not able to co-localize histone H4 or CXCL10 in these tissue samples (Fig 10A, lower panel).
To investigate if CXCL10 is systemically up-regulated during infection, the levels of the chemokine were measured in plasma samples from patients admitted at the intensive care unit (ICU) suffering from septic shock or non-septic critical illness. For comparison samples from patients with milder infections (upper respiratory tract viral infections and gastroenteritis) or from healthy individuals were used. The highest levels of CXCL10 were found in plasma samples from ICU patients (Fig 10B). Though CXCL10 concentrations in plasma samples from patients suffering from mild infection were lower than in ICU patients, their levels were still significantly increased compared to those measured in samples from healthy donors (Fig 10B). Taken together, the patient data may support our in vitro and in vivo findings that the release of histones induces the induction of CXCL10 and that patients with severe disease conditions have a systemic increase of the chemokine.
The role of extracellular histones in the pathology of severe diseases such as sepsis has lately attracted considerable attention [10, 11, 15–17, 19, 29, 30]. Extracellular histones are released into the surrounding tissue or circulation from NETotic/necrotic cells and at high concentrations they can trigger toxic responses and evoke systemic inflammatory reactions. This in turn may lead to life-threatening complications including disturbed hemostatic functions and acute organ injury [19]. Animal studies have shown that mice receiving sub-lethal doses of histones develop symptoms that are also often seen in patients suffering from severe infectious diseases [11, 15, 17, 19]. These findings point to an important role for extracellular histones in disease progression. However, little is known whether the release of lower histone levels is less devastating and perhaps even could contribute to host reactions that are involved in immune defense or wound healing.
In the present study we sought to address this issue by studying if extracellular histones can act as danger signals by triggering local host responses. Our data indeed supported this notion, as histones induced the production of the IFN-γ inducible chemokines, CXCL9 and CXCL10, respectively. Both chemokines are important chemoattractants [23, 31] and our results show that the subcutaneous injection of extracellular histones evokes an influx of inflammatory cells to the challenged sites. Further characterization revealed that histones specifically target monocytes by approaching a pattern recognition receptor, namely TLR4 and in vivo experiment with TLR4 and MyD88 knockout mice confirmed the involvement of the receptor. Additional examination of tissue biopsies from patients with tissue infections depicted co-localization of extracellular histones and CXCL10, suggesting that the chemokine is released at the site of infection under disease conditions. When analyzing plasma samples from patients suffering from infectious diseases, we found that the increase of CXCL10 in these samples correlated well with the severity state of the disease. However, unlike described for other inflammatory mediators such as IL-6 and TNF-α [17] CXCL10 concentrations did not reach pathological levels within the microgram range (cytokine storm). We can therefore conclude that extracellular histones are able to cause local and systemic release of IFN-γ inducible chemokines, however, under systemic conditions it seems that these chemokines do not contribute as much to disease progression as other cytokines.
It is noteworthy to mention that analyzing the cytokine release pattern of murine blood samples revealed that CXCL9 and CXCL10 are the most abundant inflammatory mediators detected when mice were injected with CTHs. Notably, both chemokines have in addition to their chemotactic activities other features. For instance, CXCL9 and CXCL10 have been reported to carry antimicrobial activity [32, 33] and are pro-apoptotic [34]. The two effects have in common that they help prevent an induction of systemic inflammatory complications. This further strengthens our concept that the two chemokines are released in response to a danger signal and are part of a local immune response. Taken together we here provide for the first time evidence that histones can act as DAMPs/alarmins through inducing leukocyte migration and trigger the release of the IFN-γ inducible chemokines, CXCL9 and CXCL10. These findings demonstrate a dual role of extracellular histones in local and systemic disease conditions. It will be important to understand the details of this dual role in future development of therapies targeting extracellular histones.
The study was approved by the Institutional review board (IRB) at the Lund University Hospital (Protocol #790/2005). Plasma samples and tissue biopsies (Lund University Hospital, Sweden and University of Toronto, Canada) were taken with a written informed consent, and the study included patients above 18 years of age. Animal use protocols (#M108-10, #M326-12 and #M327-12) were approved by the local Institutional Animal Care and Use Committee (IACUC) of Malmö/Lund, Sweden. All animals were handled according to the Swedish Animal Welfare Act (SFS 1988:534). The Institutional Review Boards (IRBs) of the University of Toronto (Toronto, ON) and Karolinska University Hospital (Stockholm, Sweden) approved the studies involving humans, and written informed consent was obtained from the patients and the volunteers.
Female and male Balb/c, C57BL/6 wild-type and knockout mice were purchased from Charles River (Sulzfeld, Germany) and were at least 8 weeks old at the initiation of the experiments.
Bovine calf thymus histones (CTHs) were purchased from Roche (Basel, Switzerland). Overlapping Histone H4-derived peptides were synthesized at Biopeptide Co. (San Diego, CA, USA). CTHs have been used by researchers for both in vivo and in vitro studies of the function of extracellular histones [11, 12, 15–18, 20]. LPS-free CTHs and LPS-free histone H4-derived peptides were dissolved in LPS-free water. Recombinant histone H1, H2A, H2B, H3.1 and H4 were purchased from New England Biolabs (Ipswich, MA, USA). Recombinant human TLR4/MD-2 complexes were obtained from R&D Systems (Minneapolis, MN, USA) Peroxidase-conjugated goat anti-rabbit and goat anti-mice immunoglobulin G was from Bio-Rad Laboratories (Berkeley, CA, USA). Antibodies against histone H4 (polyclonal chip-grade anti-histone H4) were from Abcam (Cambridge, UK). Rat polyclonal antibodies to human TLR4, TLR2 and isotype control were purchased from Invivogen (San Diego, CA, USA). Mouse anti-human interferon-γ (clone B27) and isotype control were purchased from BioLegend (San Diego, CA, USA). Pam3CSK was obtained from Invivogen (San Diego, CA, USA) and LPS from Salmonella enterica serotype typhimurium from Sigma-Aldrich (St. Louis, MO, USA).
Plasma samples from 30 patients enrolled at Lund University Hospital, Sweden were analyzed. Fifteen patients treated in the ICU due to septic shock (n = 8) and non-septic critical illness (n = 7) and another 15 patient with milder infections such as upper respiratory tract viral infections (n = 10) and gastroenteritis (n = 5) were enrolled. Blood samples for the analyses were collected at enrollment in 5 ml plastic vacutainer tubes containing 0.5 mL of 0.129 mol/l sodium citrate.
Peripheral venous blood samples were collected from healthy human donors into 6.0 ml sodium heparin tubes (102 IU) or 2.7 ml 0.109 M buffered sodium citrate (Becton Dickinson, Franklin Lakes, NJ, USA). Blood samples were immediately used for experiments upon collection.
The presence of LPS was measured in all histone subclasses, histone H4-peptides and CTHs using the Pierce LAL Chromogenic Endotoxin Quantitation Kit (ThermoFisher Scientific, MA, USA) according to manufacturer’s protocols. Histone H2A, H2B and H3.1 contained significant levels of LPS (S3A Fig), while no LPS was found in histone H4, CTHs or histone H4-derived synthetic peptides (S3B Fig). Therefore, only CTHs, histone H4 and histone H4-derived peptides were used throughout this work.
Human XL Cytokine Array Kit and Mouse Cytokine Array Panel A were purchased from R&D Systems (Minneapolis, MN, USA) and experiments were performed according to the manufacturer’s protocols. Human heparin blood was stimulated with CTHs (60 μg/ml, 12 h) centrifuged (2000 x g for 10 min) and the resulting plasma was used for the Human XL Cytokine Array Kit. Mice were injected with CTHs (25 mg/kg body weight, 4 h) and euthanized with CO2 inhalation followed by cervical dislocation. Blood was drawn by cardiac puncture using a syringe filled with 70 μl of buffered sodium citrate and then transferred into K2 Microtainer MAP tubes pre-coated with 1 mg EDTA (Becton Dickinson, NJ, USA). Blood was centrifuged and the mouse plasma was used for the Mouse Cytokine Array Panel A.
Analyses were performed with a BIAcore X100 instrument (GE Healthcare, Uppsala, Sweden) using Sensor Chip CM5 technology at 25°C in a HBS-EP buffer (10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.05% (v/v) Surfactant P20, pH 7.4). Histone H4 or IRR20 was diluted in sodium acetate (10 mM, pH 5.5) and immobilized via amine coupling to flow cell 2. Flow cell 1 was subjected to the coupling reaction but without protein, and was used as a control in each experiment. TLR4/MD-2 was injected over the coated surface at 30 μl/min in running buffer. Regeneration of sensorchip surfaces was obtained by injection of 30 μl 50 mM NaOH.
CTHs (60 μg/ml), Pam3CSK4 (200 ng/ml) or LPS (100 ng/ml) were incubated with heparinized blood at 37°C on rotation. At different time points, aliquots were removed and centrifuged at 2000 x g for 10 min. All plasma samples were stored at -20°C for further experiments. For inhibition experiments, the TLR4 signaling inhibitor CLI-095 (1.25 μg/ml) or antibodies with blocking activity against human IFN-γ (25 μg/ml), human TLR2 (25 μg/ml), human TLR4 (25 μg/ml) or polyclonal control antibodies (25 μg/ml) were added simultaneously with 60 μg/ml CTHs, Pam3CSK4 or LPS. Histone H4-derived synthetic peptides were incubated with heparinized blood at 37°C on rotation for 12 h. After centrifugation, CXCL10 was measured in the plasma using Human CXCL10/IP-10 Quantikine ELISA kit from R&D Systems (Minneapolis, MN, USA).
PBMCs were isolated from heparinized human blood using Polymorphprep (Axis-Shield, Dundee, Scotland) according to manufacturer’s protocols. PBMCs (8x105 cells/ml) were stimulated with different concentrations of CTHs for 6h, 10h or 12h at 37°C, 5% CO2.
Cytokine ELISA kits for human CXCL10, mouse CXCL10, human CXCL9 and human CXCL11 were purchased from R&D Systems (Minneapolis, MN, USA) and performed according to manufacturer’s instructions. ELISA kits from R&D Systems were not optimized for plasma samples from citrated blood. Therefore heparinized blood was used for these experiments. Cytokine ELISA kits for human CCL3, CCL20, CCL7 were purchased from Sigma-Aldrich (St. Louis, MO, USA).
Histone-induced cell toxicity was measured in heparinized blood or PBMCs using the Tox-7 kit (Sigma Aldrich, St. Louis, MO, USA). Cells were stimulated with CTHs (0–100 μg/ml) for 12h at 37°C, 5% CO2. Cells were pelleted and the supernatants were used for measurement of lactate dehydrogenase (LDH) according to the manufacturers’ protocol. Histone-induced LDH release was calculated as percentage of Tox-7 lysis-induced LDH release.
Citrated blood (20 μl) was diluted with 40 μl Cell Staining Buffer (BioLegend, San Diego, CA, USA). CTHs (final concentration 200 μg/ml) were added and samples were incubated for 30 min, 37°C, 5% CO2. Cells were washed in 900 μl cell staining buffer, centrifuged (500 x g for 15 min, 4°C) and the pellet was resuspended in 200 μl buffer. Histone H4 bound to cells was stained with 2 μg/ml of a primary rabbit anti-human histone H4 (Abcam, Cambridge, UK) for 30 min on ice. Erythrocytes were lysed first with 200 μl Lysis buffer A (10 min, room temperature, dark) followed by 1 ml Lysis buffer B (10 min, room temperature, dark). Supernatant containing lysed erythrocytes were removed and remaining white blood cells were resuspended in 1 ml cold cell staining buffer. A secondary FITC-conjugated goat anti-rabbit antibody (Abcam, Cambridge, UK) was added (2 μg/ml) and samples were incubated for 30 min on ice in dark. Two more wash steps were performed and each sample was analyzed in flow cytometry.
Total PBMCs were cultured with or without CTHs (final concentration 10 μg/ml) in RPMI 1640 for 6h or 10h with 1X Brefeldin A (eBioscience, San Diego, CA, USA) added during the last 4 h. Stimulated cells were collected in cold FACS buffer (PBS supplemented with 2% FCS and 2 mM EDTA) and unspecific binding was blocked with 10% mouse serum. Surface stain was done with the following antibodies; anti-CD19 (HIB19), anti-CD16 (3G8) (BioLegend, San Diego, CA, USA), anti-CD4 (RPAT4) (BD Biosciences, Franklin Lakes, NJ, USA), CD3 (UCHT1), anti-CD8 (RPAT8), anti-CD14 (61D3) and anti-CD56 (CMSSB) (eBioscience, San Diego, CA, USA) conjugated to allophycocyanin-Cy7, FITC, Alexa700, eFluor450, PE-Cy7, allophycocyanin or biotin revealed by streptavidin-PECF594 (BD Biosciences, Franklin Lakes, NJ, USA). Dead cells were excluded with Live/Dead fixable aqua dead cell stain kit (Life Technologies). Intracellular staining was done with the FoxP3 Fixation/Permeabilization kit (eBioscience, San Diego, CA, USA) using anti-IP10 (J034D6) conjugated to PE (BioLegend). Data were collected using a LSRII (BD Biosciences, Franklin Lakes, NJ, USA) and analyzed with the FlowJo software (Tree star, Ashland, OR, USA).
CTHs (100 μl, 10 μg/ml,), IRR20 (100 μM) or PBS were injected subcutaneously into an air pouch of male and female wild-type (C57BL/6), TLR4 knockout (C57BL/6), MyD88 knockout (C57BL/6), or wild-type Balb/c mice. After 16h the air pouch was rinsed with 1 ml PBS and the rinsing solution was fixed in 2% paraform aldehyde for 30 min on ice. CountBright Counting Beads (ThermoFisher Scientific, MA, USA) were diluted 1:10 into the fixed cells and the samples were counted with a FACSVerse Flow Cytometer (Becton Dickinson, NJ, USA). Cells and counting beads were gated with side scatter and forward scatter and analyzed with FlowJo 9.3.1 Software. Identification of cellular subsets was done by flow cytometry in conjunction with mAb against CD64 (X54-5/7.1), CD3 (145-2C11) (BD Biosciences, Franklin Lakes, NJ, USA), NK1.1 (PK136), Ly6G (1A8), Ly6C (HK1.4), CD11c (N418), CD11b (M1/70) (BioLegend), B220 (RA3-6B2) and CD4 (GK1.5) (eBioscience, San Diego, CA, USA). Unspecific binding was blocked using anti-FcR mAb (2.462) and dead cells were excluded with propidium iodide (PI). Data were collected using a LSRII (BD Biosciences, Franklin Lakes, NJ, USA) and analyzed with FlowJo (Tree star). Monocytes (CD11b+Ly6C+CD64+), neutrophils (CD11b+Ly6G+), eosinophils (CD11b-CD64-Ly6CintSSC-Ahigh) and NK cells (CD11b-Ly6G-NK1.1+) were identified after gating on PI- CD3- B220- events.
Wild-type C57BL/6 mice were subjected to a subcutaneous infection of S. pyogenes (AP1 strain) as previously described [35]. Briefly, mice were inoculated with 108 CFU/ml log-phase bacteria, and euthanized at different time points. To remove endogenous histones released after S. pyogenes infection, mice were treated by administration of p33 (50 μg/ml) 8h after initial infection. After 4 h, 8h or 24 h, the air pouch was rinsed with 1 ml PBS and cells and bacteria were pelleted. Samples were subjected to Western blot analysis in the cell supernatant by pooling samples from mice in the same group. Histone H4 was detected using mouse anti-histone H4 antibody (ab31830, Abcam, Cambridge, UK) and secondary goat anti-mouse HRP (1:1000, Bio-Rad Laboratories Berkeley, CA, USA). CXCL10 levels in the air pouch were measured by ELISA.
Snap-frozen tissue biopsies collected from the center of infection from three patients with deep tissue infections, i.e. severe cellulitis or necrotizing fasciitis, caused by S. pyogenes were analyzed (kindly provided by Prof. Donald E Low, Mount Sinai Hospital, Toronto, Canada). Also included were biopsies of healthy skin tissue obtained at plastic surgery at Karolinska University Hospital. The biopsies were cryosectioned to 8 μm, fixed in ice-cold acetone and immunofluorescent stainings of serial sections were conducted as described previously [36]. The following antibodies were used: rabbit polyclonal anti-histone H4 and mouse monoclonal anti-IP10 (clone 6D4) (both from Abcam, Cambridge, UK), Alexa 546 conjugated donkey anti-rabbit IgG and Alexa 488 conjugated donkey anti-mouse IgG (both from Molecular Probes, Eugene, OR, USA). Slides were mounted using DAPI supplemented mounting media (Molecular Probes, Eugene, OR, USA). Single stainings were performed to assure specificity of staining patterns. For image evaluation, a Nikon A1R confocal microscope was used (Nikon Instruments, Amstelveen, the Netherlands).
Monocyte THP-1 cells were incubated with histone H4 (50 μg/ml) for 30 min at 37°C. For immunohistochemistry and transmission electron microscopy, samples were embedded in Epon resin, sectioned and subjected to antigen retrieval with metaperiodate as recently described in detail [37]. Sections were labeled with rabbit anti-histone H4 and rat anti-TLR4 followed by gold-conjugated goat-anti rabbit (5 nm) and goat-anti rat (10 nm) antibodies. Samples were observed in a Philips/FEI CM 100 transmission electron microscope at the Core Facility for Integrated Microscopy, Panum Institute, University of Copenhagen.
Data were analyzed using GraphPad Prism 6 (GraphPad Software, San Diego, CA, USA). Statistical significance was determined by using the non-parametric Mann-Whitney U test for comparison between two groups. The level of statistical significance was defined as a two-tailed p-value of < 0.05 (*), < 0.005 (**) and < 0.0005 (***).
CXCL9, Q07325
CXCL10, P02778
CXCL11, O14625
CXCR3, P49682
IFN-γ, P01579
TLR4, O00206
TLR2, O60603
p33, Q07021
Histone H4, P62805
|
10.1371/journal.pcbi.1003378 | Enumeration of Smallest Intervention Strategies in Genome-Scale Metabolic Networks | One ultimate goal of metabolic network modeling is the rational redesign of biochemical networks to optimize the production of certain compounds by cellular systems. Although several constraint-based optimization techniques have been developed for this purpose, methods for systematic enumeration of intervention strategies in genome-scale metabolic networks are still lacking. In principle, Minimal Cut Sets (MCSs; inclusion-minimal combinations of reaction or gene deletions that lead to the fulfilment of a given intervention goal) provide an exhaustive enumeration approach. However, their disadvantage is the combinatorial explosion in larger networks and the requirement to compute first the elementary modes (EMs) which itself is impractical in genome-scale networks.
We present MCSEnumerator, a new method for effective enumeration of the smallest MCSs (with fewest interventions) in genome-scale metabolic network models. For this we combine two approaches, namely (i) the mapping of MCSs to EMs in a dual network, and (ii) a modified algorithm by which shortest EMs can be effectively determined in large networks. In this way, we can identify the smallest MCSs by calculating the shortest EMs in the dual network. Realistic application examples demonstrate that our algorithm is able to list thousands of the most efficient intervention strategies in genome-scale networks for various intervention problems. For instance, for the first time we could enumerate all synthetic lethals in E.coli with combinations of up to 5 reactions. We also applied the new algorithm exemplarily to compute strain designs for growth-coupled synthesis of different products (ethanol, fumarate, serine) by E.coli. We found numerous new engineering strategies partially requiring less knockouts and guaranteeing higher product yields (even without the assumption of optimal growth) than reported previously. The strength of the presented approach is that smallest intervention strategies can be quickly calculated and screened with neither network size nor the number of required interventions posing major challenges.
| Mathematical modeling has become an essential tool for investigating metabolic networks. One ultimate goal of metabolic network modeling is the rational redesign of biochemical networks to optimize the production of certain compounds by cellular systems. Accordingly, several optimization techniques have been proposed for this purpose. However, for large-scale networks, an effective method for systematic enumeration of the most efficient intervention strategies is still lacking. Herein we present MCSEnumerator, a new mathematical approach by which thousands of the smallest intervention strategies (with fewest targets) can be readily computed in large-scale metabolic models. Our approach is built upon an extended concept of Minimal Cut Sets, the latter being minimal (irreducible) combinations of reaction (or gene) deletions that will lead to the fulfilment of a given intervention goal. The strength of the presented approach is that smallest intervention strategies can be quickly calculated with neither network size nor the number of required interventions posing major challenges. Realistic application examples with E.coli demonstrate that our algorithm is able to list thousands of the most efficient intervention strategies in genome-scale networks for various intervention problems.
| This is a PLOS Computational Biology Methods article.
Stoichiometric and constraint-based modeling techniques such as flux balance analysis or elementary modes analysis have become standard tools for the mathematical and computational investigation of metabolic networks [1]–[4]. Although these methods rely solely on the structure (stoichiometry) of metabolic networks and do not require extensive knowledge on mechanistic details, they enable the extraction of important functional properties of biochemical reaction networks and deliver various testable predictions. The steadily increasing number of reconstructed and examined genome-scale metabolic network models of diverse organisms proves that methods for constraint-based modeling can deal with networks comprising up to several thousands of metabolites and reactions [1].
Metabolic networks consisting of m internal metabolites and n reactions can be formalized by an m×n stoichiometric matrix N. A common assumption of constraint-based methods is that the network is in steady state (i.e., the metabolite concentrations do not change) resulting in a system of homogeneous linear equations(1)where r is the vector of (net) reaction fluxes or reaction rates. In addition, the non-negativity constraints on fluxes through irreversible reactions must be fulfilled:(2)(Irrev comprises the indices of the irreversible reactions). The two constraints (1) and (2) form a convex polyhedral cone (the flux cone) in the n-dimensional space of the rate vectors r. Flux Balance Analysis (FBA; [3]) searches for optimal flux distributions within this cone that maximize a given linear objective function(3)Typical objective functions are maximization of growth (or biomass yield) or of the yield of a certain product. For FBA, the irreversibility constraint (2) can be refined to general upper and lower boundaries for each reaction rate ri:(4)Elementary-modes analysis [2], [5] is another stoichiometric technique facilitating the exploration of the space of feasible steady state flux distributions by means of particular flux vectors e fulfilling the basic constraints (1) and (2) and in addition a non-decomposability property. The latter demands that an elementary mode e is irreducible (or support-minimal), hence, there is no vector r≠0 obeying (1) and (2) and(5)Here, P(r) and P(e) represent the support of r and e, respectively, i.e., they contain the indices of the vector elements being non-zero: P(t) = {i | ti≠0}. Elementary modes (EMs) represent stoichiometrically balanced metabolic pathways or cycles and several important properties of a metabolic network can be analyzed by its unique set of EMs [2], [5]. EMs correspond to extreme rays of convex cones and can be computed as such [6], [7].
One ultimate goal of metabolic network modeling is the targeted manipulation of the network behavior. A typical application is metabolic engineering where one is interested in the optimization of the production of a certain compound by a given host organism. A number of constraint-based optimization techniques have been proposed for this purpose [2], [8], [9], [10], [11], [12]. FBA can directly be used to determine the optimal (maximal) value for a given optimization problem (e.g., maximal yield of biomass or of a certain chemical when growing on a certain substrate). This approach, however, cannot yet explain which manipulations will eventually drive the cell towards this optimum. A simple approach would be to use flux-variability analysis (FVA, [13]) to analyze how the feasible ranges of stationary fluxes in a metabolic network would change when switching from the wild-type to a desired phenotype. More sophisticated and directed FBA-based optimization routines operate on the principle put forward by the OptKnock approach [8]. Here, the key idea is to search for interventions that lead to obligatory coupling between the production of biomass and of a desired compound. Mathematically, OptKnock is a bilevel optimization problem where the inner problem defines biomass optimization as the cellular objective and where the outer optimization problem is to search for reaction removals (represented by integer variables) that lead, under consideration of the inner problem, to maximal product formation. The bi-level optimization coupling can be reformulated as a single level mixed integer linear program (MILP). Successful applications (e.g. [14]) and several refined variants of OptKnock (including, for example, RobustKnock [9] and OptORF [11]) have been published (for a review see [12]). The advantage of FBA-based approaches is that they can readily be applied to genome-scale networks. However, a potential disadvantage is that they deliver particular solutions only where often multiple alternate solutions exist which might be equally or even more relevant than the found solutions. Some methods have therefore been proposed to enumerate intervention strategies. A brute-force approach would be to test all single, double, triple … reaction knockouts with respect to their impact on the objective function [15], [16]. Suthers et al. [15] used this method to enumerate synthetically lethal reaction sets and found that this search becomes prohibitive in genome-scale networks for interventions with more than two or three reaction knockouts (the upper limit set in [16] was also three). They designed therefore a more directed search algorithm based on a bi-level optimization method formulated as a mixed integer linear program (MILP) [15]. However, to the best of our knowledge, enumerated knockout sets in genome-scale networks did not exceed a cardinality of three. This is a serious limitation because complex interventions problems may require 5, 6, 7 or more knockouts, even in medium-scale networks (see [17] and the examples in the Results section).
The method of Minimal Cut Sets (MCSs) directly addresses the enumeration of metabolic intervention strategies [10], [18], [19]. MCSs specify minimal sets of reactions whose removal (knockout) will block certain undesired (target) flux distributions. For example, one can compute (i) MCSs that block growth; (ii) MCSs that disable the production of a certain compound; (iii) MCSs that block all flux vectors where a certain compound is produced with a low (including zero) yield. In the context of MCSs, the term “minimal” refers to the property that reaction cuts specified by any proper subset of an MCS are insufficient to ensure the full repression of the undesired behaviour. In this regard, the minimality of MCSs is similar to the minimality or non-decomposability property of elementary modes specified by equation (5). In fact, there is a dual relationship between MCSs and EMs: the MCSs blocking a certain set of target flux vectors are the minimal hitting sets of the set of (target) EMs that generate these behaviors [19], [20]. By this property, each MCSs must hit (knockout) at least one utilized reaction from each EM. As a consequence, MCSs can be computed as minimal hitting sets (or so-called hypergraph transversals) of the target modes, for instance, by the Berge algorithm (see [20]) or by Binary Linear Programming [21].
Another approach to compute MCSs, which exploits the inherent dual relationship between EMs and MCSs, was recently presented by Ballerstein et al. [22]. Briefly, the MCSs of a given metabolic network can be computed as certain EMs of a dual network; the latter being derived by a simple transformation of the (primal) network. This finding makes it possible to calculate MCSs by using optimized algorithms for EM computation [7].
However, there are two potential problems related to MCSs. First, when the reactions contained in an MCS are removed, we are sure that the targeted network functions are disabled but other (desired) functions might be blocked as well. For instance, it can occur that an MCS which disables low-yield pathways synthesizing a desired product also blocks growth of the organism making this MCS impractical. To prevent such side effects, the concept of constrained minimal cut sets (cMCSs) was introduced by Hädicke and Klamt [10] where not only undesired but also desired functionalities (to be preserved) can be specified. When the EMs are available, an adapted Berge algorithm can be used to conveniently compute cMCSs by specifying in addition to the target modes (expressing the unwanted behaviour) a set of desired modes expressing the functionality that must be preserved. A cMCS will hit all target EMs and keep a (user-specified) minimal number of desired EMs. As shown in [10], cMCS provide a very flexible and powerful approach to enumerate intervention strategies; many other techniques such as Minimal Metabolic Functionality [2], [17], and the aforementioned OptKnock and RobustKnock may be reformulated as special cMCSs problems. cMCSs are also well-suited to identify knockout combinations leading to coupled growth and product formation.
The second and more serious problem of (c)MCSs is that their full enumeration in large/genome-scale networks becomes prohibitive. The algorithms requiring as inputs the target (and possibly desired) EMs are usually not applicable: despite large progress in algorithmic design [7] the full set of EMs is often not computable at genome-scale. For the same reason, the dual approach of Ballerstein et al. [22] cannot be applied either.
On the other hand, for the purpose of applying MCSs in real networks, those with the smallest number of elements are usually the most relevant. Thus, it is worthwhile to consider computing only the (c)MCSs with low cardinality. The effective enumeration of the smallest cut sets is therefore the key goal of the present work.
Usually, the unwanted/desired functionalities to be disabled/kept in a metabolic network can be described by sets of linear equalities and inequalities over the fluxes. For the purpose of computing MCSs, we could therefore use an exhaustive FBA-based scheme by testing all single, double, triple and higher knockout sets whether they are suitable cut sets or not. The formulation of FBA problems would circumvent the problem to enumerate the EMs first. However, as discussed above, this approach becomes problematic if larger knockout sets are required to solve an intervention problem, as it must test a large number of candidate sets with increasing MCS size (the number of candidates grows with where n is the number of possible cuts and k the size of cut set candidates). Therefore, it is not normally possible to find genome-scale MCSs in reasonable time with more than 4 knockouts using this scheme.
Whereas the direct calculation of smallest MCSs in large-scale networks cannot be properly addressed yet by current methods, a method for computing the smallest (or shortest) EMs in genome-scale networks was recently presented by de Figueiredo et al. [23]. This algorithm formulates the search for the EMs with fewest elements as a Mixed Integer Linear Programming (MILP) problem and delivers in the k-th iteration the k-th shortest EM (hence, it starts with shortest EM, delivers then the second shortest and so forth). As shown by the authors, this approach can readily be applied to genome-scale networks to find the first hundred or even thousand shortest EMs involving the fewest number of reactions.
The goal of the present work is to realize a similar approach for computing the k-smallest MCSs from a given network structure. We show that this can be achieved in two steps. First, the original network and the actual intervention goal are converted to its dual representation using the approach of Ballerstein et al. [22]. We then compute the shortest EMs (up to a certain size or number) in the dual network by employing a modified algorithm of de Figueiredo et al. [23]. As the EMs in the dual network correspond to the MCSs of the primal, the shortest EMs in the dual system will represent the smallest MCSs of the original network.
The paper is organized as follows: we will first briefly review the approach of de Figueiredo et al. for computing k-shortest EMs and introduce several modifications improving the performance of this algorithm. In particular, we will make use of certain features of MILP solvers provided for effective enumeration of solutions of a MILP problem.
Thereafter we will describe how the network constraints (including inhomogeneous constraints) and the intervention goal have to be translated into their dual description in which we can then enumerate the shortest EMs to obtain the smallest MCSs in the primal network. We shall also explain how constrained MCSs can be computed within this framework. Finally, to demonstrate the power of our new approach we will exemplify its use by computing relevant intervention strategies (of different complexities) in iAF1260, a genome-scale metabolic model for E.coli [24]. These benchmarks demonstrate, for example, that our approach enables us to enumerate synthetic lethals of E.coli up to size 5 which was not possible before. Moreover, we show that the algorithm facilitates the calculation of thousands of the minimal intervention strategies that lead to growth-coupled synthesis of certain compounds by E. coli.
For the sake of simplicity, throughout the manuscript we will deal with reaction cut (or knockout) sets, which must in practice be translated to gene knockout sets to construct the corresponding mutants. This transformation can be easily achieved if the corresponding gene-enzyme-reaction associations are available. The latter could also directly be included in the problem formulations given below to compute gene (instead of reaction) cut sets.
We present now the key methodological development of this work showing that the basic algorithm for enumerating shortest EMs introduced in the previous section can also be used to compute smallest MCSs. The procedure is based on the duality properties of EMs and MCSs presented by Ballerstein et al. [22] which we outline in the following. A necessary first step to establish the scheme is to describe the undesired network functionality (the “target flux vectors” r to be disabled by the MCSs) by a suitable inequality constraint(13)where t is a (n×1) vector. Usually, t corresponds to a single row with zeros except a single 1 for a target reaction (rate) whose operation is to be blocked (e.g. biomass formation if we searched for synthetic lethals). Setting in addition b to 1 we would target all flux vectors in which the rate of the target reaction is non-zero (in our context we can again set b to an arbitrary value greater than zero without loss of generality).
Constraint (13) specifying the target flux vectors can be generalized to:(14)Here, matrix T (of size t×r) together with poses t inhomogeneous inequality constraints defining the target flux polyhedron (which may be bounded becoming then a polytope). It must be made sure that the zero flux vector is not contained in the target flux polyhedron as it can not be blocked by reaction knockouts. A nice feature of (14) is that we may directly include inhomogeneous constraints to characterize target flux vectors (with maintenance ATP demand as a typical example).
In addition to (14) and to the standard network constraints (1) and (2), Ballerstein et al. augmented the system by equality constraints setting all reaction rates to zero(15)(I is the (n×n) identity matrix). These constraints ensure that the system becomes infeasible as the zero flux vector implied by (15) contradicts (14). Note that (15) can be seen as the maximal (trivial) cut set knocking out every reaction in the network. In fact, the MCSs correspond to minimal subsets of the homogeneous equations in (15) which ensure (induce) inconsistency of the inequality system posed by constraints (1), (2), (14) and (15). Minimal subsets of constraints that induce inconsistency of an inequality system are also known as irreducible inconsistent subsets (IISs; [25]). Generally, IISs can be calculated as follows: using the Farkas Lemma, the infeasible primal system is converted to its dual system which is ensured to be consistent. It can be shown that the IISs of the primal system correspond to extreme rays (and thus EMs) in the dual system which makes it possible to calculate them using methods from EM computation. Since IISs in our particular case may, in general, also contain constraints from (1) or (2), a modified algorithm was introduced in [22] to ensure that only those IISs ( = EMs in the dual system) are computed which are minimal with respect to the constraints in (15) and correspond thus to the MCSs.
We thus need to transform the primal system defined by (1), (2), (14), (15) into its dual which can be written as follows (cf. equation (8) in [22]; Ndual is the “dual stoichiometric matrix” and rdual the dual “rate” vector):(16)The (sub-)matrix contains the identity matrix for irreversible reactions of the primal system and is filled with n-|Irrev| zero rows at the position of reversible reactions (note that reversible reactions of the primal system need not to be split before dualizing the system; however, reversible reactions affected by (14) must sometimes be split to properly describe the target flux polyhedron). As described above, the MCSs in the primal correspond to particular EMs of the dual system (16) which have minimal support in the v variables. The dual variables vi, i∈{1 … n} are thus of particular importance as their values indicate which reactions participate in an MCS. Concretely, if vi≠0 then reaction i is part of the MCS (irrespective of the sign of vi), if vi = 0 then it is not. Therefore, similar as we did for reversible reactions when computing shortest EMs, both positive and negative values of vi must be checked with indicators and in order to facilitate this each vi is split into two variables, vpi and vni, both with the lower bound 0. Furthermore, since h≥0 and because the MILP can directly operate on inequalities, we can rewrite (16) to:(17)(the sub-matrices with subscript i refer to the part of the irreversible reactions and subscript r to the part of the reversible reactions of the primal system). As mentioned above, for the vpi and vni we introduce the associated indicators zpi and zni, and (in equivalence to (8)) the constraints(18)stating that vpi and vni cannot be active simultaneously. The constant c in (17) can again be set to any positive value (e.g., to 1); this will not change the set of minimal non-zero combinations of vpi and vni fulfilling (17) which are relevant for the optimization problem formulated below (eq. (19)).
After dualization, we can now compute the smallest MCSs of the primal system by applying algorithm ALGO2 in the dual system. As constraints we need to consider (17) (replacing (1) and (2) from the primal system) as well as (18) and as objective function we exchange (10) with(19)Furthermore, because the presence of one reaction in a concrete solution is now indicated by two separate variables, the exclusion constraints (11) must be adapted accordingly to ( and are the values of a given concrete solution and is a shortcut for +):(20)In this way both positive and negative values of the original vi are counted in the same way towards reaction participation in the MCS. Finally, for the same reason, the size control constraint (12) sums here over zpi+zni as in the objective function (19).
The MCSs of the primal network are eventually obtained by taking the z-vector of the solutions found in the dual; z is obtained by collapsing zpi and zni: zi = zpi+zni.
In the previous subsection we dealt with enumeration of smallest MCSs, however, we have not yet clarified how constrained MCSs can be computed by this approach. As it turns out, this is straightforward: one first enumerates the smallest MCSs blocking the undesired flux vectors as described above. We can assume that the desired flux vectors (of which at least one has to be kept) is formulated by appropriate inequalities - similar as for the targeted undesired flux vectors in (14):(21)We can then filter the true cMCS from the set of (unconstrained) MCSs by testing for each MCS with a separate linear program whether the removal of the reactions in the MCS still allows the network to perform the desired function, i.e., whether the system given by (1), (2), and (21) is feasible when setting the rates of the reactions contained in the MCS to zero. From our experience, the computational costs for these tests are negligible compared to the calculation of the smallest MCSs, even if hundred thousand MCSs have to be tested (see Results section).
The MCSEnumerator method has been integrated as a new functionality in the CellNetAnalyzer package, a MATLAB toolbox for biological network analysis [26], [27]. The implementation uses the IBM ILOG CPLEX Optimization Studio V12.4 for solving the respective MILP and LP problems. Arbitrary intervention problems can be defined by providing the respective matrices and vectors describing the network and the desired and undesired flux vectors. The resulting MILPs are set up via the JAVA-CPLEX API and MATLAB's integrated JVM while for running the LPs the MATLAB-CPLEX interface is used. A separate package containing the data and script files needed for running the iAF1260 examples discussed herein can be downloaded from http://www.mpi-magdeburg.mpg.de/projects/cna/etcdownloads.html.
We analyze basic properties of the runtime behavior of our algorithm by means of three realistic benchmark problems with different complexities. All computations were performed with the CPLEX 12.4 MILP solver. When using multiple threads deterministic parallel mode was used to get repeatable behaviour. The search tree that CPLEX dynamically constructs took up less than 3 GB of RAM for all the systems used here.
In order to compare our MILP-based MCS enumeration scheme to other approaches the same benchmark problems as in Table 1 in [22] were used. The target of the (unconstrained) MCSs in these problems is the deactivation of biomass synthesis in a smaller model of the central metabolism of E. coli for growth under different substrates (acetate, succinate, glycerol, glucose). The MCSs determined in this way will thus correspond to the synthetic (reaction) lethals for E. coli (whose compositions depend strongly on the provided substrate). Before using the different MCS calculation routines the metabolic network is compressed by combining correlated reactions (operating with a fixed ratio under steady state conditions) to single cumulated reactions [6]. The compression in the primal system can also conducted if the computation is done in the dual system. MCSs found in the compressed network must be decompressed after calculation [18].
The number of calculated MCSs and computation times are shown in Table 1. As a first observation, it is apparent that calculation of EMs followed by the Berge algorithm (computing MCSs as the minimal hitting sets of the selected target EMs; Haus et al. 2008) is the most efficient of the shown MCSs calculation methods. The approach of Ballerstein et al. to compute primal MCSs as EMs in the dual system performs similar to EM calculation+Berge algorithm in the (primal) network but in its current implementation it requires a lot of memory. For this reason, the MCSs for glucose could not exhaustively be enumerated by this approach on the computer used (with an effective memory limit of 2GB per process).
Although the MILP algorithm developed herein was actually developed to compute the smallest MCSs, we can use it here even for enumerating all of them. The EMs in the dual network (the MCSs in the primal) where computed with both MILP formulations for shortest EM calculation: ALGO1 (the original approach by de Figueiredo et al. [23] implemented with indicator variables) and the ALGO2 approach calling the populate sub-routine for fixed EM sizes. Generally, applying the MILP formulations to the dual system is at first sight comparatively slow even when using multiple threads. Nonetheless, it is apparent that solving the dual system with our new ALGO2 is more efficient (∼17 times faster) than ALGO1 based on the scheme used by de Figueiredo et al. [23]. As can be seen for the MCSs with glucose as substrate, increasing the number of threads from 1 to 4 on the same CPU decreases the time needed for computation to some extent when using ALGO1 or ALGO2. Using 12 threads on a compute cluster node yields a more noticeable speed improvement but, as in the case of 4 threads, the combined computation times of all threads is still larger than in the case where a single thread is used.
The main advantage of our new approach can be seen in the case where only the MCSs up to size 4 have to be calculated (fifth row in Table 1): here the dual approach in combination with ALGO2 is clearly the fastest way to determine small MCSs among the approaches compared.
As described in the Introduction section, the direct calculation of EMs and MCSs in genome-scale networks is normally infeasible. For this reason, the Berge algorithm and the dual system approach by Ballerstein et al. used in the previous example become impractical. In contrast, with the MILP approach enumerating shortest EMs in the dual system as proposed here, calculation of small MCSs becomes possible.
To demonstrate this, we use the E. coli genome-scale network iAF1260 [24] that accounts for 1260 ORFs and defines the reversibilities of the included reactions. In total, this network comprises 1668 internal metabolites and 2382 reactions including 304 exchange reactions with the environment and 29 spontaneous reactions. The intervention goal for the MCSs to be computed is again to disable growth (biomass formation) when glucose is available as sole carbon source. The glucose uptake rate was fixed to = 10 mmol/(gDW⋅h) and the ATP maintenance requirement was set to the standard value of = 8.39 mmol/(gDW⋅h). Analogous to Suthers et al. [15] we considered a cell viable if it has a growth rate larger than μmin≥0.01⋅μmax = 0.0093 h−1. With these inhomogeneous conditions, the MCSs will thus correspond to synthetic reaction lethals as also computed by Suthers et al. [15], where full enumeration for MCSs up to size 3 was achieved (some MCSs of size 4 could also be determined). With glucose and oxygen available 152 reactions are disabled as suggested by the gene-regulatory model included in the iAF1260 reconstruction. A subsequent flux variability analysis revealed 991 blocked reactions in total and these were removed from the network. In addition, spontaneous and exchange reactions, of which 23 resp. 97 remain after removing blocked reactions, were not allowed to take part in any MCS. After removing the blocked reactions network compression by combining correlated reaction sets was again applied by which the (primal) network could be reduced to 562 metabolites and 936 (lumped) reaction subsets of which 816 can be knocked out. By using ALGO2 in the dualized system, for the first time it was possible to fully enumerate all synthetic (reaction) lethals of sizes 1 to 5 as shown in Table 2 yielding a total set of 2486 MCSs. Although the last iteration (MCSs with 5 knockouts) took several days all of them could be determined. Comparison of the runtimes of our MCSEnumerator implementation and of SL Finder (used in [15]) for the calculation of MCSs of size two and three indicates that our algorithm is more than 100 times faster therefore allowing full enumeration of synthetic reaction lethals also of size 4 and 5.
We also tested the homogeneous version of the above intervention problem, that is, we calculated the MCSs blocking growth without the additional constraints for ATP maintenance ( is a free flux), without restriction on glucose uptake and without the minimum threshold for the growth rate (all flux vectors with biomass production >0 have thus to be blocked). As expected, for we found less MCSs of size 1–5 (1933 in total) because the target polyhedron containing the target flux vectors was expanded leading to larger MCSs with more than 5 reaction deletions. We also observed that the computation of the MCSs in the homogeneous problem was much faster (∼17 hours) than for the inhomogeneous scenario (∼430 hours) indicating that inhomogeneous constraints may complicate the whole calculation procedure.
The following third example relates to a typical problem of finding rational intervention strategies for metabolic engineering purposes. We here focus on a biotechnologically relevant application, namely to let E. coli produce a biofuel (ethanol) from glucose. The intervention goal is thus to disable flux vectors with a low ethanol yield in E. coli (undesired behavior) while retaining the capability for both maintenance and growth of the bacterium under anaerobic conditions (desired functionality). This forms a constrained MCS problem. All cMCSs that fulfill the stated requirements will lead to obligatory coupling between growth and ethanol formation.
We used again the iAF1260 genome-scale network model of E. coli metabolism but this time with the oxygen uptake removed to establish anaerobic conditions on the network. As before, glucose is the only available carbon source. To study the effect of different capacities for substrate uptake, we considered two possible limits for the glucose uptake rate: = 10 mmol/(gDW⋅h) and = 18.5 mmol/(gDW⋅h). The latter value has been measured under anaerobic conditions where E. coli tends to exhibit higher substrate uptake rates [28]. The ATP maintenance requirement was set to ≥8.39 mmol/(gDW⋅h).
With these values in mind, we formulated the following intervention goal: the task is to identify cMCSs that guarantee a minimal ethanol yield of or, in a second scenario, of . In addition, a minimum growth rate of at least μmin≥0.001 h−1 was demanded.
With these inhomogeneous constraints we can now specify the target flux polyhedron containing all undesired network behaviors to be eliminated by the cMCSs:(22)(YEth/Glc(r) denotes the ethanol yield of the reaction rate vector r). The set of desired behaviors from which we want to keep at least some flux vectors is given by:(23)(The constraints due to anaerobic growth (e.g., oxygen uptake is zero) were not restated in (22) and (23).)
With these values, several linear programs were run in a preprocessing step to explore network capabilities. For and , the maximal ethanol yield is 2 (molecules ethanol per molecule glucose). The maximum growth rate is 0.1955 h−1 (for ) and 0.4954 h−1 (for ) if we want to achieve an ethanol yield of at least 1.4 ; these values drop to 0.1356 h−1and 0.4827 h−1, respectively, for a minimal ethanol yield of 1.8 . Hence, we can be sure that the set of desired behaviors is not empty.
We then computed the cMCSs. As described in the Methods section, the calculation of cMCSs (accounting for undesired and desired behavior) based on our approach requires to first compute the MCSs blocking the undesired behavior and to keep afterwards only those MCSs that admit the desired behavior. This test is done for each found MCS by solving a separate linear program (LP) which verifies whether the remaining network supports the desired behavior. To reduce the search space, blocked reactions for the network under desired ethanol production conditions were determined and removed in a preprocessing step using flux variability analysis [13]. In addition, 104 reactions were disabled for growth on glucose as suggested by the gene-regulatory model included in the iAF1260 reconstruction. The FVA then identifies 996 blocked reactions in total, which are removed from the network. Furthermore, the remaining 19 spontaneous and 94 exchange reactions were again not allowed to take part in the MCSs. The latter can be easily achieved by setting the upper bounds of the corresponding zpi and zni indicator variables to zero.
After network compression, the (primal) network could be reduced to 562 metabolites and 958 (lumped) reaction subsets of which 845 can be knocked out.
Note also that the disruption of glucose uptake or ATP maintenance are valid MCSs deleting all undesired behaviors but they violate for trivial reasons the desired functionality (growth not possible) and can thus not be contained in any valid cMCSs. Such reactions being essential for the desired flux space could also be identified at an early stage and then be excluded from the search space.
Table 3 shows the results for the computation of the (c)MCSs for this problem. As we considered two different maximal glucose uptake rates and two different minimal ethanol yields we obtained four scenarios. We were able to enumerate all cMCSs up to size 7 in all four scenarios within 21 hours. For each scenario, after calculating first the (unconstrained) MCSs up to size 7, each MCS was tested with a LP whether the solution space of (23) is non-empty (i.e., whether the MCS is a valid cMCS). These tests took less than 7 minutes running time (single-threaded; on the same computer that was used for MCS calculation) for each of the four scenarios. Hence, the LPs account only for a negligible part of the overall computational costs.
As can be seen in Table 3, only a fraction (between 1.3% and 6%) of the computed MCSs up to size 7 turned out to be valid cMCSs. However, a large number of several thousand cMCSs could eventually be computed for each scenario.
We then analyzed the cMCSs in more detail. A first observation in Table 3 is that in three of the four scenarios considered cMCSs were found comprising only three reaction deletions; whereas for the case with smaller glucose uptake and higher demanded ethanol yield (scenario 2 in Table 3) at least 5 reaction removals are required. Generally, it is intuitive that expanding the space of undesired flux vectors in (22) and reducing the space of desired solutions in (23) by increasing can lead to larger cMCSs since (i) a larger set of undesired flux vectors must be suppressed, and (ii) due to the reduced set of desired behaviors a smaller number of MCSs become admissible cMCSs. Hence, there is no cMCS in scenario 2 that is a subset of any cMCSs in scenario 1 in Table 3, but the other way around can occur. The same relationship exists between scenarios 3 and 4. Thus, the higher the yield that we want to guarantee by an intervention strategy, the larger is the required effort in terms of number of reaction knockouts.
The situation is different in the case of increasing . While the target flux polyhedron in (22) increases potentially demanding more cuts, the space of desired behaviors in (23) expands as well meaning that an MCS that was not a suitable constrained MCS in the case with smaller could now become a suitable cMCS. Hence, when increasing , some cMCSs of a given size might disappear whereas others may arise as new solutions. This is also reflected by the cMCSs of size three which are depicted in Figure 1. All these cMCSs block central pathways for glucose degradation. An essential cut (red cross in Figure 1) for all cMCSs is that of the glucose-phosphate isomerase blocking upper glycolysis. In addition, all the considered cMCSs block the Entner-Doudoroff pathway by either cutting the phosphogluconate dehydratase or the 2-keto-3-deoxyphosphogluconate aldolase (blue crosses in Figure 1). In addition, for scenario 1 (with the smaller values for and ), we have to cut one additional reaction out of 4 reactions of the pentose phosphate pathway (dark green crosses in Figure 1) whereas for scenarios 3 and 4 (whose two cMCSs of size three are identical) the third cut is given by the pyruvate-formate lyase reaction (light green cross in Figure 1). This result confirms that increasing (from scenario 1 to scenario 3) may remove existing cMCSs but also produce new ones.
The cMCSs for scenario 1 (the red cut, one of the two blue cuts and one of the four dark green cuts in Figure 1) also illustrate the difference between reaction and enzyme/gene cut sets. Since two of the four reactions with a green cross are catalyzed by the same enzyme (transketolase) knocking out the corresponding two genes (there are two different transketolases in E. coli) would actually cut two reactions at the same time for which the model predicts that E. coli can not grow anymore. Thus, only four of the eight cMCSs remain valid on gene basis. However, as already explained earlier, those effects can be taken into account based on gene-enzyme-reaction associations.
The fact that three reaction or gene knockouts may suffice to induce a high ethanol yield of more than 1.8 (scenario 4) is a surprising fact on its own. Previous work on computing intervention strategies for ethanol overproduction in a smaller (core) network of E. coli showed that more than three reaction knockouts would be required to ensure a large ethanol yield (see, e.g., [10]). Given the results with three knockouts made herein, this might be a bit confusing since much more inefficient pathways will exist in a genome-scale network which must all be blocked. However, similar as discussed above for a scenario with increased substrate uptake rates, a larger network may also have additional high-yield metabolic routes (allowing coupled biomass and ethanol synthesis) not contained in the smaller network which could ‘survive’ a cut set for blocking the low-yield pathways. We can thus conclude that genome-scale network models may reveal metabolic engineering strategies that are smaller than those found in small- or medium-scale subnetworks. Importantly, one always has to keep in mind that an MCS predicts an intervention purely from stoichiometric relationships. Whereas blockage of the undesired flux vectors can be guaranteed if the network structure is correct, it can not ensure that the remaining pathways will have the capacity to carry a flux that is large enough to fulfil the requirements of the desired flux vectors. In addition, unknown regulatory constraints may further reduce the space of desired behaviors by which some cMCSs may become invalid.
We mention here that two other intervention strategies with three knockouts for production of ethanol by E.coli were presented in [9]. However, these solutions ensure high ethanol yield only if the cell grows at maximal growth rate whereas our interventions are more stringent since they guarantee a high ethanol yield for any growth rate of the mutant.
Having exhaustively enumerated the cMCSs up to a given size enables one to analyze essential features and performance measures of all found intervention strategies by which eventually the optimal knockout strategy can be selected. Figure 2 shows exemplarily two such performance studies. Figure 2A displays for each cMCS of scenario 3 the relationship between (i) maximal growth rate, (ii) minimal (guaranteed) product yield (shown for maximal substrate uptake rate; the lower boundary for arbitrary substrate uptake rates still holds to be 1.4), and (iii) number of required reaction deletions (cut set size). It can be seen that most cMCSs (including those with the smallest size 3) achieve relatively low growth rates (lower than 0.1 h−1) and that in order to have a growth rate larger than 0.1 h−1 it is necessary to use cut sets with a least 6 knockouts. If higher growth rates and/or smaller cut sets are required the minimal product yield would have to be lowered.
Other performance measures of designed mutant strains can be studied as well. One such proposed measure is substrate-specific productivity (SSP) which is the product of the growth-rate and the product yield [29]. Figure 2B shows the SSP of all cMCSs computed for scenario 3. It can be seen that highest SSP values can only be achieved with cut sets of size 6 or 7. This illustrates again that a trade-off between number of knockouts and certain performance measures has sometimes to be made when eventually selecting an intervention strategy for implementation. Such a screen is greatly facilitated if all cut sets have been enumerated up to a certain size. More advanced screening methods for evaluating strain design strategies have been suggested in [30] and could readily be applied to calculated cMCSs.
As a technical note, it is not absolutely mandatory to have all MCSs (up to a maximal size) enumerated before running the LP checks for testing the “survival” of some desired flux vectors: these checks could be (independently) performed as soon as an MCSs has been found by the MILP solver. In fact, it is in principle possible to integrate the LP into the MILP so that the cMCSs are computed directly which offers the advantage that far fewer exclusion constraints need to be integrated while the enumeration proceeds. In practice, however, this approach showed a markedly inferior performance for the system studied here. One reason is that the LP adds further degrees of freedom to the solution space and leads to redundant solutions for the cMCSs which requires a more intricate control of the populate procedure to suppress these redundant solutions. Whether the integrated approach can be reformulated in a manner that facilitates a more efficient calculation of its cMCSs solutions is a potential topic for further investigation.
To summarize the results of this sub-problem, our algorithm enabled the enumeration of all reaction knock-out sets up to size 7 that lead to coupled ethanol and biomass synthesis in E.coli. To the best of our knowledge, this exceeds by far other attempts to enumerate such metabolic engineering strategies in large-scale networks.
If more computational capacity is available, one might try to find even larger cMCSs. However, the best knockout strategy to be implemented is likely to be contained among the up to 8819 smallest cMCSs found as the number of required interventions will be one (though not the only) key criterion when deciding for a concrete strain design.
One large-scale study to evaluate the growth-coupled production potential in E.coli has been presented by Feist et al. [29]. The aim was to identify strain designs based on reaction knockouts with a maximum production rate at optimal growth for a number of substrate/product pairs. This was achieved by first applying OptKnock [8] with a knockout limit of either three or five and then using the results in the initial population for OptGene which employs genetic programming as optimization method [31]. OptGene was then run with a time limit of one week to find additional strain designs with up to 10 knockouts. As underlying E.coli model the iAF1260 reconstruction [24] was taken and in order to reduce the search space the knockouts were restricted to a subset of about 150 reactions in the network. As minimum growth rate for the strains a limit of 0.1 h−1 was chosen and an ATP maintenance of 8.39 mmol/(gDW⋅h) required. Both glucose and oxygen uptake were limited to 20 mmol/(gDW⋅h). Given this setup it was possible to calculate strain designs for many substrate/product pairs but for some of them strains with only low productivity or even no strains with growth-coupled product synthesis were found.
Here we wanted to test the potential of our method for some of the intervention problems. We focused on the aerobic production of either fumarate or serine from glucose which both have a potential for high yield as calculated by FBA. However, growth-coupled strains for the production of fumarate only achieved 20% (5 knockouts, OptKnock) respectively 23% (7 knockouts, OptGene) of the theoretical maximum while for serine no growth-coupled strains could be identified in [29]. We therefore applied our approach to look for (additional) strain designs for these two configurations.
To demonstrate the power of our method in dealing with large-scale systems, we increased the search space drastically compared to [29] by allowing all reactions to be knocked out except for those that are either spontaneous or essential for the production condition. Since glucose is taken up under aerobic conditions, the same 152 reactions as for the calculation of the synthetic lethals above have also been removed. This results in 718 (fumarate) resp. 719 (serine) knockout candidates (compared to 150 candidates used in [29]). As the results in [29] suggested that growth coupling will be difficult for fumarate and serine production we chose a comparatively low minimal product yield of 0.5. This constraint together with the ATP maintenance requirement und the uptake limits was used to calculate MCSs that disable flux vectors with product yields below 0.5. Afterwards, only those (constrained) MCSs were kept that fulfil the minimal growth rate requirement.
For fumarate production, the MCSs up to size 7 were calculated (taking 13.6 h) from which 30 cMCSs (all of size 7) could be extracted. Applying those cMCSs would result in production strains exhibiting – at maximal substrate uptake rates – a guaranteed (minimal) fumarate yield between 0.71 and 0.89 corresponding to minimal production rates between 40.9% and 51.3% of the theoretical maximum of 34.68 mmol/(gDW⋅h) (note that the minimal yield for any substrate uptake rate is ensured to be 0.5 as demanded by the constraints for the desired flux vectors). As for the ethanol study, all these values are independent of the assumption of optimal growth. Likewise, in the case of serine production, the MCSs up to size 6 were calculated (taking 3.1 h) from which 140 cMCSs (all of size 6) could be extracted. These would result in strains with with a guaranteed serine yield between 0.71 and 0.91 (at maximal substrate uptake rate) corresponding to minimal production rates between 36.6% and 47.0% of the theoretical maximum (38.71 mmol/(gDW⋅h)). Hence, our results show that significantly larger fumarate production rates can be achieved with 7 knockouts than computed by OptGene. In case of serine where no suitable knockout strategy could be identified in [29], our method proves the existence of strain designs for coupled biomass and product synthesis and that 6 reaction knockouts would be theoretically sufficient to guarantee a serine yield of 47% of the theoretically maximal value. Moreover, tens of the smallest strain designs with 6 knockouts could be identified by our algorithm in a comparably fast way and larger ones could also be determined if desired.
In this work we presented MCSEnumerator, a new algorithmic approach to enumerate the smallest (c)MCSs up to a given size in genome-scale networks. This approach is based on a MILP problem calculating the shortest EMs in the dual representation of the metabolic network eventually yielding the smallest cMCSs. The whole procedure can be summarized by five steps:
With these five steps, MCSEnumerator provides a generic approach for enumerating smallest intervention strategies; one just has to plugin the corresponding matrices in equation (17) and can then start the calculation using ALGO2.
Apart from the combination of dualization and shortest EM calculation in step 3, another key development made herein is the improvement of the required sub-routine for computing shortest EMs (ALGO2) which is now based on a more efficient enumeration of feasible EMs with fixed size and which consequently makes use of available enumeration features of modern MILP solvers. Appropriate integration of such functionalities could also be useful to effectively solve other enumeration problems in the field.
Despite the fact that calculation of all (c)MCSs with our approach is slower compared to other approaches requiring EMs to be calculated in a first step, it has the advantage that the smallest (c)MCSs, which are often the most interesting ones, can be found first and that no EMs need to be calculated beforehand. This property renders (c)MCSs calculation feasible in genome-scale networks. Also, the number of elements in an MCS has no major impact on the performance as it would have in brute-force enumerations (that exhaustively test all reaction subsets) and as it has been observed also for several directed search algorithms.
The main drawback of using a MILP stems from the fact that constraints have to be continuously added to remove already found MCSs and their supersets from the solution space. Hence this method is bound to slow down with increasing number of constraints which explains the inferior performance when computing all MCSs. However, the shown application examples demonstrated that our approach is capable to compute hundreds of thousands of smallest MCSs and several thousand smallest constrained MCSs in genome-scale networks (Table 3) which has not been achieved before. The large set of smallest cMCSs should suffice to characterize the space of the most efficient intervention strategies from which, in metabolic engineering applications, the most promising ones can be selected, possibly by screening the cMCSs via certain performance parameters.
The algorithmic advantage of the presented approach lies thus in the possibility to quickly (compared to other approaches) calculate the smallest (c)MCSs with neither network size nor the number of elements in the (c)MCSs posing major challenges. With these results and due to the fact that the approach of (c)MCSs allows the setup of complex intervention problems in a flexible and convenient way, we expect that a large number of metabolic network studies can benefit from our conceived framework.
An interesting aspect for future work will be to investigate how far ALGO2 (the sub-routine used for shortest EM calculation) can be generalized to enumerate also other elementary sets arising in different contexts of computational biology (e.g., for calculating minimal intervention sets in signaling or regulatory networks [32]).
|
10.1371/journal.pntd.0001893 | Direct Interaction between EgFABP1, a Fatty Acid Binding Protein from Echinococcus granulosus, and Phospholipid Membranes | Growth and maintenance of hydatid cysts produced by Echinococcus granulosus have a high requirement for host lipids for biosynthetic processes, membrane building and possibly cellular and developmental signalling. This requires a high degree of lipid trafficking facilitated by lipid transporter proteins. Members of the fatty acid binding protein (FABP) family have been identified in Echinococcus granulosus, one of which, EgFABP1 is expressed at the tegumental level in the protoscoleces, but it has also been described in both hydatid cyst fluid and secretions of protoscoleces. In spite of a considerable amount of structural and biophysical information on the FABPs in general, their specific functions remain mysterious.
We have investigated the way in which EgFABP1 may interact with membranes using a variety of fluorescence-based techniques and artificial small unilamellar vesicles. We first found that bacterial recombinant EgFABP1 is loaded with fatty acids from the synthesising bacteria, and that fatty acid binding increases its resistance to proteinases, possibly due to subtle conformational changes induced on EgFABP1. By manipulating the composition of lipid vesicles and the ionic environment, we found that EgFABP1 interacts with membranes in a direct contact, collisional, manner to exchange ligand, involving both ionic and hydrophobic interactions. Moreover, we observed that the protein can compete with cytochrome c for association with the surface of small unilamellar vesicles (SUVs).
This work constitutes a first approach to the understanding of protein-membrane interactions of EgFABP1. The results suggest that this protein may be actively involved in the exchange and transport of fatty acids between different membranes and cellular compartments within the parasite.
| Echinococcus granulosus is the causative agent of hydatidosis, a zoonotic infection that affects humans and livestock, representing a public health and economic burden in many countries. Since the parasites are unable to synthesise most of their lipids de novo, they must acquire them from the host and then deliver them by carrier proteins to specific destinations. E. granulosus produces in abundance proteins of the fatty acid binding protein (FABP) family, one of which, EgFABP1 has been characterised at the structural and ligand binding levels, but it has not been studied in terms of the mechanism of its interaction with membranes. We have investigated the lipid transport properties and protein-membrane interaction characteristics of EgFABP1 by applying biophysical techniques. We found that EgFABP1 interacts with membranes by a mechanism which involves direct contact with them to exchange their cargo. Given that the protein has been found in the secretions of the parasite, the implications of its direct interactions with host membranes should be considered.
| Hydatidosis is a highly pathogenic infection with an almost global incidence caused by the larval stage (metacestode) of the cestode Echinococcus granulosus. In endemic areas it has serious health effects on humans, livestock and wildlife, representing a major public health and economic burden in many countries [1]–[3]. Echinococcus species, as do other tapeworms of mammals, require two hosts to complete their life cycle. The E. granulosus eggs containing the infective oncosphere are shed in the faeces of wild and domestic carnivores that are the definitive hosts harbouring the dwarf adult tapeworms. Once a suitable intermediate host ingests the eggs, they hatch and the oncosphere is released, escaping from the intestine to establish hydatid cysts in liver and lungs. The cyst produces thousand of protoscoleces, each of which can progress to the adult form when ingested by the definitive host [4], but it is the hydatid cysts in intermediate hosts that cause significant pathology and death. Hydatid disease in humans is highly pathogenic and is particularly difficult to treat successfully, especially so when cysts develop and proliferate in the lungs.
Fatty acid binding proteins (FABPs) are small proteins (14–15 kDa) that bind non-covalently to hydrophobic ligands, mainly fatty acids (FA) and retinoids. FABPs are confined to the interior of the synthesising cells, the only known exceptions to this being in nematodes [5], [6]. Several tissue-specific FABP types have been identified in vertebrates, each named after the tissue in which they are predominantly expressed, and have also been given numeric designations [7]. In mammals they are implicated in intracellular uptake, storage and transport of FAs in lipid metabolism and membrane building, as well as protection from the membrane-disruptive effects of free long chain FAs [8]. In addition, the non-FA-binding retinoid-binding isoforms contribute to regulation of gene expression [9]. However, the precise function of each FABP type remains poorly understood, but sub-specialization of functions is suggested by the tissue-specific and temporal expression, in addition to ligand preferences [10]. Despite very similar tertiary structures, FABPs have been found to interact with membranes in different ways that might reflect how they acquire and deliver their cargoes. The fluorescence-based biophysical approaches used for this have shown that most FABPs from mammals (adipocyte FABP, intestinal FABP, heart FABP, keratinocyte FABP, myelin FABP, etc.) and one from Schistosomes (Sj-FABPc) exhibit a collisional mechanism of ligand exchange, meaning that they interact by direct contact with a membrane in ligand transfer. In contrast, only liver FABP and cellular retinol binding protein II from mammals transfer ligands in a diffusional mechanism, meaning that transfer occurs without requiring direct contact between protein and membrane but through release of ligand into the aqueous phase followed by its intercalation into the membrane. Proteins like liver FABP may therefore be more involved in lipid storage and regulation in the cytoplasm rather than in direct transport of FAs [8], [11], [12].
FABPs of parasitic platyhelminths are interesting because these parasites are unable to synthesise most of their own lipids de novo, in particular long-chain FAs and cholesterol [13], [14]. Such lipids must therefore be acquired from the host, and then delivered by carrier proteins to specific destinations within the parasite. Whether they are involved extracellularly in lipid acquisition from, or delivery to, host cells, remains to be seen. It is noteworthy that EgFABP1 has been found in hydatid cyst fluid and in protoscolex secretions [15], [16]. A final reason for interest in FABPs is their potential role in drug delivery and the fact that they have been assayed as vaccine candidates [17]–[23].
EgFABP1 is considered to be a member of the heart FABP subfamily [24], [25], whose members are believed to be involved in lipid oxidation processes [8]. The ligand-binding properties of EgFABP1 have been investigated by the displacement of cis-parinaric acid by a set of hydrophobic ligands [26], and its crystal structure reveals the 10-stranded β-barrel fold typical of the family of intracellular lipid-binding proteins [27].
The objective of this study was to investigate the lipid transport properties and protein-membrane interaction characteristics of EgFABP1. We characterise the biophysical properties of the protein in a number of ways, and show that the protein exchanges FAs through a collisional, direct contact, mechanism with acceptor membranes, indicating that it may indeed be involved in FA dynamics within the parasite, but that it may also engage in direct, non-specific interactions with host cell membranes.
The cDNA encoding EgFABP1 (UniProtKB/Swiss-Prot Q02970) was subcloned into pET11b. The expression of the protein was carried out in E. coli BL21(DE3) by induction with 0.4 mM isopropyl-beta-D-thiogalactoside for 3 hours at 37°C in Luria Bertani medium in presence of 100 µg/mL of ampicillin. Cells were lysed by sonication and the lysate clarified by ultracentrifugation (25 min, 61700× g, 4°C). Following clarification, the supernatant was subjected to salting out incubating the protein for 2 hours at 4°C with 0.5 volume of a saturated ammonium sulphate solution. After centrifugation, the obtained protein solution was applied into a size exclusion chromatographic column (Sephadex G-50, Pharmacia Biotech Inc.). The fractions containing EgFABP1 were subsequently subjected to ionic exchange chromatography employing a MonoQ column (Pharmacia Biotech Inc.) in order to remove nucleic acids contamination. Delipidation was carried out using a Lipidex 1000 column (Sigma) at 37°C in a high ionic strength buffer (10 mM phosphate (K2HPO4 6 mM+KH2PO4 4 mM), 1 M KCl).
As an approach for studying binding preferences of EgFABP1, the lipid moiety of recombinant non-delipidated EgFABP1 was extracted according to Bligh & Dyer's method [28] and analysed on a TLC plate using a mobile phase for resolving neutral lipids (hexane∶diethyl-ether∶acetic acid at 80∶20∶1, v∶v∶v). The FA composition of EgFABP1 lipid fraction was analysed by GC of their methyl esters derivatives methylated with BF3-Methanol according to the method described by Morrison & Smith [29], employing an HP 6890 device Hewlett Packard) as described previously by Maté et al. [30].
In order to analyse possible conformational changes between apo- and holo- forms, EgFABP1 was subjected to limited proteolysis experiments. The protocol was a modification of that described by Arighi et al. [31]. Briefly, clostripain (ArgC, Sigma) was activated by preincubation in 10 mM phosphate (K2HPO4 6 mM+KH2PO4 4 mM), 150 mM KCl, pH 7.4 and 1 mM DTT for 2 hours. Prior to digestion, delipidated EgFABP1 (0,5 mg/ml) was incubated for 30 min with either myristic acid, palmitic acid, stearic acid or oleic acid in ethanol (4∶1 mol∶mol ligand∶protein) to obtain holo-forms. As a control of the FA solvent used, an equal volume of ethanol was added to the apo-form. Additional 15 min incubation with 1 mM DTT was carried out previous to the addition of the protease. At fixed intervals, samples were collected and frozen for subsequent analysis by SDS-PAGE. SDS-PAGE was carried out according to Schägger and von Jagow [32] in 16.5% acrylamide Tris-Tricine. After Coomassie Blue staining digital images were collected employing an ImageQuant 350 device (GE Healthcare).
Fatty acid binding to EgFABP1 was assessed employing a fluorescent titration assay [33]. Briefly, 0,5 µM anthroyloxy-fatty acid (AOFA, Molecular Probes) was incubated at 25°C for 3 min in 40 mM Tris, 100 mM NaCl, pH 7.4 buffer (TBS) with increasing concentrations of EgFABP1. The AOFAs employed for binding assays were 12-(9-anthroyloxy)stearic acid (12AS) and 16-(9-anthroyloxy)palmitic acid (16AP). Fluorescence emission at 440 nm was registered after excitation at 383 nm in a Fluorolog-3 Spectrofluorometer (Horiba-Jobin Yvon). An exact equilibrium n-sites binding model was fitted to fluorescence data (using Microcal ORIGIN software) as previously described [34].
For AOFA transfer experiments, small unilamellar vesicles (SUVs) were prepared by sonication and ultracentrifugation as described previously [35]. The standard vesicles were prepared to contain 90 mol % of egg phosphatidylcholine (EPC) and 10 mol % of N-(7-nitro-2,1,3-benzoxadiazol-4-yl)-phosphatidylcholine (NBD-PC), which served as the fluorescent quencher. To increase the negative charge density of the acceptor vesicles, either 25 mol % of phosphatidylserine (PS) or cardiolipin (CL) was incorporated into the SUVs in place of an equimolar amount of EPC. Vesicles were prepared in TBS except for SUVs containing CL which were prepared in TBS with 1 mM EDTA. SUVs containing 64 mol % EPC, 10 mol % egg phosphatidylethanolamine (EPE), 25 mol % CL and 1 mol % dansyl-phosphatidylethanolamine (DPE) were prepared in 20 mM Tris, 0.1 mM EDTA, pH 7.4 for protein-membrane interaction assays.
Large unilamellar vesicles (LUVs) of EPC were prepared (1 mM in phospholipids) by extrusion through polycarbonate membranes of 100 nm pore diameter (Avestin Inc., Ottawa, Canada) as described previously [35]. All lipids were purchased from Avanti Polar Lipids.
Ligand partition between the protein and NBD-containing SUVs was determined by measuring AOFA fluorescence at different protein∶SUVs ratios obtained by adding SUV to a solution containing 10 µM EgFABP1 and 1 µM 12AS in buffer TBS at 25°C [36]. The relative partition coefficient (KP) was defined as:(1)Where [Ligand-SUV] and [Ligand-FABP] are the concentration of AOFA bound to membrane and EgFABP1, respectively, and [FABP] and [SUV] are the concentration of protein and vesicles. The decrease in AOFA fluorescence as a function of SUV is related to KP by(2)Where Frel, [SUV], [FABP], Kp, a and b are the relative fluorescence, the molar concentration of SUV, the molar concentration of EgFABP1, the partition constant and fitting parameters, respectively [37]. The partition coefficient was used to establish AOFA transfer assay conditions that ensure essentially unidirectional transfer, as detailed below.
A Förster Resonance Energy Transfer assay was used to monitor the transfer of 12AS from EgFABP1 to acceptor model membranes as described in detail elsewhere [11], [33], [38]. Briefly, EgFABP1 with bound 12AS was mixed at 25°C with SUVs, prepared as above, using a stopped-flow RX2000 module (Applied Photophysics Ltd.) attached to the spectrofluorometer. The NBD moiety is an energy transfer acceptor of the anthroyloxy group donor; therefore, the fluorescence of the AOFA is quenched when the ligand is bound to SUVs that contain NBD-PC. Upon mixing, transfer of AOFA from protein to membrane is directly monitored by the time-dependent decrease in anthroyloxy group fluorescence. Different SUVs and buffer compositions were employed in order to analyse the ligand transfer mechanism. Transfer assay conditions were 15∶1 mol∶mol EgFABP1∶AOFA ratio. SUVs were added ranging from 1∶10 mol∶mol to 1∶40 mol∶mol EgFABP1∶SUVs. Controls to ensure that photobleaching was eliminated were performed prior to each experiment, as previously described [38]. Data were analysed employing SigmaPlot and all curves were well described by an exponential decay function. For each experimental condition within a single experiment, at least five replicates were done.
To analyse the putative association of EgFABP1 with vesicles, an assay that exploits the well known membrane-interactive properties of cytochrome c was employed. The binding of cytochrome c to acidic membranes can be monitored by using a resonance energy transfer assay [39] in which the dansyl fluorescence of DPE-labelled SUV is quenched upon binding of cytochrome c, which contains the heme moiety quencher. Competition of EgFABP1 with cytochrome c for binding to SUVs was determined by the relief of cytochrome c-related quenching of the dansyl fluorescence. In a final volume of 200 µl, 0–48 µM EgFABP1 was added to 15 µM SUV in 20 mM Tris.HCl/0.1 mM EDTA, pH 7.4. After a 2 min equilibration, fluorescence emission at 520 nm was measured (λex = 335 nm). Cytochrome c (Sigma) was then added (1 µM final concentration), and the mixture equilibrated an additional 2 min period before monitoring again fluorescence emission at 520 nm. In the absence of bound FABP, the dose-dependent quenching of dansyl fluorescence is observed. An inhibition of cytochrome c-dependent quenching is interpreted as evidence for EgFABP1 interaction with SUVs, i.e., EgFABP1 prevention of subsequent cytochrome c interaction with the bilayer.
This assay was performed in order to determine which lipid classes bind to EgFABP1 in a cellular environment. Despite E. coli's cytoplasm not being the natural environment of EgFABP1, this approach could contribute to the assignment of the protein's natural ligands as it analyses the preference of EgFABP1 for different hydrophobic compounds present in the bacterial cytoplasm. TLC analysis showed that only FAs were bound to the recombinant protein (data not shown). Among them, palmitic acid (16:0) and stearic acid (18:0) are important ligands, although myristic (14:0), pentadecanoic (15:0), palmitoleic (16:1 n-7), 7-hexadecenoic (16:1 n-9), oleic (18:1 n-9), vaccenic (18:1 n-7), and linoleic acid (18:2) were also detected (Figure 1). The latter may come from culture media, as E. coli is not able to synthesise polyunsaturated FAs, at least during log-phase growth [40], [41]. The distribution of FAs bound to FABP may be related to the relative abundance of each of them in E. coli, and it correlates well with the reported FA composition of E. coli grown in equivalent conditions [41]. As in previous in vitro displacement of fluorescent ligand studies where palmitic and stearic acids are among those that produce moderate displacement percentages (>50%) [26], this experiment shows that EgFABP1 is able to bind many FAs of different chain length and degree of insaturation. In addition, in agreement with these results, the crystal structure of recombinant EgFABP1 revealed an electronic density inside the cavity, which was interpreted as being palmitic acid [27]. We therefore proceeded to investigate protein:membrane transfer using fluorophore-tagged fatty acid analogues.
Partial proteolysis can provide information related to conformational changes in proteins since this technique may reveal the differential exposure of proteolytic sites in apo and holo forms. We analysed the peptide pattern obtained by digestion of EgFABP1 in its apo- or different holo-forms with Clostripain (ArgC). The FAs selected, following to the analysis of ligands bound to recombinant EgFABP1 (Figure 1), were myristic, palmitic, stearic and oleic acids. The enzyme hydrolyses the polypeptide chain at the C-terminal end of arginine residues. Qualitative differences were evident between apo-EgFABP1 and the different complexes (Figure 2). Results show that binding of FAs gives EgFABP1 significant relative protection against cleavage. After 5 minutes of proteolysis the apo-protein shows several bands corresponding to proteolytic fragments, while the holo-forms show mainly the band corresponding to full-length EgFABP1 and less intense bands corresponding to proteolytic fragments (Figure 2A). This suggests that ligand-binding results in a different exposure of proteolytic sites. It is interesting to note that after 16 hours of proteolysis the holo-proteins do not seem to be further proteolysed while the apo-protein is almost completely degraded (Figure 2B). Previous results obtained for other members of the family of FABPs have suggested that binding of ligands involves conformational changes, especially on the portal region of FABPs [10], [31], [42]. Furthermore, in silico simulations show that, upon ligand binding, subtle conformational changes can be detected inside the cavity, in the surface and in the portal region of EgFABP1 (Esteves, unpublished data). These changes could make cleavage sites less accessible to the protease.
As an additional approach to investigate conformational changes between apo- and holo-protein, we analysed the circular dichroism (CD) spectra of EgFABP1 in the far (200–250 nm) and near (250–320 nm) UV regions. Two different ligands were employed for the generation of holo-EgFABP1: palmitic and oleic acid. Results indicated that the far-UV spectra of apo- and the two holo-forms did not show appreciable differences as can be seen in Figure S1. These data could be interpreted to show that no significant changes in overall secondary structure content are caused by ligand binding. On the other hand, the near-UV CD spectra (Figure S1) showed differences upon ligand binding, especially with oleic acid, indicating a likely alteration in the environment of aromatic residues resulting from proximity to ligand and/or a change in the conformation of the protein. So, ligand binding to EgFABP1 could elicit a change in the tertiary structure of the protein that could be correlated to the relative resistance of the holo form to proteolytic attack observed in the previous experiment.
In preparation for experiments on the interaction of EgFABP1 with phospholipid vesicles, binding experiments were performed using fluorescent analogues of stearic and palmitic acids, 12AS and 16AP, respectively. Anthroyloxy probes are useful indicators of binding site characteristics because their spectral properties are environment-sensitive. These probes usually have very low fluorescence intensity in buffer, which becomes dramatically enhanced upon interaction with a FABP [43]. 12AS showed a large increase in fluorescence emission accompanied by a substantial blue shift upon binding to EgFABP1. On the other hand, 16AP's fluorescence was surprisingly decreased when bound to EgFABP1, but also accompanied by a distinct blue shift in emission (Figure 3). This blue-shift indicates that the fluorophore had entered an apolar environment, almost certainly the hydrophobic binding pocket rather than a superficial, non-specific site of the protein. Following addition of artificial 100 mol % phosphatidylcholine LUVs to the 16AP:EgFABP1 complex, the intensity of fluorescence emission increased, indicating that the quenching of 16AP's fluorescence emission was reversed upon transfer to the different, lipidic, environment of the vesicles. In both cases (12AS and 16AP) the titration described curves that reached saturation, in accordance to a ligand binding phenomenon consistent with 1∶1 binding, with a Kd of 0.12±0.02 µM for 12AS, and 0.013±0.006 µM for 16AP. 12AS was chosen as a ligand for the following analysis of transfer kinetics due to its fluorescence emission characteristics when bound to protein being more typical of that observed in other studies on protein to membrane transfer [37], [44], [45]. However, the quenching effect observed with 16AP will be very useful to analyse FA transfer between EgFABP1 and other proteins that show a typical increase of AOFA fluorescence upon binding. Regarding this, another lipid binding protein from E. granulosus which is very abundant in the hydatid fluid, Antigen B, has been investigated in its binding properties, showing that it binds 16AP with a 30-fold fluorescence enhancement of the probe [46].
The apparent partition coefficient that describes the relative distribution of 12AS between EgFABP1 and EPC-SUVs was determined by adding SUVs containing NBD-PC to a solution of 12AS:EgFABP1 complex. As a result of this experiment, a KP value of 0.48±0.23 was obtained employing Eq. 2 (see Materials and Methods), which indicates that there is preference of the AOFA for the phospholipid membranes.
In a collisional transfer, the limiting step is the effective protein-membrane interaction, and the rate increases as the acceptor membrane concentration increases. In a diffusional mechanism in which the rate limiting step is the dissociation of the protein-ligand complex, no change in rate is observed [33], [37], [38], [44], [45], [47]. The values of Kd and KP were used to set the conditions for the transfer assay. The proportion of protein and ligand was such that less than 1% of AOFA remained free in the preincubation solution. On the other hand, KP value was used to calculate the final concentrations of protein and SUVs for which unidirectional transfer prevailed. Figure 4 shows that when constant concentrations of the EgFABP1-12AS donor complexes were mixed with increasing concentrations of EPC-SUV, the 12AS transfer rate from EgFABP1 to EPC-SUV increased proportionally to vesicle concentration in the SUV: EgFABP1 ratios (10∶1 to 40∶1) tested. In these conditions, the increase in transfer rate ranged from 0.04±0.01 sec−1 to 0.12±0.03 sec−1. These results strongly suggest that the FA transfer from EgFABP1 occurs via a protein-membrane interaction rather than by simple aqueous diffusion of the free ligand.
Considering the hypothesis that FA transfer from EgFABP1 occurs by collisional contact with an acceptor membrane, changes in membrane properties should modify the transfer rate. If the mechanism relied on aqueous diffusion alone, then the characteristics of acceptor membranes should be irrelevant to the transfer rate, since the rate-determining step in such a transfer process (ligand dissociation into the aqueous phase) is a physically and temporally distinct event from processes involving the membrane. Figure 5 shows that 12AS transfer rate from EgFABP1 to membranes increased when acceptor membranes contained 25% of negatively charged phospholipids (PS or CL). In agreement with the behaviour we have previously observed for collisional mammalian FABPs [38], [44], [45], [47], EgFABP1 shows a large increase in FA transfer rate to CL vesicles compared with zwitterionic vesicles. To investigate further the effect of negative charge of the acceptor vesicles on the FA transfer mechanism from the protein, we analysed the modification of transfer rates with increasing concentrations of negatively charged acceptor vesicles. The rate of FA transfer from EgFABP1 always, and independently of the net charge of the vesicles, showed the classical proportional increase in transfer rate with acceptor concentration (Figure 6).
Transfer of 12AS from EgFABP1 to membranes was examined as a function of increasing concentrations of NaCl. The results show that an important increase in transfer rate from EgFABP1 to neutral membranes was detected with increasing ionic strength of the aqueous phase (Figure 7A). It is generally thought that electrostatic interactions at surfaces are diminished and hydrophobic interactions are stimulated as a function of increasing ionic strength. The effect of ionic strength on the rate of AOFA transfer from EgFABP1 to zwitterionic vesicles suggests that the elimination of electrostatic interactions by salt shielding is compensated by an increase in hydrophobic interactions.
When negative charge was added to the acceptor lipid vesicles, a drastic decrease was observed at high salt concentrations (Figure 7B). As shown in Figure 5, EgFABP1 exhibited approximately a 60-fold increase in AOFA transfer rate to CL vesicles compared with EPC vesicles at low ionic strength. Upon increasing the ionic strength, a marked decrease from the very high values observed at low ionic strength was found (Figure 7B). This suggests a masking of electrostatic interactions, which play a very important role at low ionic strength, caused by the high salt content of the buffer.
FA transfer experiments suggest that the interaction of EgFABP1 with membranes is sensitive to surface charge density. As cytochrome c is known to interact as a peripherally associating protein with acidic membranes [48], we analysed the ability of EgFABP1 to compete with cytochrome c for binding to membranes containing CL. Cytochrome c quenches dansyl fluorescence in a concentration-dependent manner (ref. [49] and Figure 8). Results show that preincubation of CL-containing vesicles with EgFABP1 was effective in preventing cytochrome c binding in a concentration-dependent manner (Figure 8). When EgFABP1 (48 µM) was added to CL-containing SUVs, the dansyl fluorescence was twice that obtained in the absence of EgFABP1 and with 1 µM cytochrome c.
We show that recombinant EgFABP1 is able to bind FAs of different chain lengths from E. coli, mainly palmitic and stearic acids. This is clearly an incomplete inventory of ligands that it may transport in the parasite, but it does illustrate the propensity of the protein to bind FAs when exposed to an environment rich in a wide range of small hydrophobic compounds. The analysis of the natural ligands bound by EgFABP1 in the parasite environment is currently being undertaken in our laboratory.
Our main finding in this work was that the protein engages in a collisional mechanism in ligand transfer, as do various FABP isoforms from mammals, and one from Schistosomes, that have been investigated in this way [8], [11], [12]. This involvement of direct contact between protein and membrane for this transfer was found by altering electrostatic and hydrophobic conditions in the transfer experiments. The results indicated that the interaction event is mediated by both charge and hydrophobic factors, and it would seem reasonable that the initial interaction is ionic, between the protein and charged phospholipid headgroups, followed by direct, transient hydrophobic interaction with the apolar layer of the membrane. The interaction of the protein with membranes has also been demonstrated by the competition with cytochrome c for membrane binding.
The tertiary structure of EgFABP1 is virtually superimposable on FABPs that engage in collisional transfers [27], in which the two alpha-helices adjacent to the portal of ligand entry in FABPs are important in engaging contact with membranes [38]. It may be no coincidence that EgFABP1 has, like these other collisional FABPs, a prominent pair of bulky hydrophobic amino acid sidechains (Phe27, Val28) extending into solvent from helix II, immediately adjacent to the portal. Such a ‘sticky finger’ could attract and orient ligand for entry into the protein, or be involved in the protein's interaction with membranes or other proteins [50].
Our results suggest that EgFABP1 is likely to be an active participant in the transport and exchange of lipids in vivo, which could involve uptake of FAs directly from, and delivery to, membranes within the parasite, potentially resourcing the developing protoscoleces within the hydatid cysts. This might also be the case for Antigen B, which belongs to a new family of hydrophobic ligand binding proteins of cestodes and has been proposed as a lipoprotein involved in lipid trafficking [46], [51]. Furthermore, our proteolysis experiments with EgFABP1 and the analysis of CD spectra of apo- and holo-forms indicated that ligand binding would induce a conformational change in the protein. Such a change might modify the mechanism of interaction of EgFABP1 with membranes to facilitate upload or download of their cargo. A conformational change could also function as a signal to target the protein to different destinations, as has been suggested for other members of the FABP family [52], [53]. The possibility that it also interacts with host cell membranes is more contentious, particularly since EgFABP1 does not have a secretory leader peptide, as is also the case for FABPs from any group of animals other than nematodes [5], [6], so should be confined to the interior of cells. However, if EgFABP1 appears in cyst fluid in vivo and in excretion/secretion products of protoscoleces [15], [16] (but not as a result of cell damage during fluid collection or imperfect culture conditions in the collection of excretion/secretion products) then the possibility that it does interact with host cells beyond the cyst wall must be considered. Host proteins are known to cross hydatid cyst walls [15], so it is conceivable that this permeability (if a unidirectional transfer system is not in operation) could mean that EgFABP1 leaves the cyst to interact with host membranes for return to the parasite, or to deliver lipids to host tissues for immunomodulation. These hypotheses remain to be tested. In this regard, future studies should also include protein interaction analysis with membranes that mimic parasite and host composition.
This work is a first approach to understand the functional properties of EgFABP1 and constitutes the basis for further expanding our knowledge about this protein. This has been the case for other members of the FABP family, where this kind of studies has contributed to the understanding of the mechanisms of ligand transfer to membranes, protein-membrane and protein-protein interactions [8], [54].
|
10.1371/journal.pgen.1005679 | A Comprehensive Genomic Analysis Reveals the Genetic Landscape of Mitochondrial Respiratory Chain Complex Deficiencies | Mitochondrial disorders have the highest incidence among congenital metabolic disorders characterized by biochemical respiratory chain complex deficiencies. It occurs at a rate of 1 in 5,000 births, and has phenotypic and genetic heterogeneity. Mutations in about 1,500 nuclear encoded mitochondrial proteins may cause mitochondrial dysfunction of energy production and mitochondrial disorders. More than 250 genes that cause mitochondrial disorders have been reported to date. However exact genetic diagnosis for patients still remained largely unknown. To reveal this heterogeneity, we performed comprehensive genomic analyses for 142 patients with childhood-onset mitochondrial respiratory chain complex deficiencies. The approach includes whole mtDNA and exome analyses using high-throughput sequencing, and chromosomal aberration analyses using high-density oligonucleotide arrays. We identified 37 novel mutations in known mitochondrial disease genes and 3 mitochondria-related genes (MRPS23, QRSL1, and PNPLA4) as novel causative genes. We also identified 2 genes known to cause monogenic diseases (MECP2 and TNNI3) and 3 chromosomal aberrations (6q24.3-q25.1, 17p12, and 22q11.21) as causes in this cohort. Our approaches enhance the ability to identify pathogenic gene mutations in patients with biochemically defined mitochondrial respiratory chain complex deficiencies in clinical settings. They also underscore clinical and genetic heterogeneity and will improve patient care of this complex disorder.
| Mitochondria play a crucial role in ATP biosynthesis and comprise proteins encoded in both the nuclear and mitochondrial genomes. Although more than 250 mitochondrial disease-causing genes have been reported, the exact genetic causes in patients remain largely unknown. Here, we aimed to provide further insights into the pathogenic mechanisms of mitochondrial disorders. We investigated the genes encoded in the nuclear and mitochondrial genomes using comprehensive genomic analysis in 142 patients with mitochondrial respiratory chain complex deficiencies. We identified 3 novel disease-causing mitochondria-related genes (MRPS23, QRSL1, and PNPLA4) as well as other disease-causing genes and novel pathogenic mutations in known mitochondrial disease-causing genes. All pathogenic mutations in this study are validated by genetic and/or functional evidence. Our findings, including the achievement of firm genetic diagnoses for 49 of 142 patients (34.5%), were higher than the general diagnosis rate of approximately 25% and demonstrated the value of comprehensive genomic analysis. Accordingly, we have shed light on the genetic heterogeneity underlying mitochondrial disorders.
| Human oxidative phosphorylation (OXPHOS) disease has the highest incidence among congenital metabolic disorders characterized by a biochemical respiratory chain complex deficiencies and is thought to occur at a rate of 1 in 5,000 births[1]. No more than 15–30% of pediatric diseases diagnosed as mitochondrial disorders show mitochondrial DNA (mtDNA) abnormalities[2,3]; the remaining cases occur because of defects in genes encoded in the nucleus. A certain amount of nuclear-encoded gene products are present in the mitochondria, and roughly 1,500 are thought to play important roles in mitochondrial function[4,5].
It is particularly difficult to diagnose patients with OXPHOS disease at the molecular level because of the massive numbers of potentially involved nuclear genes and genes not yet linked to human disease. Therefore, identification of the causative genes and an understanding of the pathogenic mechanisms of OXPHOS disease remain unsolved challenges.
Recent studies[6,7] have shown that heterogeneous genetic backgrounds as well as genes previously not linked to mitochondrial functions or localization are associated with this disease. However, because of phenotypic and locus heterogeneity, only a fraction of patients has been identified to date. Limitations in target resequencing have motivated us to apply a comprehensive genomic analysis for more accurate molecular diagnosis and for the identification of novel causative genes.
Here, we aimed to determine whether a comprehensive genomic analysis approach could be used to reveal the broad spectrum of genetic background of the disease[8]. One hundred and forty-two unrelated individuals with displayed childhood-onset mitochondrial respiratory chain complex deficiencies were selected. We applied long-range polymerase chain reaction (PCR)-based whole mtDNA sequencing, whole exome sequencing (WES), and high-density oligonucleotide arrays to identify single-nucleotide variants (SNVs), small insertions or deletions (indels), and chromosomal aberrations for comprehensive genomic analyses.
In this study, 142 patients with childhood-onset mitochondrial respiratory chain complex deficiencies were enrolled and subjected to comprehensive genomic analyses (detailed clinical characteristics are described in S1 Table). A schematic workflow of these analyses is shown in Fig 1. Comprehensive genomic analyses included three approaches: (i) amplicon-based whole mtDNA sequencing for pathogenic mutations and large duplications/deletions, (ii) WES for pathogenic mutations in nuclear DNA, and (iii) high-density oligonucleotide arrays for chromosomal aberrations. The prioritized variants derived from each approach are described below.
After comprehensive genomic analysis shown in Fig 1, rare variants were filtered out and prioritized on the basis of the strategy described below. For mtDNA variants, we targeted variants confirmed and reported in MITOMAP[9]. Exome sequencing covered 89% (ranged: from 70%–to 98%) of the targeted bases, with more than 20-fold coverage. Detailed sequence statistics is shown in S2 Table.
The precise strategy for WES variant prioritization is shown in S2 Fig. We evaluated our prioritization pipeline to validate whether it could feasibly enrich known OXPHOS disease-causing genes or mitochondria-related genes (S3 Fig). Known OXPHOS disease-causing genes were clearly enriched in disease cases, whereas no prioritized genes were detected in healthy controls. Compared with healthy controls, mitochondria-related genes also exhibited a 1.64-fold enrichment. No enrichment was observed in randomly selected genes. These results suggest that mitochondria-related gene enrichment is caused by unidentified causative genes.
To analyze chromosomal aberrations, we focused on rather large (>100 Kb) deletions or duplications. For prioritizing candidate aberrations, we filtered out deleted or duplicated regions found in the 524 in-house controls and manually curated the pathogenicity of the aberrations by referring to the OMIM, DGV, and DECIPHER databases.
A breakdown of the 142 patients according to prioritized variants is shown in Fig 2. Of the 142 patients with mitochondrial respiratory chain complex deficiencies, 102 (71.8%) harbored at least 1 prioritized mtDNA mutation, nuclear gene mutation, or chromosomal abnormality. Ten (7.0%) patients harbored mtDNA mutations, including one large deletion (S4 Fig). In 29 patients (20.4%), firm molecular diagnoses were made in 20 genes previously linked to mitochondrial disorders. We newly confirmed 3 mitochondria-related genes (MRPS23, QRSL1, and PNPLA4) as causative genes of mitochondrial respiratory chain complex deficiencies. Three patients (2.1%) harbored mutations in genes known to cause monogenic diseases (MECP2 and TNNI3). Intriguingly, 4 patients (2.8%) had pathogenic chromosomal deletions previously linked to other disorders (6q24.3-q25.1, 22q11.21, and 17p12) but not linked to mitochondrial respiratory chain complex deficiencies.
In 53 (37.3%) patients, we identified and designated candidate genes or loci as prioritized variants of unknown significance (pVUS) because these variants have possibilities to be pathogenic but have insufficient evidence to support a disease linage. The current lack of functional validation for linking these genes with mitochondrial disorders led us to classify these variants as inconclusive with respect to pathogenicity (S3, S4 and S5 Tables). The remaining 40 (28.2%) patients lacked prioritized nuclear variants, mtDNA variants, and chromosomal abnormalities.
Twenty-two genes were prioritized in 31 patients (Table 1). Of these, 29 patients harbored 41 disease-causing mutations in 20 genes known to cause OXPHOS disease: ACAD9, BOLA3, COQ4, COX10, EARS2, ECHS1, GFM1, GTPBP3, KARS, MPV17, NDUFA10, NDUFAF6, NDUFB11, NDUFS4, RARS2, RRM2B, SCO2, SUCLA2, TAZ, and TUFM. All such mutations were confirmed through Sanger sequencing and haplotype phasing. In particular, 8 patients had homozygous mutations, 19 had compound heterozygous mutations, and 2 had hemizygous mutations. Of the 41 mutations, 37 were novel and 4 were reported as pathogenic in the Human Gene Mutation Database[10] (HGMD, professional version 2013.10).
BOLA3, which plays an essential role in iron–sulfur cluster production, was mutated in 4 unrelated patients with severe lactic acidosis and combined respiratory chain complex deficiencies (MIM 614299). Three of these patients (Pt045, Pt268, and Pt314) exhibited multiple organ failure; Pt268 and Pt314 had hypertrophic cardiomyopathy, and Pt045 developed seizures. All 4 patients exhibited decreased complex II activity and harbored the c.287A>G (p.H96R) mutation. Pt314 and Pt286 patients showed clear long contiguous stretches of homozygosity (LCSH) (2.8 Mb, 3.2 Mb respectively) around this p.H96R mutation. Pt268 also showed a short contiguous stretch of homozygosity (0.3 Mb). This homozygous region encompassing BOLA3 was shared between these unrelated individuals. Sanger sequencing identified the parents for these three patients as heterozygous carriers of this mutation. No p.H96R carriers were found in NHLBI GO Exome Sequencing Project (ESP6500), and 1 Japanese carrier in 1000 Genomes Project (1KG) was found. We screened for mutations that violated the Hardy–Weinberg principle and only identified the p.H96R mutation. These results suggest that p.H96R is common in the Japanese population and has originated from a single founder (S5 and S6 Figs).
NDUFAF6, which plays an important role in complex I assembly, was mutated in 4 unrelated patients: Pt101, Pt512, and Pt598 exhibited regression, whereas Pt330 exhibited muscle atrophy. All patients had complex I deficiency (MIM 256000). Pt101 shared 1 allele with Pt512 and another with Pt598. Pt330 harbored homozygous mutation c.820A>G (p.R274G) located in 1.3 Mb LCSH. Sanger sequencing identified the parents as heterozygous carriers of this mutation. Only 1 family was reported to harbor a mutation in this gene[18] (S7, S8 and S9 Figs).
NDUFB11, recently reported as causative gene for microphthalmia with linear skin defects syndrome (MIM 300952) and encoding a complex I component, was mutated in Pt067, a boy born to non-consanguineous parents under conditions of intrauterine growth restriction; this patient presented with heart failure, respiratory failure, complex I deficiency, and lethal infantile mitochondrial disorder (LIMD). He harbored a hemizygous de novo mutation, c.361G>A (p.E121K), and there was no NDUFB11 protein expression in his fibroblasts (S10 Fig). Because the p.E121 residue is highly conserved in this gene, we performed functional in vivo assays using a dndufb11-knockdown Drosophila model (S11 Fig); compared with controls, the mean lifespan was significantly reduced and the metabolic rate was lower in knockdown flies. Blue-native polyacrylamide gel electrophoresis (BN-PAGE) analysis showed a loss of complex I assembly, and lactate and pyruvate levels were increased in the knockdown flies. The in vivo dndufb11-knockdown Drosophila experiment further supported this conclusion. While preparing this manuscript, two girls harboring mutations in NDUFB11 with microphthalmia with linear skin defects were reported by van Rahden et al[19]. Our patient was a male and died 55 h after birth. He presented with redundant skin but had no linear skin defects.
Pt459, a boy with lactic acidosis, developmental delays, hypertrophic cardiomyopathy, seizure, and combined complex deficiencies (I and IV), harbored the compound heterozygous mutations c.1343T>A (p.V448D) and c.953T>C (p.I318T) in KARS. KARS is a lysyl-transfer RNA synthetase that produces 2 proteins that localize to the cytosol and mitochondria. A cDNA complementation assay revealed that mitochondrial KARS successfully rescued the enzyme defect, but not cytosolic form (S12 Fig). Detailed information and evidential support for other known genes are described in S1 Text.
Five (MRPS23, C1QBP1, ALAS2, SLC25A26, QRSL1) genes were identified as novel candidate genes (Tables 2 and S3). These genes were previously reported links to mitochondrial function but not mitochondrial respiratory chain complex deficiencies. Of these, we obtained pathogenic support for mutations in MRPS23 and QRSL1. In addition, candidate genes that have no evidence of functional involvement in current mitochondrial biology are good targets for underlying novel mitochondrial biological functions. In one such case, we identified PNPLA4 as a novel causative gene for mitochondrial respiratory chain complex deficiencies and proved its mitochondrial localization for the direct evidence of mitochondrial functions. The supportive evidence included (i) the identification of independent mutations in candidate genes in unrelated individuals with exquisitely similar phenotypes, (ii) rescue of patients’ cellular phenotypes in a cDNA complementation assay, and (iii) identification of a de novo mutation in the candidate gene. Other pVUS for candidate genes are shown in S3 Table.
A component of the highly conserved mitochondrial ribosome small subunit MRPS23[22] was mutated in Pt276, a boy with hepatic disease and combined respiratory chain complex deficiencies. In this patient, enzyme activities in complexes I and IV were decreased by 28% and 14% of the normal fibroblastic values, respectively. The patient was born to a non-consanguineous family. However, high-density oligonucleotide array analysis identified an approximately 500 kb contiguous stretch of homozygosity encompassing MRPS23. No other candidate gene was prioritized in our comprehensive genomic analysis. Pt276 harbored a homozygous c.119C>G (p.P40R) mutation in MRPS23 (NM_016070) (Figs 3A and S13). Sanger sequencing identified the parents as heterozygous carriers of this mutation. A complementation assay rescued the defect in complexes I and IV (Figs 3B and S13) and restored mitochondrial 12S rRNA/16S rRNA expression (Fig 3C).
Pt250, a girl with tachypnea, hypertrophic cardiomyopathy, adrenal insufficiency, hearing loss, and combined respiratory chain complex deficiencies (I, II, III, and IV), harbored a homozygous mutation c.398G>T (p.G133V) in QRSL1 (NM_018292) (Figs 3D and S14). Her older brother, also ill, harbored the same homozygous mutation. Sanger sequencing identified the parents as heterozygous carriers of this mutation. The high-density oligonucleotide array analysis identified a shorter 100 kb contiguous stretch of homozygosity encompassing QRSL1. QRSL1 (hGatA) is a glutaminase that produces ammonia, which is then transferred to misacylated Glu-charged tRNAGln to synthesize Gln-tRNAGln, which interacts with PET112L (hGatB) and GATC (hGatC) to form a trimeric enzyme hGatCAB[23]. Additional screening also identified an independent patient (Pt860) harboring the compound heterozygous mutations c.350G>A (p.G117E) and c. 398G>T (p.G133V) (Fig 3E). In vitro reconstitution of Gln-tRNAGln formation using recombinant hGatCAB revealed strongly decreased transamidation activity in both mutant (G117E or G133V) hGatA (Fig 3F).
PNPLA4 has both triacylglycerol lipase and transacylase activities. Pt712 is a boy who inherited a hemizygous nonsense variant c.559C>T (p.R187X) in PNPLA4 (NM_001142389) from his mother (Fig 4A and 4B). The colocalization of PNPLA4 and mitochondrial markers was identified by immunofluorescence microscopic observation (Fig 4C). We confirmed PNPLA4 protein loss in the fibroblasts of this patient by qRT-PCR (Fig 4D), sodium dodecyl sulfate (SDS)-PAGE/Western blotting (Fig 4E), and immunohistochemistry (Fig 4C). We found reduced complex I, III and IV assemblies of Pt712 fibroblasts under low glucose medium conditions (Fig 4G). The expression of PNPLA4-V5 cDNA in the fibroblasts of Pt712 recovered an amount of complex III and IV assemblies under low glucose medium conditions (Fig 4F and 4G).
In our cohort, all patients showed mitochondrial respiratory chain complex deficiencies. Intriguingly, we identified 3 cases having mutations in two genes that were previously reported to cause other monogenic diseases but not linked to canonical mitochondrial disease. These are MECP2 and TNNI3 (Table 2).
Pt053, a boy with complex I deficiency and seizures, diarrhea, arrhythmia, regression, respiratory failure, liver dysfunction, and hearing loss, harbored the hemizygous de novo mutation c.806delG (p.G269fs, rs61750241) in MECP2 (NM_004992) (S15 Fig), a gene reported to cause Rett syndrome (MIM 312750). We also identified another boy, Pt369, who harbored the hemizygous de novo mutation c.17_18insG (p.A6fs) in MECP2 (NM_001110792) (S15 Fig).
Pt827 was diagnosed with restrictive cardiomyopathy and complex I deficiency, and harbored the heterozygous de novo mutation c.575G>A (p.R192H, rs104894729) (S16 Fig) in TNNI3 (NM_000363); this exact mutation was reported to cause autosomal dominant familial restrictive cardiomyopathy (MIM 115210). Electron microscopic examination also revealed abnormally shaped mitochondria with concentric cristae (S16 Fig).
It has long been thought that patients with mitochondrial respiratory chain complex deficiencies rarely suffer chromosomal rearrangements but instead harbor mtDNA mutation, deletion, and depletion or nuclear DNA mutation. We subjected our entire cohort to a high-density oligonucleotide array to precisely evaluate the presence of any copy number variations (CNVs) of >100 kb. Intriguingly, we identified 13 patients (9.2%) harboring rare CNVs (Tables 3 and S5).
Pt369 and Pt657, 2 boys with complex IV deficiency from independent families, harbored similar deletions (1,429 and 1,387 kb) in chromosome 17p12. These 17p12 deletion disrupted the last 2 exons of COX10 in both patients (Fig 5A). This region causes hereditary neuropathy with liability to pressure palsies (HNPP) (MIM 162500). Whole exome analysis of Pt657 revealed an additional mutation c.683G>A (p.R228H) on the remaining allele of COX10 (Fig 5B and 5C). Notably, p.R228 is highly conserved among species. The PolyPhen2 and SIFT algorithms predicted this mutation as “probably damaging. In both patients, fibroblastic COX10 mRNA expression was reduced (Fig 5D). The complementation study using wild-type COX10 confirmed recovery of the complex IV deficiency in Pt657 (Fig 5E, 5F and 5G). Taken together, we concluded that Pt657 is a primary mitochondrial disorders. In Pt369, we concluded that de novo frameshift insertion mutation in MECP2 as a primary causative based on phenotype information, and classified 17p12 as pVUS.
We identified de novo 6q24.3-q25.1 deletions (S17 Fig) in Pt452 and Pt695, unrelated patients who harbored congenital heart defects. This region has been associated with the chromosome 6q24-q25 deletion syndrome (MIM 612863) and congenital heart defects[24].
Pt587, a boy diagnosed with LIMD and complex IV deficiency, harbored a deletion in 22q11.21. This deletion, which has been linked to DiGeorge syndrome (DGS, MIM 188400) and velo-cardio-facial syndrome (VCFS, MIM 192430), was confirmed as a de novo mutation in this patient (S17 Fig).
We performed comprehensive genomic analyses, including whole mtDNA and exome sequence analyses using high-throughput sequencing and CNV screening using high-density oligonucleotide arrays, for 142 patients with childhood-onset mitochondrial respiratory chain complex deficiencies. We ultimately identified 41 mutations, of which 37 were novel, in 20 genes that were previously reported to cause OXPHOS disease and 3 novel mitochondria-related genes (MRPS23, QRSL1, and PNPLA4) as causative genes of mitochondrial respiratory chain complex deficiencies. We also found 9 previously confirmed mtDNA mutations and 1 large mtDNA deletion. We further identified 2 genes known to cause monogenic diseases (MECP2, and TNNI3) and 3 chromosomal aberration regions (17p12, 6q24.3-q25.1, and 22q11.21) in our cohort. Collectively, this study defined firm genetic diagnoses in 49 of the 142 patients (34.5%). While the overall diagnostic rates for major and minor subgroups were similar (33.9% in the major subgroup and 36.4% in the minor subgroup), 35 out of 49 genetically diagnosed patients showed biochemical defects in their fibroblasts (71.4%), indicating a much higher genetic diagnostic yield in patients with such cellular defects. This is the first report to comprehensively assess patients diagnosed clinically and biochemically.
MRPS23 is a component of the small mitochondrial ribosome subunit (28S ribosome). Mutations in MRPS16[25] and MRPS22[26] cause mitochondrial respiratory chain complex deficiencies because of reductions of 12S rRNA, a 28S ribosome component. One patient exhibited a reduced 12S rRNA/16S rRNA ratio that was restored in a complementation study. This was the first case of MRPS23-induced mitochondrial respiratory chain complex deficiencies.
QRSL1 (GatA) is involved in Gln-tRNAGln formation. No mitochondrial glutaminyl-tRNA synthetase (GlnRS) has been identified in mammals; therefore, Gln-tRNAGln synthesis was proven to occur via an indirect pathway[23]. In particular, mt tRNAGln is first misaminoacylated by mt glutamyl-tRNA synthetase (GluRS) to form Glu-tRNAGln, followed by transamidation to form Gln-tRNAGln. This transamidation is processed by a human homolog of the Glu-tRNAGln amidotransferase hGatCAB heterotrimer. We clearly showed that mutations in QRSL1 (GatA), a component of hGatCAB, observed in our patients were associated with severe transamidation activity defects.
PNPLA4 encodes a calcium-independent phospholipase A2η (iPLA2η) that acts as an acylglycerol and retinol transacylase, triglyceride hydrolase. PNPLA4 has never been reported to associate with the mitochondria. Nine patatin-like phospholipase domain-containing proteins (PNPLA1–9) are encoded in the human genome. iPLA2γ (PNPLA8) is known to be involved in cardiolipin biosynthesis and mitochondrial respiration[27,28]. Recently, mutations in human PNPLA8 identified in a young girl with a suspected mitochondrial myopathy[29]. She presented with progressive muscle weakness, hypotonia, seizures, poor weight gain, and lactic acidosis. A deficiency in iPLA2β (PNPLA9) was previously shown to cause abnormal phospholipid metabolism and mitochondrial defects in mice[30]. Here we demonstrated the mitochondrial localization of iPLA2η using immunohistochemistry and restored of the amount of complex IV in Pt712 fibroblast cells via the exogenous expression of wild-type PNPLA4. We assume that PNPLA4 is also required for mitochondrial phospholipid metabolism and respiratory chain function.
Although all patients were diagnosed with mitochondrial respiratory chain complex deficiencies, we identified 2 disease-causing genes and 2 pathogenic CNVs known to cause other genetic disorders in our cohort. These included MECP2, TNNI3, 6q24.3-q25.1 deletion, and 22q11.21 deletion. Because all of these patients had complex II activities within the normal range (percentage of protein and citrate synthase ratio), we concluded that their defects were not artefactual[31,32]. Because these genes and loci are not directly linked to the respiratory chain complex, we consider the mitochondrial respiratory chain complex deficiencies are caused by secondary. Pt053, Pt369, and Pt827 were classified as having major ETC reductions in affected tissues, whereas Pt452, Pt695, and Pt587, harbored deletions are all classified as minor. The fact these heterozygous deletions are all classified as minor suggests that the mitochondrial defects in these patients might be caused indirectly through haploinsufficiency.
Because Pt053 and Pt369 harbored MECP2 mutations known to cause Rett syndrome, we re-evaluated the phenotypes of both patients and found phenotypes that overlapped with Rett syndrome characteristics (seizures, microcephaly, cerebral atrophy, and hearing loss). Previous studies also reported mitochondrial dysfunction in Rett syndrome[33,34]. Although Pt827 was enrolled with a diagnosis of mitochondrial disease, after comprehensive genomic analyses, the clinical diagnosis was changed to cardiomyopathy, familial restrictive (OMIM: 115210) caused by a mutation in TNNI3. Jia et al reported a link between Tnni3 and mitochondrial dysfunction using knockout mice [35].
Two independent patients from our cohort showed 6q24.3-q25.1 deletions. Pt452 exhibited a phenotype similar to that of cases reported cases in OMIM (MIM 612863). Pt695 presented with respiratory distress and a congenital heart defect. We classified these patients as having chromosome 6q24-q25 deletion syndrome. The enrichment of this deletion supports the suggested link with mitochondrial dysfunction. Pt587 was difficult to diagnose based on clinical information, because he did not have the facial anomalies and cleft palate characteristic of DGS/VCFS. The 22q11.21 deletion includes some mitochondria-related genes (PRODH, SLC25A1, MRPL40, TXNRD2, COMT, TANGO2, ZDHHC8, and AIFM3), suggesting a link between this deletion and mitochondrial dysfunction. The inclusion of a patient with features of DGS/VCFS and complex I deficiency in a study by Calvo et al[36] also indicates a link with mitochondrial dysfunction.
With these in mind, we should be mindful that some patients with mitochondrial respiratory chain complex defects will have mutations in genes apparently unrelated to mitochondrial functions.
Previous reports of mitochondrial disorders can be classified as either target resequence studies[37,38,7] or whole exome approaches[39,40,41]. When comparing target resequencing groups, our approaches are advantageous for the identification of mutations in other disease-causing genes, and the detection of chromosomal aberrations. WES groups[40,41] and our group detected pathogenic mutations in genes not linked to mitochondrial disorders. When comparing WES groups, our approaches are advantageous in terms of mtDNA sequencing and chromosomal aberration analysis. Our analysis could detect mtDNA heteroplasmy using long-range PCR-based sequencing and also revealed established pathogenic chromosomal deletions. Accordingly, we identified a composite combination of COX10 SNV and 17p12 deletion (Pt657). Previous WES and target exome reports achieved molecular diagnoses in 20%–60% of their cohorts. A precise comparison of overall diagnostic rates with previous studies is difficult, given the existence of several biases that affect the diagnostic rate, including prior mtDNA/nDNA genetic screening, population characteristics, phenotyping accuracy, and study design. In particular, the reports by Taylor et al[40] described a high rate of diagnosis (approximately 60%) in their cohort, although their patient group appeared to be enriched by consanguineous families (12 of 28 diagnosed cases). The report by Wortmann et al[41] described a rate of diagnosis (38%) similar to ours. In our study, we emphasized functional analyses to conclude disease causality against pVUS and attempted to present molecular evidence of pathogenicity;in contrast, some previous studies lacked sufficient molecular evidence of pathogenicity. We designated the variants without any molecular evidence of pathogenicity as pVUS, even when the gene had been reported as a causal gene for mitochondrial disorders. We consider molecular evidence to be indispensable for a conclusive firm genetic diagnosis.
We found that approximately 28.2% of patients lacked any prioritized variants. We likely missed pathogenic mutations in these unresolved cases for the following reasons, as discussed in a report by Calvo et al[6]: first, we may have missed pathogenic mutations because of a lack of sensitivity from low sequence coverage. Second, pathogenic mutations may be located in uncovered genomic regions (e.g., uncovered exons, introns, or regulatory regions not targeted by whole exome platforms). Third, our filtering strategy may have filtered true pathogenic mutations, although some were recovered by manual curation. Fourth, the hereditary assumption may be wrong. More dominant-acting cases may exist. Digenic/polygenic inheritance may also exist beyond our expectation.
In conclusion, for suspected mitochondrial disorders, comprehensive analyses such as those in this study are worthwhile. We expanded the clinical disease spectrum and revealed the genetic landscape of this disorder.
In total, 142 patients with childhood-onset and enzymatically diagnosed mitochondrial respiratory chain complex deficiencies were enrolled in this study. Informed consent was obtained from the patients and their families before participation in the study. Patients with suspected mitochondrial respiratory chain complex deficiency were referred to the Saitama Medical University Hospital and Chiba Children’s Hospital in Japan from 2007 to 2013. The inclusion criterion was a biochemical diagnosis of mitochondrial respiratory chain complex activity in a clinically affected tissue (skeletal muscle, liver, or heart) or fibroblasts in patients younger than at the age of 15 years. Patients with known nuclear or mtDNA mutations at the time of recruitment were excluded. The 142 included patients had not received a prior molecular diagnosis, despite varying degrees of exposure to genetic testing. This cohort included 3 non-Japanese cases: Pt346 (father, American; mother, Japanese), Pt298 (Brazilian), and Pt223 (Vietnamese). Enzyme activity[42] was measured on the basis of spectrophotometric enzyme assays using fibroblasts from patient’s skin or biopsy specimens from diseased organs of patients with clinically suspected mitochondrial respiratory chain disorders[43]. All enrolled patients in this study had biochemical mitochondrial respiratory chain complex deficiencies; the enzymatic diagnoses are shown in S1A Fig. In brief, complex I deficiency was most common (61 patients, 43.0%), followed by (in decreasing order of prevalence) combined respiratory chain complex deficiencies (46 patients, 32.4%), complex IV deficiency (27 patients, 19.0%), MTDPS (5 patients, 3.5%), and complex III deficiency (3 patients, 2.1%); no patients exhibited complex II deficiency. Diagnoses of mitochondrial respiratory chain complex deficiencies were assessed as “major” or “minor” on the basis of biochemical complex activity. Based on the Bernier criteria, severity was defined as major (<20% in a tissue, <30% in a fibroblast cell line, or <30% in ≥2 tissues) or minor (<30% in a tissue, <40% in a fibroblast cell line, or <40% in ≥2 tissues) in accordance with the residual mean citrate synthase or complex II activities relative to those of normal controls (S1B Fig). The distribution of age of onset of these patients was as follows: 45.7% (65 patients) before 1 month, 19.7%(28 patients) within 1–6 months, 19.7%(28 patients) within 6–24 months, 12,7% (18 patients) within 2–10 years, and 2.1% (3 patients) within 10–15 years (S1C Fig). The clinical diagnoses of 142 patients are also shown in S1D Fig. The most common diagnosis was mitochondrial cytopathy (27 patients, 19.0%), followed by Leigh’s disease (25 patients, 17.6%), LIMD (23 patients, 16.2%), sudden unexpected death (17 patients, 12.0%), non-lethal infantile mitochondrial disorder (NLIMD) (16 patients, 11.3%), cardiomyopathy (11 patients, 7.7%), hepatic disease (11 patients, 7.7%), enteropathy (6 patients, 4.2%), neurodegenerative disorder (4 patients, 2.8%), and short stature (2 patients, 1.4%). The male:female ratio was 76:66. There were no consanguineous relationships among our cohort. Detailed clinical characteristics are described in S1 Table.
DNA was isolated from cultured fibroblast cells using the QIAamp DNA Blood mini Kit (QIAGEN). Blood genomic DNA was isolated by phenol–chloroform extraction according to the standard protocol. Total RNAs were purified from HEK293FT cells, fibroblast cells using the SV Total RNA Isolation System (Promega). cDNAs were synthesized from total RNAs using ReverTra Ace (Toyobo). Total RNA was extracted from flies using TRIzol reagent (Invitrogen), and RNA was reverse transcribed by SuperScript VILO transcriptase (Invitrogen).
To avoid the contamination of mitochondrial-origin nuclear genome sequences [44] (NUMTs), a long-range mtDNA polymerase chain reaction (PCR) method was used in this study. DNA were extracted from patients skin fibroblast cells. These DNAs were checked for large-scale mtDNA rearrangements and subjected to large mtDNA deletion mapping using long-range PCR with amplicon 1 (rCRS 619–8988) and amplicon 2 (rCRS 8749–895) primers; 5′-GACGGGCTCACATCACCCCATAA-3′ and 5′-GCGTACGGCCAGGGCTATTGGT-3′ for amplicon 1, and 5′-GCCACAACTAACCTCCTCGGGCTCCT-3′ and 5′-GGTGGCTGGCACGAAATTGACC-3′ for amplicon 2.
Indexed PCR fragment libraries were prepared from patient mtDNA using the Nextera XT DNA Sample Prep Kit (Illumina) according to the manufacturer’s protocol. Sequencing library concentrations were estimated using a library quantification kit (Kapa Biosystems). Sequencing was performed with 150-bp paired-end reads on MiSeq (Illumina). Read alignments to the 1000 Genomes Project phase II reference genome (hs37d5.fa), which contains rCRS sequences, were performed with the Burrows–Wheeler Aligner[45] (BWA, version 0.7.0). PCR duplicate reads were removed using Picard (version 1.89); non-mappable reads were removed using SAMtools[46] (version 0.1.19). After filtering out those reads, we applied the Genome Analysis Toolkit[47] (GATK version 2.4-9-nightly-2013-04-12-g3fc5478) base quality score recalibration and performed SNP and INDEL discovery (UnifiedGenotyper). Confirmed pathogenic mutations and reported variants in MITOMAP and mtDNA deletions detected through reference-based alignment with BWA mapping were prioritized (S4 Table).
De novo mtDNA sequence assembly was performed using SPAdes (version 3.0.0)[48] with iterations over values of 3 kmer sizes (k = 75, 95, and 115). Each assembly was aligned to the mitochondrial genome sequence of hs37d5.fa using BLASTN (version 2.2.29+) with default settings and was manually inspected to identify aberrations (deletions, duplications, and rearrangements).
A large mtDNA deletion in Pt334 was validated using long-range PCR with primers 5′-GCCACAACTAACCTCCTCGGGCTCCT-3′ and 5′-GGTGGCTGGCACGAAATTGACC-3′. The mtDNA was also sequenced (using primers 5′-ACTACCACTGACATGACTTTCCAA-3′ and 5′-TGTTGTTTGGATATATGGAGGATG-3′ for amplification and 5′-CTTATCCAGTGAACCACTATCACG-3′ for sequencing) closer to the breakpoint as described above.
Quantitative PCR[49] was used to determine whether mtDNA depletion was present in patients with decreased activity levels for multiple respiratory chain enzymes (mtDNA gene MT-ND1 was compared against a nuclear gene [CFTR exon 24]). A diagnosis of MTDPS was made when the relative copy number ratio of mtDNA to nuclear DNA was less than 35% of that in healthy control tissues in 4 independent experiments.
Quantitative reverse transcription PCR (qRT-PCR) was performed for the analysis of mRNA (NDUFB11, TTC37, and PNPLA4) and mitochondrial rRNA expression[50]. Primers were designed with the Primer3 software[51]. Primer sequences used in the qRT-PCR analysis are listed in S6 Table. qRT-PCR of cDNA extracted from human cells was performed using SYBR Premix Ex Taq (Takara), Power SYBR Green PCR Master Mix (Life technologies), and Mx3000P (Agilent Technologies). The relative mRNA concentration was normalized to the average of two housekeeping genes (β-actin and GAPDH). qRT-PCR of cDNA extracted from flies was performed using SYBR Premix Ex Taq II (Takara) and Chromo 4 Four-Color Real-Time System (Bio-Rad). Results were normalized to the rp49 mRNA level.
Indexed genomic DNA (gDNA) libraries were prepared from patient gDNA, and exomes were captured using TruSeq (Illumina), SeqCap EZ (Roche AG, Basel, Switzerland), and SureSelect (Agilent Technologies) exome enrichment kits according to the manufacturers’ protocols. Sequencing library concentrations were estimated using a library quantification kit (Kapa Biosystems). Sequencing was performed using 100-bp paired-end reads on a HiSeq2000 or GAIIx (Illumina). The precise exome platforms used in this study are listed in S2 Table. The raw sequence read data passed the quality checks in FASTQC (see URLs). Read trimming via base quality was performed using Trimmomatic[52]. Read alignment was performed with BWA, the hs37d5 reference genome, Picard, and SAMtools as described above. GATK was also used for insertion and deletion realignment, quality recalibration, and variant calling. Detected variants were annotated using both ANNOVAR (version 2013Feb21)[53] and custom ruby scripts. Prediction scores were obtained from dbNSFP[54].
Variants that passed quality control were prioritized according to the following strategies (S2 Fig). We only retained variants predicted to modify protein function; these included nonsense, splice site, coding indel, or missense variants. We removed variants with minor allele frequencies (MAFs) of >1.0% for dbSNP 137 without known medical impact (allele frequencies were extracted from common_no_known_medical_impact_20130808.vcf.gz), >0.1% for ESP6500 (provided by ANNOVAR program) database, >1.0% for 1KG (these data are based on a phase 1 release v3 called from the 20101123 alignment and provided by ANNOVAR), >0.1% for the Exome Aggregation Consortium (ExAC, accessed on December 2014), and >0.4% in HGVD (contains genetic variations determined by exome sequencing of 1,208 individuals in Japan; see URLs). Variants that were too common among cases (≥10 alleles) were also excluded. Careful inspection of the reads using the Integrative Genomics Viewer[55,56] and NextCODE clinical sequence analyzer (see URLs) excluded doubtful genes from prioritized candidate genes when 2 sequence variants were present in the same read (or read-pair). Variants that appeared to be mapping artifacts (called by suspicious reads or end positions of NGS reads) were also excluded through a manual inspection of NGS reads. Variants located within segmental duplication regions were excluded. In addition to these filters, Sorting Intolerant From Tolerant (SIFT) scores > 0.15 and Genomic Evolutionary Rate Profiling (GERP) scores < 2.5 were used for further prioritization. We also excluded variants inconsistent with a recessive mode of inheritance. Two (or more) variants on a single haplotype as identified by Sanger sequencing were also excluded. Finally, we filtered remaining genes based on MAF and genotype information in the 1,070 whole-genome reference panel database (1KJPN) constructed in the Tohoku Medical Megabank Project in Japan (http://ijgvd.megabank.tohoku.ac.jp/). The details of the project and analysis are described in the 1KJPN literature[57]. To recover true mutations that were filtered out using current pipeline applying stringent conditions, we also applied this pipeline without a segmental duplication filter, SIFT filter, or GERP filter, followed by focusing on mitochondria-related genes.
Enrichment analysis was conducted to evaluate our exome pipeline and included 128 cases and 175 ethnically matched healthy controls whose sequence reads exceeded 50 million; these were adjusted to 50 million reads per individual. We used a simplified exome analysis pipeline that did not include a manual inspection step, Sanger sequencing validation step, and 1KJPN filtering step. We also omitted the HGVD filtering step because these control data were included in the HGVD samples. The other steps were the same as those described above for the exome analysis pipeline. After processing the controls and cases, we calculated the percentage of individuals harboring prioritized genes. Comparisons of the percentages of controls and cases on the basis of known disease genes and mitochondria-related genes are shown in S3A Fig. To consider the background rate of this simplified pipeline, we also evaluated enrichment using randomly selected 908 genes with no strong mitochondrial relationships in their annotations. The gene set was generated 1000 times via random selection from all genes after excluding those known to cause OXPHOS disease and those listed in MitoCarta[18]. The results plus standard deviations are shown in S3B Fig.
Prioritized variants were independently validated by Sanger sequencing. PCR products were either directly sequenced using GENEWIZ, ABI 3130XL, and BigDye v3.1 Terminators (Applied Biosystems) per the manufacturer’s protocols or sequenced after gel purification using the MinElute Gel Extraction Kit (QIAGEN). Sequencing primers are listed in S7 Table. All compound heterozygous variants described in the main text were confirmed on different alleles (phased) using sequenced, cloned gDNA or cDNA derived from the patients’ fibroblasts. Patients’ familial DNA was also sequenced for haplotype phasing when available. All information about the experimentally confirmed localization of compound variants within a separate allele is presented in S3 Table.
Amino acid sequences of orthologous genes were downloaded from the HomoloGene database (see URLs). Amino acid sequence alignments were constructed with the ClustalW2 program[58].
Cells were cultured at 37°C and 5% CO2 in Dulbecco's modified Eagle’s medium (DMEM 4.5 g/l glucose or 1.0 g/l glucose; Nacalai tasque) supplemented with 10%–20% fetal bovine serum. Normal neonatal human dermal fibroblasts (NHDFs; Takara) and normal fetal human dermal fibroblasts (fHDFs; Toyobo) were used as control fibroblast cells. Open reading frames (ORFs) of candidate genes (ACAD9, BOLA3, COX10, KARS, MRPS23, NDUFA10, NDUFAF6, PNPLA4, and TUFM) were PCR amplified from cDNA. Primer sequences used for cDNA cloning are listed in S8 Table. ORFs and pTurboRFP-mito (TurboRFP fused to a mitochondrial targeting sequence derived from the subunit VIII of human cytochrome C oxidase; Evrogen) were cloned into the CS-CA-MCS lentiviral vector with a C-terminal V5 tag, CAG promoter for mammalian cell expression, and blasticidin resistance using the In-Fusion HD Cloning Kit (Clontech Laboratories, Inc.). Following this, 2 × 106 HEK293FT cells were seeded in 6-cm plates and co-transfected with ViraPower Packaging vectors (pLP1, pLP2, pLP/VSVG; Invitrogen) and a pCA-CS-ORF(candidate gene)-blast vector. Transfection was performed using Lipofectamine 2000 (Invitrogen). Transfection medium was replaced with fresh medium 24 h after transfection. Supernatant containing the viral particles was collected 48 h after transfection and filtered through a 0.45 μm filter. Patients’ skin fibroblasts were infected with the viral supernatant and 5 μg/ml polybrene (Sigma) for 24–48 h. After 5–7 days, selection was initiated with 1–2 μg/ml blasticidin. After 1–3 months of selection, mitochondria were harvested from the cells for enzyme assays or BN-PAGE.
To prepare enriched mitochondria, cell pellets were resuspended in ice-cold MegaFb Buffer (250 mM sucrose, 2 mM HEPES, 0.1 mM EGTA, pH 7.4) and homogenized with 20 strokes. The homogenates were centrifuged for 10 min at 600 g. Supernatants were centrifuged for an additional 10 min at 14,400 g. Pellets were resuspended in 400 μl MegaFb buffer, and 200 μl aliquots were frozen and thawed 3 times for complex II + III and complex III assays and protein estimation. The remaining samples were resuspended in hypotonic buffer (25 mM potassium phosphate, pH 7.2, 5 mM MgCl2) for complex I, II, and IV, citrate synthase, and protein concentration assays and centrifuged for 10 min at 14,400 g. Pellets were resuspended in Hypotonic Buffer and subjected to 3 freeze–thaw cycles. These samples were stored at −80°C prior to enzyme assays. Respiratory chain enzyme activities were measured using cary300 (Agilent Technologies) as described previously[42]. Complex I, II, II + III, III, and IV activities were expressed as percentages of citrate synthase activity.
To isolate mitochondria, cell pellets were suspended in mitochondria isolation buffer A (220 mM mannitol, 20 mM HEPES, 70 mM sucrose, 1 mM EDTA, pH 7.4, 2 mg/ml bovine serum albumin, 1× protease inhibitor cocktail) and homogenized with 20 strokes on ice. Homogenates were separated into cytosolic and nuclear fractions after centrifugation at 700 g for 5 min at 4°C. The supernatants were centrifuged at 10,000 g for 10 min at 4°C. Mitochondrial pellets were rinsed twice with mitochondria isolation buffer B (220 mM mannitol, 20 mM HEPES, 70 mM sucrose, 1 mM EDTA, pH 7.4, 1× protease inhibitor cocktail). Mitochondria were isolated from adult flies as described previously[59]. Fifty flies were homogenized in 1 ml of chilled mitochondrial isolation medium (MIM; 250 mM sucrose, 10 mM Tris pH 7.4, 0.15 mM MgCl2). The samples were centrifuged twice for 5 min at 1,000 g at 4°C to remove debris. The supernatant was centrifuged again for 5 min at 13,000 g at 4°C.Mitochondrial protein levels were determined using a bicinchoninic acid (BCA) assay. For SDS-PAGE analyses, enriched mitochondria and cell pellets were solubilized in M-PER Mammalian Protein Extraction Reagent (Thermo Fisher Scientific) and denatured for 30 min at 37°C. Prepared samples were separated by electrophoresis on 8%, 10%, and 15% SDS-PAGE gels, depending on the size of the detected protein. For BN-PAGE analyses, The NativePAGE Novex Bis-Tris Gel System (Life Technologies) was used according to the manufacturer’s protocol. Mitochondrial fractions were solubilized in NativePAGE sample buffer containing 0.5% Triton-X100 and separated on 4%–16% NativePAGE gels. The BN-PAGE analyses of Drosophila were performed as previously described[60]. Immunoblot analysis was performed as described previously[61]. Anti-NDUFA9 (Complex I), anti-70 kDa Fp Subunit (Complex II), anti-core 1 (Complex III), anti-subunit 1 (Complex IV), and anti-V5 antibodies were purchased from Life Technologies. Anti-Lamin A/C antibody was purchased from BD biosciences. Anti-HSP60, anti-NDUFA10, anti-ACAD9, and anti-COX10 antibodies were purchased from Abcam, and anti-NDUFB11 antibody was purchased from Santa Cruz Biotechnology. Anti-TTC37 antibody was purchased from ProteinTech. Anti-PNPLA4 (GS2) was purchased from GeneTex. Anti-tafazzin and anti-α/β-tubulin antibodies were purchased from Cell Signaling Technology. Anti-β-actin antibody was purchased from Sigma. Anti-MECP2 antibodies were purchased from Acris Antibodies and Merck Millipore.
Normal human dermal fibroblast cells and patient cells were seeded in a 35-mm glass-bottom dish. Mitochondria were stained with 500 nM MitoTracker Orange CMXRos (Molecular Probes) for 30 min in DMEM containing 10% fetal bovine serum. Cells were fixed with 4% paraformaldehyde for 20 min and permeabilized by incubation in 0.2% Triton X-100. After blocking with 3% bovine serum albumin, fluorescent staining was performed with rabbit anti-PNPLA4 antibody (GeneTex) or mouse anti-V5 antibody (Life Technologies) and secondary Alexa Fluor 488 antibody (Molecular Probes) or secondary FITC antibody (Sigma). Cells were visualized with a Leica TCS SP8 confocal microscope.
For siRNA transfection, Lipofectamine RNAiMAX (Invitrogen) and 120 pmol of siRNA were prepared according to the manufacturer’s instructions and directly added to a 10 cm culture dish of NHDF fibroblasts. Mitochondria were isolated after 6 days, and the assembly levels of respiratory chain complexes were analyzed using BN-PAGE and Western blotting. The Stealth RNAi siRNA (Life Technologies) sequences used for NDUFB11 knockdown are as follows: HSS147694 (#94), ACC CAG ACU CCC AUG GUU AUG ACA A; HSS147695 (#95), UCC AAG AGC GUG GGA UGG GAU GAA A; HSS147696 (#96), CCU CUU CUC AGA GCA CCU AAU UAA A. Stealth RNAi siRNA Negative Control, Med GC (cat no. 12935–300) was used as the negative control. The Silencer Select RNAi siRNA (Life Technologies) used for MECP2 knockdown are as follows: s8644 (#44), s8645 (#45), s8646 (#46). Silencer Select RNAi siRNA Negative Control, No.2 (#2) (cat no. 4390847) was used as the negative control.
Flies were reared at 25°C in a standard glucose yeast agar medium containing propionic acid and n-butyl p-hydroxybenzoate as mold inhibitors. arm-Gal4 was obtained from the Bloomington Drosophila Stock Center. UAS-dndufb11 (NP15.6)-IR (5717) was obtained from the Vienna Drosophila RNAi Center. UAS-GFP-IR (GFP-IR-2) was obtained from the National Institute of Genetics Fly Stock Center.
Newly eclosed flies were housed in a glass vial containing the standard glucose yeast medium and were transferred to fresh media every 2 days; the numbers of dead flies were counted at the time of transfer. At least 100 flies per genotype were used for experiments.
Eight flies were placed in an 8-lane cell vial (1 cell; H 7 mm × W 8 mm × D 70 mm) and bumped to the bottom. Pictures were taken at 5 s after bumping and used to measure the distance climbed by each individual. For each sample, the average climbing activity of 10 trials was determined.
CO2 production was measured as described previously[62]. In brief, 10 adult flies were placed in a 1-ml plastic syringe that contained a small amount of CO2-absorbent material (Soda lime), which was connected to a 200-μl glass disposable micropipette. A small amount of black ink was placed at the end of the micropipette as an indicator of CO2 production. The apparatus was kept on a flat surface at 25°C, and measurement was initiated after 10 min. The amount of CO2 produced by the flies was calculated according to changes in the air volume during 1 h of measurement. Assays were performed at least 3 times per genotype.
Lactate and pyruvate measurements were performed as described previously[63].
The cDNAs for hGatA with C-terminal SBP-HA-tag and hGatC with C-terminal HA-tag were cloned into pENTR/D-TOPO (Invitrogen). Each of the pathogenic point mutations (G117E and G133V) was introduced into the hGatA gene in the entry clone by site-directed mutagenesis using PrimeSTAR HS DNA polymerase (Takara) with primers 5′-GATCAGGGAGCTCTACTAATGGAAAAAACAAATTTAGA-3′ and 5′-TCATCTAAATTTGTTTTTTCCATTAGTAGAGCTCCCTGATC-3′ for G117E, and 5′-GATCTGGGAGCACAGATGTTGTATTTGGACCAGTTAAAAAC-3′ and 5′-GTTTTTAACTGGTCCAAATACAACATCTGTGCTCCCAGATC-3′ for G133V. The cDNAs for hGatA (WT, G117E or G133V) and hGatC were transferred from each entry clone to pHAGE to generate the expression vector[64] by LR reaction (Invitrogen).
HEK293T cells were co-transfected with lentiviral vectors (TAT, VSVG, RRE, or REV), pHAGE-hGatA (WT, G117E or G133V) and pHAGE-hGatC. The transformants were cultured at 37°C for 3 days. Cells were harvested and suspended in lysis buffer [50 mM HEPES-KOH (pH 7.5), 200 mM KCl, 1 mM PMSF, 0.1% TritonX-100, 1 mM DTT, 2.5 mM MgCl2] containing complete protease inhibitor cocktail (Roche) and were disrupted by sonication at 0°C. The hGatCA complex in the cell lysate was captured with streptavidin-Sepharose beads (GE Healthcare) and was eluted from the beads with 4 mM of biotin according to the manufacturer’s instructions. The eluted hGatCA complex was subjected to SDS-PAGE, stained by SyproRuby, and quantified with a FLA-7000 imaging analyzer (Fujifilm) with BSA as a standard. Recombinant hGatB was expressed in Escherichia coli and was purified as described previously[23].
As human mt GluRS strictly recognizes the post-transcriptional modification at the anticodon first position (position 34) of human mt tRNAGln for glutamylation[23], in vitro-transcribed human mt tRNAGln cannot be aminoacylated by human mt GluRS. However, Thermotoga matritima nondiscriminating GluRS can efficiently glutamylate tRNAGln bearing unmodified C at position 34[65]. We accordingly prepared in vitro-transcribed human mt tRNAGln with C34 for glutamylation by T. maritima GluRS. Human mt tRNAGln with C34 was transcribed in vitro by T7 RNA polymerase from the template DNA PCR-amplified with the synthetic DNAs 5′-GCTAATACGACTCACTATATAGGATGGGGTGTGATAGGTGGCACGGAG-3′, 5′-ATAGGTGGCACGGAGAATTCTGGATTCTCAGGGATGGGTTCGAT-3′, and 5′-TGGCTAGGACTATGAGAATCGAACCCATCCCTGA-3′, as described previously[66,67]. The aminoacylation reaction was performed at 37°C for 30 min in a mixture containing 50 mM HEPES-KOH(pH 7.5), 20 mM KCl, 10mM MgCl2, 2 mM ATP, 1 mM DTT, 1 mM spermidine, 20 μM [14C]L-glutamine(9.36 GBq/mmol), 0.02 A260 unit of tRNA transcript, and 1.88 μM T. maritima GluRS. The [14C]Glu-tRNAGln was extracted by phenol–chloroform treatment under acidic conditions followed by ethanol precipitation. Residual ATP in the reaction was removed using a Nap5 gel filtration column (GE Healthcare). A small part of the mixture was spotted onto Whatman 3MM filter discs, followed by washing with 5% trichloroacetic acid, and the radioactivity was measured by liquid scintillation counting.
In vitro reconstitution of Gln-tRNAGln formation by hGatCAB was performed as described previously[23]. The reaction was performed at 37°C in a mixture of 100 mM HEPES-KOH (pH 7.5), 30 mM KCl, 12 mM MgCl2, and 2.5 mM DTT, 5 mM ATP, 6.3 nM recombinant hGatCA (WT, G117E, or G133V), 1.03 μM recombinant hGatB, 65 nM [14C]Glu-tRNAGln and 2 mM glutamine. Over time, aliquots of the reaction mixture were taken at 0, 1, 5, 10, and 15 min, and were mixed with phenol–chloroform to extract aminoacyl-tRNAs under acidic conditions, followed by ethanol precipitation and removal of ATP using a Nap5 column. The amino acids attached to the tRNA were deacylated at 37°C for 30 min in 0.3% aqueous ammonia. The [14C] labeled amino acids were analyzed by thin-layer chromatography (TLC) on a cellulose plate (Melck) using a basic solvent system (28% ammonia solution:chloroform:methanol, 1:3:4). The TLC plate was exposed to an imaging plate, and the radioactivity was visualized using FLA-7000 image analyzer (Fujifilm).
Samples were processed in accordance with the manufacturer’s instructions. In brief, two aliquots of 250 ng genomic DNA were digested with Nsp1 and Sty1, and ligated to adaptors. Generic primers recognizing the enzyme-specific adaptor sequences were used to amplify adaptor-ligated DNA. After purification, 270 μg of the PCR product was fragmented and labeled with biotin. Hybridization was performed in an Affymetrix GeneChip Hybridization Oven 640, and the arrays were washed and stained in an Affymetrix GeneChip Fluidics Station 450. Arrays were scanned with an Affymetrix GeneChip Scanner 3000 7G. Hardware scripts were enabled and image processing performed using the Affymetrix GeneChip Command Console software (AGCC). Genotypes were called using the Affymetrix Genotyping Console software v4.1.1 GTC with the Birdseed algorithm and a default-calling threshold of 0.1. Samples were required to have an average minimum Quality Control SNP call rate of 99.7%.
All samples were analyzed with GTC v4.1.1. The predicted copy numbers as well as the start and end of each CNV segment were determined using the Hidden Markov Model. In all datasets, hg19 was used. All large CNVs were manually curated. The CNV calls were also generated using the PennCNV software[68].
Results are presented as mean ± SEM or SD for the number of experiments indicated in the figure legends. Statistical analysis of continuous data was performed with 2-tailed Student’s t test, as appropriate. p < 0.05 was considered statistically significant.
The study was approved by the ethics committee of the Saitama Medical University. Written informed consent was obtained from all subjects prior to inclusion in this study.
1000 Genomes Project, http://www.1000genomes.org/; hs37d5.fa ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/technical/reference/phase2_reference_assembly_sequence/; DECIPHER, https://decipher.sanger.ac.uk; DGV, http://dgv.tcag.ca/dgv/app/home; ExAC [Dec., 2014 accessed], Cambridge, MA, http://exac.broadinstitute.org; ESP6500 [accessed via ANNOVAR 2013Feb21 version], http://evs.gs.washington.edu/EVS/; FASTQC, http://www.bioinformatics.babraham.ac.uk/projects/fastqc/; GERP, http://mendel.stanford.edu/SidowLab/downloads/gerp/; The Human Genetic Variation Database (HGVD), http://www.genome.med.kyoto-u.ac.jp/SnpDB/index.html; Hgvd2annovar, https://github.com/misshie/hgvd2annovar; Homologene, http://www.ncbi.nlm.nih.gov/homologene; MitoCarta http://www.broadinstitute.org/pubs/MitoCarta/index.html; mtDB, http://www.mtdb.igp.uu.se/; NextCODE, http://www.nextcode.com/; OMIM, http://www.omim.org; Picard, http://broadinstitute.github.io/picard/; PolyPhen-2, http://genetics.bwh.harvard.edu/pph2/; R for statistical analysis, http://www.R-project.org/; Ruby, https://www.ruby-lang.org/en/; SIFT, http://sift.jcvi.org/.
|
10.1371/journal.pgen.1007072 | A mutation in the viral sensor 2’-5’-oligoadenylate synthetase 2 causes failure of lactation | We identified a non-synonymous mutation in Oas2 (I405N), a sensor of viral double-stranded RNA, from an ENU-mutagenesis screen designed to discover new genes involved in mammary development. The mutation caused post-partum failure of lactation in healthy mice with otherwise normally developed mammary glands, characterized by greatly reduced milk protein synthesis coupled with epithelial cell death, inhibition of proliferation and a robust interferon response. Expression of mutant but not wild type Oas2 in cultured HC-11 or T47D mammary cells recapitulated the phenotypic and transcriptional effects observed in the mouse. The mutation activates the OAS2 pathway, demonstrated by a 34-fold increase in RNase L activity, and its effects were dependent on expression of RNase L and IRF7, proximal and distal pathway members. This is the first report of a viral recognition pathway regulating lactation.
| Using ENU-mutagenesis in mice we discovered a pedigree with lactation failure. Mammary development through puberty and pregnancy appeared normal in mutant animals, but the activation of lactation failed in the immediate post partum period and no milk reached the pups. Failure of lactation was accompanied by greatly diminished milk protein synthesis, decreased epithelial cell proliferation, increased epithelial cell death and a robust interferon response. A non-synonymous mutation in Oas2 (I405N) in the viral sensor Oas2 was found and expression of mutant Oas2 in mammary cells recapitulated these phenotypes. RNase L, the most proximal effector of OAS2 action, was activated in the mammary glands of mutant mice and in mammary cells expressing mutant Oas2. Knockdown of RNase L, or the distal pathway member IRF7, prevented these effects, indicating that the mutation in OAS2 caused activation of the viral signaling pathway. These results show that viral detection in the mammary gland can prevent lactation.
| The oligoadenylate synthetase (OAS) enzymes are activated by detection of double stranded RNA produced during the viral life cycle, and in response polymerize ATP into 2´-5´ linked oligoadenylates (2-5A) of various lengths. The 2-5As then bind and activate latent RNase L, which degrades viral and host single stranded RNAs, so disrupting the viral life cycle [1]. It has been reported that OAS1 has antiviral activity independently of RNAse L [2], that OAS2 binds to NOD2 [3], and that OASL binds RIG-I [4], pointing to additional mechanisms of action. Although mechanistic detail is lacking, it is proposed that OAS enzymes can activate anti viral responses via mechanisms independently of 2-5A production, by direct interactions within the viral signaling complex. For example, this complex is tethered to the mitochondrial outer membrane by the scaffold protein MAVS, and contains RIG-I, related helicase MDA5, and possibly OAS family members [5]. OAS family members may also mediate apoptosis outside the context of viral infection [6,7]. Here we report a mutation of OAS2 that produces lactation failure in an otherwise normal mouse. This is the first demonstration that a viral recognition pathway can regulate lactation.
Using N-ethyl-N-nitrosourea (ENU) mutagenesis and a screen for failed lactation we established a mouse line in which heterozygous (wt/mt) dams showed partial penetrance of poor lactation, producing litters that failed to thrive, while homozygous (mt/mt) dams experienced complete failure of lactation (Fig 1A), providing a dominant pattern of inheritance. Development of the mammary ductal network during puberty, and of the lobulo-alveolar units during pregnancy, was normal in mt/mt dams (S1A and S1B Fig). The onset of milk protein synthesis also showed no defects during pregnancy by immunohistochemistry or western blot (S1C Fig and S1D Fig). Lactation failure in mt/mt mice at 2 days post-partum (2dpp) was seen as failure of alveolar expansion and retention of lipid droplets and colostrum (Fig 1B–1G). Western blotting for milk (Fig 1H and S1C and S1D Fig) showed greatly reduced expression of all the major milk components at 2dpp relative to the level of the epithelial cell marker cytokeratin 18. Quantitative PCR for the mRNAs for the milk proteins whey acidic protein (WAP) and β-casein (β-Cas) showed reduced levels in mt/mt dams (mt) compared to wt/wt dams (wt) at 18 days post-coitus (dpc) and especially at 2dpp (Fig 1I and 1J). The number of cleaved Caspase-3 positive epithelial cells increased (Fig 1K) and BrdU incorporation by the epithelium was reduced, indicating increased cell death rate and decreased cell proliferation respectively (Fig 1L).
We used immunohistochemistry to examine STAT1 activation as it is an interferon regulated gene involved in mammary gland involution [8]. In wt/wt dams at d18.5 of pregnancy and 2 days post partum, we observed scattered regions of phosphorylated Stat1 staining in tightly packed areas of small and unexpanded alveoli (Fig 1M). These regions were very rare at the other stages of development examined. In mt/mt animals Stat1 phosphorylation was again seen within regions of small unexpanded and tightly packed alveoli (Fig 1N), but at day 18.5 of pregnancy, these regions of STAT1 phosphorylation occurred at a far greater frequency than in wt/wt glands, and instead of receding in the post partum period like wt/wt glands, the frequency of this pattern of staining increased further (Fig 1O). We examined Stat1 phosphorylation in mammary glands formed by transplant of epithelium from mt/mt or wt/wt animals into the mammary fat pads of prepubescent wild type mice cleared of endogenous epithelium. We again observed a statistically significant increase in Stat1 phosphorylation in mt/mt transplants in the pre-partum period (transplants can’t interrogate the post partum period), demonstrating that the ENU-mutation operates autonomously via the mammary epithelial cell (Fig 1P).
We used the Affymetrix Mouse Transcriptome Assay (MTA) 1.0 GeneChip to measure changes in gene expression underlying these events. We profiled RNA transcripts in the mammary glands at 18dpc and 2dpp from wt/wt and mt/mt mice. A Gene Set Enrichment Analysis (GSEA) of genes was carried out using the Limma t-statistic as a measure of ranked differential expression and visualized with the Enrichment Map plugin for Cytoscape. We compared gene expression changes between mt/mt and wt/wt mice at 18dpc or 2dpp (Fig 2, shown in detail S2 Fig). This identified a robust enrichment of a prominent cluster of gene sets involved in the interferon response in postpartum mt/mt but not wt/wt mammary glands, which increased in magnitude between 18dpc and 2dpp. Genes in these sets included the interferon-induced genes Isg15, Mx1, Rsad2, Oas1, Oas2 and OasL1. Interferon-induced genes involved in the molecular pattern response pathway were also induced, such as Ddx58 (RIG-1), Dhx58 (RIG-1 regulator), Mavs and Nlrc5 (NOD5). Additional downstream transcriptional regulators of the interferon response, such as Stat1, Irf7 and Irf9, were upregulated. In mt/mt glands this was accompanied by increased expression of a broad range of mitochondria-associated cell death genes such as Tnsfs10 (TRAIL), Acin1, Birc2, Traf2, Bcl2l1 (BCL-XL), Bcl2l11 (BIM), Apaf1, Dffb, Xaf and Ripk1. Very similar results were obtained using an independent analysis technique based on self-organizing maps (S3 Fig). These results are also presented as a.txt table (S1 Table) of the 5000 probes showing most-changed expression. This transcriptional data indicates that a robust interferon response is induced by the mutation.
PCR genotyping of polymorphic markers and their co-inheritance with lactation failure narrowed the mutation to a 4Mb region of chromosome 5 between rs3662655 and rs2020515. We expected 4–8 ENU mutations per 4Mb and our strategy was to sequentially sequence exomes and then to experimentally validate when an exonic mutation was discovered. Sequencing revealed a T to A base change in Oas2, resulting in a non-conservative isoleucine to asparagine amino acid substitution (I405N; Fig 3A and S4A Fig) in a conserved region of the OAS2 catalytic domain (S4B and S4C Fig). In wt/wt animals Oas2 was expressed at a relatively low level until the establishment of lactation, when the level of Oas2 mRNA increased by 20 fold (S5A Fig) and subsequently fell during early involution. Changes in Oas2 expression in wt/wt animals compared to mt/mt animals are shown in S1E Fig. Using immunohistochemistry we observed corresponding changes in OAS2 levels in the mammary epithelium (S5B Fig).
We measured RNase L activity in the mammary glands of wt/wt and mt/mt mice at 18 days post coitus (dpc) and 2 days post partum (dpp) using a recently developed technique [9]. In wt/wt mice we observed a fall in RNase L activity from pre lactation at 18 dpp to lactation at 2 dpp despite the rise in OAS2 over this period (Fig 3B top panel). In contrast mt/mt animals showed an increase in RNase L activity over this period, so that at 2 dpp, RNase L activity was 34 fold higher in mt/mt animals. PCR for RNase L-cleaved rRNA showed a six-fold increase RNase L activity (Fig 3B lower panel), while non-RNase L generated cleavage was negligible. Bioanalyzer profiles of RNA (S5C Fig) showed increased RNA degradation in mt/mt animals, but not to the extent that appreciable loss of the 18S or 28S ribosomes was seen, and which may be a result of both RNase L dependent and independent mechanisms. Although robust activation of RNase L can cause the loss of the 18S and 28S ribosomes [10], recent findings show that ribosomal degradation is not required for RNase L to stop protein synthesis [9]. To determine if the mutation altered OAS2 enzyme activity we purified the mutant and wild type forms of mouse OAS2 expressed in HeLa cells by FLAG-immunoprecipitation. Using a cell-free system we observed that both mutant and wild type forms of OAS2 showed induction of enzyme activity by the double-stranded RNA mimic poly (I:C), seen as the formation of a series of 2-5A species resolved by denaturing PAGE. Both mutant and wild type forms of OAS2 showed similar sensitivity to increasing poly (I:C) concentrations (Fig 3C, quantified in Fig 3D). Western blotting showed that the immunoprecipitates used in these experiments had similar OAS2 levels (Fig 3E). These experiments show that the ENU-induced mutation in OAS2 does not change the size range of oligoadenylates that it produces, its capacity for 2-5A synthesis, or its sensitivity to activation by poly (I:C). This assay uses a cell free system, so we cannot exclude a mechanism where mutant OAS2 activates RNase L activity via an indirect effect to increase the active 2-5A pool without altering its rate of synthesis, such as reduced 2-5A depletion or loss of 2-5A sequestration. Another possibility is that mutant OAS2 has an altered molecular interaction with a species that increases its enzymatic activity, but which is lost in the immunoprecipitation of OAS2 in this assay. Regardless, the mutation in Oas2 activates RNase L in mice and tissue culture models.
We constructed a model of doxycycline (Dox)-inducible expression of mutant or wild type Oas2 in T47D human breast cancer cells (Fig 4A). These models produced a 20-fold induction of Oas2 expression (Fig 4B). Western analysis showed the appearance of mouse OAS2 protein following Dox administration just below endogenous human OAS2, both above a non-specific band (Fig 4C). Thus although PCR showed a small amount of leakiness in this system it seems negligible by western blot. Cells expressing either mutant or wild type Oas2 showed a similar sensitivity to poly (I:C) that was not changed significantly by induction with Dox (Fig 4D). Induction of mutant, but not wild type Oas2 for 72 h reduced cell number (Fig 4E). Increased cell detachment was observed, but the magnitude of this effect was highly variable between experiments using mutant cells and so did not reach statistical significance at p<0.05 (Fig 4F). Reduced re-plating efficiency however following trypsinization was significant, indicting that cell surface re-expression of adhesion molecules following their trypsin digestion was impaired (Fig 4G). Western analysis of two of these molecules, Beta-1 Integrin (ß1) and E-Cadherin (EC), showed reduced expression in response to Dox-induction of mutant Oas2, especially for Beta-1 Integrin, shown in the far right hand side lane (Fig 4H). We used flow cytometry to simultaneously measure cell viability by propidium iodide exclusion and cell death by cell surface expression of Annexin V, in response to Oas2 expression. While induction of wild type Oas2 expression produced no apoptotic response, induction of mutant Oas2 produced a doubling in the number early apoptotic cells within the cultures (Fig 4I). Induction of wild type Oas2 did not alter the distribution of cells among the phases of the cell cycle, while induction of the mutant produced a shift of cells out of S-phase and into G1 (Fig 4J). Thus the effects of mutant Oas2 expression in T47D cells reproduce the phenotypes of cell death and reduced cell proliferation seen in the mouse, and indicate that epithelial cell adhesion may also be affected.
Mouse HC-11 cells express milk proteins in response to withdrawal of EGF and the addition of prolactin and dexamethasone, providing a way to examine the effects of mutant and wild type Oas2 expression on milk protein expression. The inducible vector system used successfully in T47D cells (Fig 4A) proved to be very leaky in HC-11 cells, resulting in high baseline expression of Oas2 in the pooled clones without DOX treatment. Cloning, in an attempt to find cells without leaky expression, was unsuccessful, but resulted in cell lines with similar levels of constitutive expression of mutant or wild type Oas2 that was many fold greater than seen in untransfected cells (Fig 4K). Treatment with prolactin and dexamethasone induced beta casein levels in the cell line expressing wild type Oas2, and this effect was comparable in magnitude to that seen in parental HC-11 cells, but in the two lines expressing mutant Oas2 the induction of beta casein was greatly reduced (Fig 4L), reproducing the suppression of milk protein synthesis seen in the ENU-mutant mouse. Transient expression of wild type and mutant Oas2 in HC11 cells also showed an increase in the basal rate of cell death, reproducing the cell death phenotype (Fig 4M).
We used Affymetrix Human Transcriptome Assay 2.0 GeneChips to profile the changes in gene expression that occurred in T47D cells when either wild type or mutant Oas2 was induced for 72h, presented as GSEA/Cytoscape (S6 Fig), self organizing maps (S7 Fig) and as table containing the top 500 differentially expressed genes (S2 Table). We compared the transcriptional effects of mutant Oas2 in T47D cells to the effects in the ENU mouse shown in Fig 2 using Cytoscape (S8 Fig), or self-organizing maps (S7 Fig). The transcriptional effects of mutant OAS2 in T47D cells were very similar to those observed in the ENU-mutant mouse, with the interferon response most prominent. This demonstrates that expression of mutant but not wild type Oas2 in T47D cells reproduces the molecular phenotypes observed in the ENU mutant mice. While the phenotype in mice is likely to involve additional cells of the immune system, these effects in T47D cells show that the transcriptional phenotype can be elicited via the innate immune response of the mammary epithelial cell, in agreement with the findings made using transplanted ENU-mutant mammary epithelium into wild type mice (Fig 1P).
OAS2 activates RNaseL. In T47D cells we used siRNA against human RNASEL to knockdown its expression in the context of Dox-induction of mutant or wild type mouse Oas2. In these experiments the induction of Oas2 in response to Dox was robust and knockdown of RNASEL was very effective, as demonstrated by qPCR (Fig 5A and 5B) and by western blot (Fig 5C). Induction of wild type Oas2 had no effect on RNase L activity, cell death or cytokine levels and knockdown of RNASEL was without consequence to these endpoints. In contrast, induction of mutant Oas2 produced a large increase in RNase L activity, cell death, and interferon gamma and GM-CSF protein secretion, changes that were prevented by knockdown of RNASEL (Fig 5D–5G).
Expression of the IRF transcription factors, especially IRF7, was increased by mutant OAS2. We knocked down IRF7 (Fig 5H) and found a similar prevention of cell death (Fig 5I), indicating that the signaling pathway activated by mutant OAS2 also involves IRF7, a distal member of the viral-detection signaling pathway. Knockdown of IRF3 (Fig 5J), which often acts together with IRF7, had the opposite effect (Fig 5K), suggesting IRF3 acts to oppose signaling via the OAS2 pathway.
These experiments show that the Oas2 mutation caused activation of OAS2 driven signaling to prevent the activation of lactation in the post partum period. The effect of the mutation could be detected via Stat1 activation from mid pregnancy and was most apparent in the post partum period, and was only required in the mammary epithelial cell for effect. The mutation increased RNase L activity but the enzymatic activity of mutant OAS2 was unaltered. Thus RNase L activation must occur via mechanisms that increase the effect of 2-5A without a change in its production, such as by reducing 2-5A degradation, increasing the efficiency of 2-5A interaction with RNase L or OAS2 interaction with dsRNA, or by causing relief of a mechanism that sequesters 2-5A. The activation of RNAse L is not sufficient to degrade the ribosomes, indicating that the loss of milk production does not occur via a generalized loss of translation. Thus while RNase L expression is required for activity of the mutation, the mutation may act via regulatory mechanisms that do not require 2-5A activation of RNase L. RNase L may be simply permissive of an alternative mechanism of action, such as altering interactions of OAS2 with its cellular binding partners, by changing its subcellular localization, or by decreasing the rate of OAS2 degradation. Thus it is possible that RNase L and OAS2 could also both be involved in as yet undiscovered molecular complexes that initiate activation of this pathway. For example OAS2 has been reported to bind NOD2 [3], and the composition and mechanism of action of this mitochondrial-signaling complex is currently the subject of intense worldwide study, but its definition is proving to be elusive. The non catalytic OAS1b [11] and OASL1 [12] have mechanisms independent of 2-5A production involving molecular interactions. This is the first genetic demonstration that OAS2 can signal in ways other than by alterations in enzyme activity. This mutation may prove to be important for the discovery of the mechanisms signaling the detection of viral infection, which remain largely unknown, because it provides a single point of pathway activation, unlike the existing reagents used for this purpose. Like other family members, OAS2 may regulate apoptosis independently of its function to control viral replication [6,7].
Lactation failure and milk stasis characterize mastitis, raising an interesting new avenue of investigation opened by our findings. The major consequence of mastitis is reduced weight-gain of the infant, precipitating a switch to bottle-feeding where available, or reduced neonatal health where it is not. Our results raise the possibility that the OAS2 pathway may be involved in its pathogenesis. Bacterial infection is commonly thought to be the cause of mastitis but the evidence resoundingly shows that bacterial infection is the sequelae of an unknown primary cause of the disease. For example, in women the severity of symptoms of mastitis do not correlate with the level of bacterial infection, the disease is often observed in the absence of bacteria in the milk, bacteria are often found in the milk of healthy mothers, and meticulous hygiene or prophylactic antibiotics do not prevent mastitis (reviewed [13]). Recent Cochrane Systematic Reviews concluded that there is insufficient evidence to support antibiotic use for the prevention [14] or treatment [15] of mastitis. The strongest risk factors for mastitis in women involve incomplete or interrupted milk flow from one or more galactophores [13] and the World Health Organization recognizes milk stasis as the cause of mastitis [16]. Thus bacterial infection most likely represents progression of mastitis to a more pathogenic form involving abscess formation, but it is not the primary cause.
The concept of physiological inflammation as the primary cause of mastitis was proposed in 2001, though no mechanism was proposed at the time [17], and the unavailability of breast tissue from women with mastitis makes the study of mechanism near impossible. Using mice, Ingman and colleagues hypothesize that molecular pattern receptors like Tlr4 recognize molecules released by tissue damage caused by milk engorgement, which trigger an innate immune response and milk stasis [13,18]. Alleles of Tlr4, a bacterial associated molecular pattern receptor, are linked with the occurrence of mastitis in cattle [19]. Tlr4 has also been linked to a number of the systemic symptoms of mastitis [13]. As we show, stimulation of the OAS2 pathway can produce the accepted cause of mastitis, milk stasis, opening a new avenue of investigation into human mastitis as a disease amenable to anti-inflammatory therapy. Our findings also open the question of the role of viruses in the initiation of mastitis. Even non-infectious forms could play a role. Fragments of the mouse mammary tumor virus are present in the genome of all laboratory mice and they continue to produce transcripts in response to the hormones of pregnancy, while homologous fragments exist in the human genome [20,21], which may promote milk stasis and inflammation via OAS2 activation.
This is the first time that a viral recognition pathway has been implicated in the regulation of lactation. Transmission of viruses via milk is a well-documented phenomenon and the evolution of mechanisms to prevent it would be expected. This would not necessarily be lethal for the neonate as all mammals have multiple and independent lactation systems. Mice, for example have 10 mammary glands each containing a single ductal system. Each human breast contains between 6 and 8 independent ductal systems, exiting at the nipple without joining. Viral infection in one ductal system, or one mammary gland, could initiate milk stasis in that system, leaving the others to continue lactation. Social systems in humans and mice allow the feeding of neonates by multiple mothers. There could be an intriguing evolutionary twist here resulting from the evolutionary arms-race between viruses and their hosts [22]. Since HIV transmission via the milk occurs far more frequently if mastitis is present [23], could viruses have adapted to this defense and learned to induce a limited mastitis to aid viral transmission?
A molecular mechanism is suggested by our results (S9 Fig) that requires further investigation. STAT1 activation (Fig 1), presumably resulting from the production of interferon due to OAS2 pathway activation, would be expected to cause the induction of the SOCS proteins, which inhibit STAT phosphorylation via targeting the JAK kinases, including JAK2 which phosphorylates STAT5 in response to prolactin, the major hormone driving the onset of lactation. Many aspects of this pathway have been demonstrated in mice such as the regulation of lactation by prolactin via STAT5 [24,25] and the SOCS proteins [26–28], the induction of STAT1 in conditions of sterile mastitis [29] and the ability of STAT1 to regulate prolactin signaling [30]. In the T47D transcript profiling (S6–S8 Figs and S2 Table) we observed increases in the levels of SOCS 1,4,5 and 6. In the ENU mice we observed a decrease in STAT5 phosphorylation. So it is possible that OAS2 pathway stimulation, resulting from the natural rise in OAS2 at d18.5 of pregnancy (S1 and S5 Figs), produces a persistent interferon response in OAS2 mutant animals, because mutant OAS2 activates RNase L which via the resulting interferon response maintains high OAS2 levels, establishing a positive feed-back loop which then persistently prevents prolactin from activating STAT5 (maybe via SOCS) to induce the activation of milk secretion during the post partum period.
All mice were housed in specific pathogen-free conditions at the Australian Phenomics Facility and the Garvan Institute, with all animal experiments carried out according to guidelines contained within the NSW (Australia) Animal Research Act 1985, the NSW (Australia) Animal Research Regulation 2010 and the Australian code of practice for the care and use of animals for scientific purposes, (8th Edition 2013, National Health and Medical Research Council (Australia)) and approved by either the Australian National University or Garvan/St Vincent’s Animal Ethics and Experimentation Committees (approval number 14/27).
ENU mutagenesis and pedigree construction was carried out as previously described [31]. The Oas2 mutation was discovered in a single G1 female and heritability of the phenotype confirmed by breeding with CBA CaJ male and cross fostering of pups. For quantification of lactation failure litters were standardized to 7 pups per dam. Pups were weighed, as a group, at the same time daily. Mice were injected with BrdU dissolved in H2O (100μg BrdU per gram body weight) 2 h prior to sacrifice by CO2 asphyxiation, and mammary glands were collected. Mammary glands were either whole-mounted and stained with Carmine alum or snap frozen in liquid nitrogen for mRNA and protein analyses. All animals were housed with food and water ad libitum with a 12-h day/night cycle at 22°C and 80% relative humidity.
A complete analysis of the histology and pathology of the Jersey strain was conducted by the Australian Phenomics Network (APN) Histopathology and Organ Pathology Service, University of Melbourne. Eight week and a 31 week female sibling pairs, comprised of mt/mt and wt/wt siblings, were examined macro and microscopically. Mammary tissue, ovaries, oviducts, bicornuate uterus, cervix, urinary bladder, liver/gall bladder, cecum, colon, spleen/pancreas, mesenteric lymph node, stomach, duodenum, jejunum, ileum, kidney/adrenal, salivary glands/lymph nodes, thymus, lungs, heart, skin, eyes, brain, spinal cord, skeletal muscle, skeletal tissue/hind leg were macro and microscopically examined.
A pool of 15 affected N2 mice and a pool of 15 unaffected N2 backcrossed mice were screened with a set of ~130 markers polymorphic between C57BL/6 and CBA/CaJ mice that spanned the genome at 10–20 Mb intervals. Allele specific SNP primers were designed from a set of validated SNPs available at www.well.ox.ac.uk/mouse/INBREDS/. SNP genotyping was performed using the Amplifluor kit (Chemicon) as per the manufacturers instructions. The confirmation and fine mapping were performed using Amplifluor markers designed to amplify C57BL/6 x CBA/CaJ SNPs within the linkage interval in individual affected and unaffected mice. Markers were designed at approximately 1–2 Mb distances within the initial map interval. More than 250 mice were screened from many successive cohorts of mice from backcrosses to CBA/CaJ to narrow the region to a 3 Mb interval. Sequencing of candidate genes was performed to locate the causal ENU base substitution using an affected mouse homozygous for the linkage region. Primers were designed for candidate genes to amplify all exons +/- 15 bp to cover splice junctions. Sanger sequencing was used to identify the causal mutation by comparing the sequence of the affected individual against a C57BL/6 mouse reference genome. Mutations were confirmed in a second affected individual and a C57BL/6wild type mouse.
The mapping was performed with the gsMapper program, which is part of the 454 software suite. The two samples (defined by Jersey_F4IC140 and Jersey_pool) were mapped against the full region of the mouse genome on chromosome 5. The sequence used as reference is from genbank build37/UCSC mm9. Further analysis then focused on the reads mapped onto the target region: 118710087–123738720 on chromosome5.
Variation analysis detected where at least 2 reads differ either from the reference sequence or from other reads aligned at a specific location. SNPs, insertion-deletion pairs, multi-homopolymer insertion or deletion regions, and single-base overcalls and under calls are reported. Also, in order for a difference to be identified and reported, there have been at least two non-duplicate reads that (1) show the difference, (2) have at least 5 bases on both sides of the difference, and (3) have few other isolated sequence differences in the read. Variations were classified as high-confidence if they fulfilled the following rules: 1. There must be at least 3 non-duplicate reads with the difference. 2. There must be both forward and reverse reads showing the difference, unless there are at least 5 reads with quality scores over 20 (or 30 if the difference involves a 5-mer or higher). 3. If the difference is a single-base overcall or under call, then the reads with the difference must form the consensus of the sequenced reads (i.e., at that location, the overall consensus must differ from the reference).
After identification of the causative mutation genotyping was performed using the following primers:
The Oas2 mutant colony was maintained by breeding heterozygous males (wt/mt) with wt/wt females. For the generation of homozygous experimental animals wt/mt males were bred with wt/mt females and their offspring removed at 1dpp and fostered on a control mother.
Total RNA was isolated using Trizol reagent (mouse tissues; Gibco/Invitrogen Vic) or RNeasy extraction kit (cell pellets; Qiagen) according to the manufacturer’s instructions. All total RNA samples were quantified with a Nanodrop 1000 Spectrophotometer (ThermoFisher) prior to loading 100 ng of total RNA on a 2100 Bioanalyzer (Agilent) total RNA analysis. Single stranded cDNA was produced by reverse transcription using 1 μg of RNA in a 20μl reaction and diluted 1:5 with H2O (Promega). Quantitative PCR was performed using the Taqman probe-based system (Table 1) on the ABI 7900HT as per the manufacturer’s instructions (Applied Biosystems).
tRNA cleavage by RNaseL was performed using the technique developed by JD and AK [9]. Briefly, RNA was ligated with 2’,3’-cyclic phosphate to an adaptor with RtcB as described in [9]. Reactions were stopped by adding EDTA and used as a template for reverse transcription with Multiscribe RT. A primer with a 3’-end complimentary to the adaptor and a 5’-overhang that serves as a universal priming site (5’-TCCCTATCAGTGATAGAGAGTTCAGAGTTCTACAGTCCG- 3’) was used for reverse transcription. SYBR green-based qPCR was conducted using a universal reverse primer that binds to the cDNA overhang (underlined) and cleavage-site specific forward primers designed for tRNA [9]. qPCR was carried out for 50 cycles using 62°C annealing/extension for 1 min. U6, which has a naturally occurring 2’,3’-cyclic phosphate and an RNase L independent cleavage site in tRNA-His (position 18, transcript numbering; [9]) was used for normalization.
HeLa cells were maintained in MEM + 10% FBS in a humidified 5% CO2 atmosphere. Cells cultured in 10 cm dishes were transfected at 80–90% confluence with 10 μg empty pcDNA4/TO or N-terminally FLAG-tagged WT or I405N mouse OAS2 in pcDNA4/TO (Life Technologies) using Lipofectamine 2000 (Life Technologies). Cells were harvested by trypsinization 24h post-transfection, resuspended in complete media, and washed 2 x 10 mL ice-cold PBS. Cell pellets were lysed in buffer A (20 mM HEPES pH 7.5, 100 mM NaCl, 0.1% Triton X100, and 1x complete protease inhibitor cocktail (Roche) for 10 min with end-over-end rotation at 4°C. Lysates were cleared by centrifugation at 16,000 x g, 15 min, 4°C and the supernatants subjected to immunoprecipitation with M2-α-FLAG magnetic beads (Sigma) for 2h at 4°C followed by 4 x 1mL washes, 10 min per wash, with buffer A without protease inhibitors. After the fourth wash the beads were washed with 2 x 1 mL storage buffer (20 mM HEPES pH 7.5, 100 mM NaCl, 10% glycerol), the supernatant removed, storage buffer added to 100 μL. Inputs and IPs were blotted with α-FLAG to verify expression and IP of FLAG-OAS2. 5% of each IP was incubated with 1 mM ATP, trace-labeled with 3 nM 32P-α-ATP, in the absence or presence poly (I:C) (Sigma) for 2h at 37°C. Reaction volumes were 20 μL and contained 20 mM HEPES pH 7.5, 70 mM NaCl, 10 mM MgCl2, 10% glycerol, and 4 mM DTT. After incubation, the reactions were quenched by adding 120 μL stop buffer (8M urea, 0.1% SDS, 1 mM EDTA, 0.02% bromophenol blue, 0.02% xylene cyanol). Equal portions of each reaction were resolved by 20% denaturing PAGE and visualized by phosphoimaging. Gels were quantitated using GelQuant.NET software.
The N—terminal and C—terminal cDNAs of mouse Oas2 were obtained as a gift of Yoichiro Iwakura (Institute of Medical Science, University Tokyo) and subcloned into pcDNA3.1 and pBluescript vectors. Site directed mutagenesis was performed using Phusion Site-directed mutagenesis (Thermo Scientific) as per the manufacturer’s instructions using the following primers (Forward Jer: TATATGTTCCTTCCTTAAAAATGTCTGC and Reverse AGGATTTCGTCTTGTTCCTTCGACAACTGTA). Wildtype and mutant (I405N) mouse Oas2 clones were then subcloned into pShuttle and finally into the pHUSH ProEx tetracycline inducible retroviral expression system [32]. Retrovirus was then packaged by transfecting Phoenix cells with pHUSH containing either mouse wildtype (wt) or Oas2 I405N mutant (mt) cDNAs using Fugene transfection reagent (Promega). T47D breast cancer cells were then infected with filtered viral supernatants and stable cell lines selected using Puromycin. T47Ds were maintained sub-confluent in RPMI complete media (Gibco) containing 10% tetracycline free FCS and supplemented with 10μg/ml Insulin. Mouse Oas2 wt or mt expression was induced with 100 ng/ml Doxycline (DOX) or vehicle control daily in the media and cells were harvested at 72 hours after plating. Cell counts of viable and non-viable cells (identified by the incorporation of 0.4% Trypan Blue at a 1:2 dilution) were performed in triplicate from 3 independent experiments. Annexin V PI staining was performed using the Annexin V-FITC Apoptosis Kit (Biovision, CA USA) as per the manufacturers instructions. Human inflammatory cytokines were analyzed using the Multi-Analyte ELISArray (Qiagen).
Retrovirus was made packaged as above by transfecting Phoenix cells with pHUSH containing either mouse wildtype (wt) or Oas2 I405N mutant (mt) cDNAs using Fugene transfection reagent (Promega). HC11 normal mouse mammary cell lines were then infected with filtered viral supernatants and stable constitutive cell lines selected using Puromycin and then clonal colonies established by titrating single cells into 96 well plates. HC11s were maintained in maintenance RPMI media containing 10% FCS and supplemented with 5μg/ml Insulin and 10ng/ml human recombinant epidermal growth factor (EGF, Sigma-Aldrich). For differentiation assays, HC11 cells were plated sub confluent in maintenance media for 2 days until confluent and then media replaced with RPMI media containing 10% FCS and supplemented with 5μg/ml Insulin and supplemented with 5μg/ml sheep pituitary Prolactin (Sigma Aldrich) and 1μM Dexamethasone (Sigma Aldrich) daily for 3 days before RNA harvest and quantitative PCR analysis as above. For transient transfections apoptotic assays wildtype (wt) or Oas2 I405N mutant (mt) cDNAs were cloned into pIRES-EGFP vectors and transiently transfected with 2μg DNA/well using X-tremeGENE transfection reagent (Roche) in maintenance media as per the manufacturers instructions. After 24 hours the media replaced with RPMI media containing 10% FCS and supplemented with 5μg/ml Insulin and cells were harvested 96 hours after transfection for Annexin V PI staining as above.
T47D Oas2 wt or mt cells were plated in 10cm dishes and treated daily for 48 hours with 100 ng/ml DOX or vehicle. At 48 hours the cells harvested by trypsinization, retreated with DOX or vehicle, counted, plated at a density of 5000 cells/well and simultaneously transfected with Poly (I:C) (11 point titration) using RNAiMAX (Invitrogen) in opaque 96 well plates. 24 hours after transfection the plates were analysed using the CellTiter-Glo Luminescent Cell Viability Assay Protocol (Promega). Inhibitory dose curves were plotted in Prism 6 statistical software and normalized data analyzed using the sigmoidal-dose response function. The mean was calculated from quadruplicate replicates.
T47D Oas2 wt or mt cells were plated in T150 flasks and mouse Oas2 wt or mt expression was induced with 100 ng/ml DOX or vehicle for 48 hours prior to trypsinisation by 0.25% Trypsin (no EDTA) with phenol red (Life Technologies) for 3 mins. 1 x 106 cells from each cell line was treated with DOX or vehicle and then plated in 6-well plates and allowed to adhere for 4 hours. After 4 hours the cells were gently washed, trypsinised and the number of adherent cells counted and expressed as a proportion of the total number of cells plated.
T47D Oas2 wt or mt cells were plated in 6-well plates and mouse Oas2 wt or mt expression was induced with 100 ng/ml DOX or vehicle for 72 hours. Cells were harvested by trypsinisation, washed in 1 ml of PBS and fixed by adding 10 mls of 100% cold ethanol drop wise onto 1ml re-suspended cells and incubated at 4°C overnight. Cells were then pelleted, washed and incubated at 90°C for 5 mins and then re-suspended in a FACS buffer containing 0.5ng/ml RNase (Qiagen) and 1μg/ml Propidium iodide. Flow cytometry was performed and G1, S phase and G2/M phases for each experimental group estimated using the propidium iodide fluorescence intensity histograms. The mean of 5 independent experiments was calculated.
T47D Oas2 wt or mt cells were reverse transfected with ON-TARGET plus SMARTPOOLS of siRNA against RNaseL (L-005032-01-05) or Non Targeting controls (D-001810-10-05) using RNAiMAX (Invitrogen) in 10cm dishes as per the manufacturers specification. 24 hours after transfection, cells were washed, media replaced and either treated with 100 ng/ml DOX or vehicle daily for 3 days after which they were harvested by trypsinisation. Annexin V PI staining was performed using the Annexin V-FITC Apoptosis Kit (Biovision, CA USA) as per the manufacturers instructions. Cell pellets were also collected for RNA isolation and western blotting and the supernatant collected, filtered with a 0.22μm filter and stored at -80°C. Human inflammatory cytokines were analyzed using the Multi-Analyte ELISArray™ Kits as per the manufacturers instructions.
Wt/wt or mt/mt mice were time mated and mammary glands collected at day 18 of pregnancy or 2 days after partuition (2dpp) and snap frozen in liquid N2. Total RNA was isolated using Trizol reagent (Gibco/Invitrogen, Vic) and measured on the 2100 Bioanalyzer (Agilent). From the cell experiments, total RNA was extracted using the RNeasy extraction kit (Qiagen) for cells with or without DOX induction of the wt or mt mOas2 gene. Total RNA from the mouse mammary glands was then labeled and hybridized to the Mouse Transcriptome Array (MTA) 1.0 as per the manufacturer’s instructions (Affymetrix Ca, USA) at Ramiaciotti Centre for Genomics (UNSW Sydney, Australia). Likewise, total RNA from the T47D cells was labeled and hybridized to the Affymetrix Human Transcriptome Array (HTA) 2.0 as per the manufacturer’s instructions (Affymetrix Ca, USA) at Ramiaciotti Centre for Genomics (UNSW Sydney, Australia). All mouse and T47D samples were prepared in biological triplicate for each experimental grouping, except for the T47D mt -DOX group where analysis was performed in duplicate due to one of the samples failing quality control. Microarray data are freely available from GEO: GSE69397 http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?token=ohkleoespjotzoh&acc=GSE69397
Quality control was performed using the Affymetrix Expression Console. Normalisation and probe-set summarization was performed using the robust multichip average method of the Affymetrix Power Tools apt-probeset-summarize software (version 1.16.1) (using the -a rma option). The transcript clusters with official HGNC symbols were then extracted from the HTA 2.0 arrays, resulting in 23532 gene transcript clusters. Differential expression between experimental groups was assessed using Limma [33] via the limmaGP tool in GenePattern. Functionally associated gene-sets were identified using Gene Set Enrichment Analysis (GSEA) [34] on a ranked list of the limma moderated t-statistics, from each pair-wise comparison, against a combined set of 6947 gene-sets from v4.0 of the MSigDB [35] and custom gene-sets derived from the literature. Mouse gene-symbols were mapped to their human orthologs using the ensembl database. The Enrichment Map plugin [36] for Cytoscape [37] was used to build and visualize the resulting regulatory network of gene-signatures, with conservative parameters: p = 0.001; q = 0.05; overlap s = 0.5.
10μg reduced protein was loaded in each well of 12% NuPAGE SDS polyacrylamide gels (Life Technologies) and separated using electrophoresis. Proteins were transferred to Immun Blot PVDF (Biorad) and Western blotted for mouse Oas2 (M-105, sc99098 Santa-Cruz), RNaseL (H-300, sc25798 Santa Cruz), E-cadherin (610182 BD Biosciences) and beta-ACTIN (AC-74, A5316, Santa Cruz)
The limma F-test statistic [33], with a Benjamini-Hochberg adjusted p-value threshold of 0.05, was used to identify differentially expressed transcripts across the four experimental groups in the mouse expression arrays (wt/wt 2dpc, wt/wt 2dpp, mt/mt 2dpc, mt/mt 2dpp) and T47D cell-line expression arrays (wt–Dox, wt +Dox, mt–Dox, mt +Dox). This resulted in 660 and 135 significant transcript clusters from the mouse and T47D arrays, respectively.
Self-organising maps (SOMs), consisting of 6 nodes, were used to identify clusters of genes in both the mouse and T47D cells. The z-scores of the log2 normalised gene-expression values, for each transcript cluster, were used as input to the biopython SOM algorithm implementation [38]. The somcluster() parameters used were: iterations = 50,000; nx = 2, ny = 3, inittau = 0.02, dist = Euclidean.
The db2db() function from the BioDBNet database [39] was used to convert gene-symbols to Ensembl gene IDs for input into DAVID. Functional annotation clustering was carried out using the getTermClusterReport() function from the DAVID web services interface (Jiao et al., 2012), with the following parameters: overlap = 3, initialSeed = 3, finalSeed = 3, linkage = 0.5, kappa = 50.
DAVID databases used: (BBID, GOTERM_CC_FAT, BIOCARTA, GOTERM_MF_FAT, SMART, COG_ONTOLOGY, SP_PIR_KEYWORDS, KEGG_PATHWAY, INTERPRO, UP_SEQ_FEATURE, OMIM_DISEASE, GOTERM_BP_FAT, PIR_SUPERFAMILY)
A Hypergeometric test was used to calculate the level of gene overlap between the genes identified in each SOM cluster and the MSigDB gene-set collections [35] and our custom functional signature gene-sets. A background set number, of 45956, as described on the MSigDB website, was used. A Benjamini-Hochberg (BH) corrected p-value was calculated for each set and a threshold of BH<0.05 was considered a significant enrichment. Mouse gene-symbols were mapped to their human orthologs using the ensembl database.
Mouse mammary glands were harvested from wt/wt and mt/mt mice and fixed in 4% buffered formalin for 4 hours. Glands were defatted in 3–4 changes of acetone before being dehydrated and stained in Carmine alum as previously described [40]. Glands were then dehydrated in a series of graded alcohols and embedded in Paraffin for sectioning. Sections were either stained with haematoxylin and eosin for routine histochemistry or stained with antibodies to the following antigens using immunohistochemistry protocols as detailed in Table 2.
|
10.1371/journal.pgen.1003792 | Nebula/DSCR1 Upregulation Delays Neurodegeneration and Protects against APP-Induced Axonal Transport Defects by Restoring Calcineurin and GSK-3β Signaling | Post-mortem brains from Down syndrome (DS) and Alzheimer's disease (AD) patients show an upregulation of the Down syndrome critical region 1 protein (DSCR1), but its contribution to AD is not known. To gain insights into the role of DSCR1 in AD, we explored the functional interaction between DSCR1 and the amyloid precursor protein (APP), which is known to cause AD when duplicated or upregulated in DS. We find that the Drosophila homolog of DSCR1, Nebula, delays neurodegeneration and ameliorates axonal transport defects caused by APP overexpression. Live-imaging reveals that Nebula facilitates the transport of synaptic proteins and mitochondria affected by APP upregulation. Furthermore, we show that Nebula upregulation protects against axonal transport defects by restoring calcineurin and GSK-3β signaling altered by APP overexpression, thereby preserving cargo-motor interactions. As impaired transport of essential organelles caused by APP perturbation is thought to be an underlying cause of synaptic failure and neurodegeneration in AD, our findings imply that correcting calcineurin and GSK-3β signaling can prevent APP-induced pathologies. Our data further suggest that upregulation of Nebula/DSCR1 is neuroprotective in the presence of APP upregulation and provides evidence for calcineurin inhibition as a novel target for therapeutic intervention in preventing axonal transport impairments associated with AD.
| Alzheimer's disease (AD) is a debilitating neurodegenerative disease characterized by gradual neuronal cell loss and memory decline. Importantly, Down syndrome (DS) individuals over 40 years of age almost always develop neuropathological features of AD, although most do not develop dementia until at least two decades later. These findings suggest that DS and AD may share common genetic causes and that a neuroprotective mechanism may delay neurodegeneration and cognitive decline. It has been shown that the amyloid precursor protein (APP), which is associated with AD when duplicated and upregulated in DS, is a key gene contributing to AD pathologies and axonal transport abnormalities. Here, using fruit fly as a simple model organism, we examined the role of Down syndrome critical region 1 (DSCR1), another gene located on chromosome 21 and upregulated in both DS and AD, in modulating APP phenotypes. We find that upregulation of DSCR1 (Nebula in flies) is neuroprotective in the presence of APP upregulation. We report that nebula overexpression delays the onset of neurodegeneration and transport blockage in neuronal cells. Our results further suggest that signaling pathways downstream of DSCR1 may be potential therapeutic targets for AD.
| Virtually all Down syndrome (DS) adults develop progressive neurodegeneration as seen in Alzheimer's disease (AD), and overexpression of the amyloid precursor protein (APP), a gene located on chromosome 21, is thought to contribute to AD in DS [1]–[3]. Consistently, duplication of a normal copy of APP is sufficient to cause familial AD [4], [5], confirming that it is a key gene in AD neuropathologies seen in DS. This well-known connection between AD and DS provides a unique opportunity to identify the genetic and molecular pathways contributing to AD. In addition to APP, another gene likely to play a crucial role in both AD and DS is the Down syndrome critical region 1 gene (DSCR1, also known as RCAN1). Intriguingly, post-mortem brains from AD patients show increased DSCR1 both at mRNA and protein levels [6]–[8]. Studies have also shown that oxidative stress and Aβ42 exposure can induce DSCR1 expression [8], [9]. DSCR1 is located on human chromosome 21 and encodes a highly conserved calcineurin inhibitor family called calcipressin [10]–[15]. DSCR1 has been implicated paradoxically in both promoting cell survival in response to oxidative stress and in inducing apoptosis [8], [9], [16], [17]. The role of DSCR1 in AD thus remains unclear and an important question is whether DSCR1 contributes to AD or plays a role in combating the toxic effects of APP overexpression.
To elucidate the role of DSCR1 in modulating APP-induced phenotypes, we used Drosophila as a model system, which has been used successfully to investigate various human neurodegenerative diseases including AD, Parkinson's, and polyglutamine-repeat diseases [18]–[27]. Overexpression of APP in both fly and mouse models have previously been shown to cause age-dependent neurodegeneration and axonal transport defects [28]–[31]. Furthermore, impaired transport of essential organelles and synaptic vesicles caused by APP perturbation is thought to be an underlying cause of synaptic failure and neurodegeneration in AD [32]–[34]. However, mechanisms for how APP induces transport defects remain unclear. Here, we show that Nebula, the fly homolog of DSCR1, delays neurodegeneration and reduces axonal transport defects caused by APP overexpression. We report that Nebula enhances anterograde and retrograde axonal trafficking as well as the delivery of synaptic proteins to the synaptic terminal. We find that APP upregulation elevates calcineurin activity and GSK-3β signaling, but Nebula co-upregulation corrects altered signaling to restore axonal transport. Together, our results indicate that Nebula/DSCR1 upregulation is neuroprotective in the presence of APP overexpression and further suggest that Nebula/DSCR1 upregulation may delay AD progression. In addition, our results for the first time link defective calcineurin signaling to altered axonal transport and imply that restoring calcineurin and GSK-3β signaling may be a feasible strategy for treating AD phenotypes caused by APP upregulation.
To examine the role of DSCR1 in modulating APP-induced neurodegeneration and axonal transport defects, we generated transgenic flies containing UAS-APP (APP) in the presence or absence of UAS-nebula (nlat1) [15]. Targeted expression of human APP in the fly eyes using the Gmr-GAL4 driver caused age-dependent degeneration of the photoreceptor neurons, consistent with a previous report by Greeve et al [35]. As seen in Fig. 1A, staining with an antibody specific for the photoreceptor neurons (24B10) and antibody against the APP protein (6E10) revealed the presence of vacuoles in the retina (arrow). Surprisingly, overexpression of nebula together with APP (APP;nlat1) reduced neurodegeneration (as determined by calculating the fold change in the percentage of area lost), suggesting that Nebula upregulation is neuroprotective (Figs. 1A and 1B). By 45-days of age, flies expressing both nebula and APP started to show increased vacuole formation, but the extent of degeneration was significantly reduced compared to that of APP overexpression, further implying that Nebula delays the onset of neurodegeneration rather than completely preventing it.
To confirm that Nebula indeed protects against neurodegeneration caused by APP upregulation, we expressed APP in nla1, a previously characterized nla hypomorphic mutant [15]. Note that because nebula null alleles are lethal [15], nebula hypomorphs were examined. Fig. 1 shows that decreasing Nebula level enhanced APP-induced neurodegeneration in the retina (APP;nla1), thus highlighting the importance of endogenous Nebula protein in conferring neuroprotection. We did not detect significant neurodegeneration in nla1 mutant and nla overexpression flies even by 45 days of age (data not shown), indicating that APP is necessary for the observed phenotype. In addition, mitigation of photoreceptor degeneration by Nebula upregulation is not due to altered expression level of APP, since UAS-LacZ transgene was included to balance out the number of transgenes (we found Gmr-GAL4 is particularly sensitive to number of transgenes). The level of APP protein in each fly line is also further confirmed by staining with the 6E10 antibody (Fig. 1A) and Western blot analyses (Fig. S1). Comparable level of APP was detected in all transgenic lines, suggesting that rescue by nebula overexpression is not due to altered APP level.
We next determined if Nebula rescues functional defects in photoreceptor by measuring the ability of flies to see light. Flies are normally phototactic and will move toward light when placed in test tubes with light source on the opposite end [36]. We find that the severity of the vacuole phenotype was paralleled by impairments in phototactic behavior (Fig. 1C). Flies overexpressing APP showed age-dependent decline in phototaxis that is delayed by APP and nebula expression (Fig. 1C). Taken together, these results imply that nebula overexpression protects neurons structurally as well as functionally against the toxic effects of APP overexpression.
We also noticed that APP overexpression caused formation of APP aggregates in the photoreceptor axons as detected by 6E10 antibody (Fig. 1D; yellow arrow head). Previous studies have shown that APP phosphorylated on threonine 668 (pT668-APP) is preferentially transported in axons [37], we thus further monitored the distribution of pT668-APP. We found that overexpression of APP led to pT668-APP accumulations in the photoreceptor axons, whereas APP and nebula co-overexpression significantly enhanced the delivery of pT668-APP to synaptic terminals in the medulla (Fig. 1D). These results suggest that APP overexpression may lead to blocked transport that is alleviated by Nebula.
Axonal transport abnormalities are thought to precede the onset of AD [30], and APP overexpression has been shown to cause synaptic vesicle accumulations indicative of blocked axonal transport [28], [29]. We thus further investigated the role of Nebula in modulating APP-induced vesicle aggregation in larval motor axons, which is an excellent system for monitoring vesicle transport because of the long axons and stereotypical innervation of the neuromuscular junction (NMJ). As seen in Fig. 2A, APP overexpression in neurons using the Elav-GAL4 driver caused synaptic vesicle accumulation as detected by synaptotagmin staining in the motor axons, suggesting abnormal vesicle transport. Staining using the 4G8 antibody to detect APP revealed that APP aggregates frequently colocalized with synaptotagmin aggregates, implying that synaptotagmin and APP are either comparably inhibited by physical blockade within the nerve or that they are transported together as suggested by recent reports [38], [39]. Co-upregulation of Nebula and APP significantly prevented APP-induced synaptotagmin and APP accumulations. Decreasing Nebula by crossing it into nla1 background increased the number of synaptotagmin and APP aggregates slightly, although not significantly (Fig. 2B). As nla1 only reduces Nebula level by about 30% and that nla null alleles are lethal [15], we used RNAi strategy to further decrease Nebula level (Fig. S2). Figs. 2A–2B show that greater reduction in Nebula level using the UAS-nla-RNAi transgene (RNAi-nla) further exacerbated the APP-induced aggregation phenotype. To ensure that the observed rescue in phenotype is not due to altered APP overexpression, we monitored the level of neuronal APP protein, as well as Nebula, in different fly lines. As seen in Fig. S3, APP level was unaltered in flies containing different number of transgenes, and Nebula manipulations in APP overexpression background showed the expected changes. Similar results were obtained when performing western blot analyses using brains dissected from 3rd instar larvae (Fig. S4). Together, these results confirm that rescue of APP phenotype by Nebula is not due to altered APP expression. In addition, we examined the effect of altering Nebula levels alone on vesicle accumulation. Manipulations of Nebula levels alone did not cause synaptotagmin aggregate accumulation in nerves, suggesting the observed phenotype is APP-dependent (Figs. S5A and S5B).
To verify that the synaptotagmin aggregate accumulation phenotype is not due to a non-specific effect of expressing human APP, we also monitored the effect of Nebula on modulating endogenous fly Appl gene function. Fig. 2C shows that upregulation of APPL in neurons also caused synaptotagmin accumulation in axons. Nebula co-upregulation significantly reduced the number of synaptotagmin aggregates, whereas Nebula reduction using RNAi significantly exacerbated the phenotype (Figs. 2C and 2D). Together, our results support earlier finding that mammalian APP and Drosophila APPL are functionally conserved [40], and further indicate that APP and APPL-induced axonal transport defects are regulated by Nebula in a similar fashion.
To determine to what degree aggregate accumulation corresponded to altered delivery of synaptic proteins to the synaptic terminal, we evaluated the levels of both synaptotagmin and APP in the NMJ. As demonstrated in Figs. 3A–3B, APP upregulation significantly reduced the level of average synaptotagmin intensity in the synapse while nebula co-overexpression enhanced the delivery of both synaptotagmin and APP to the synaptic terminal. This change is not due to altered overall synaptotagmin or APP levels (Figs. S4 and S5C). Note that the 4G8 antibody does not detect endogenous fly APPL; therefore, we normalized the level of APP delivered to the synapse to flies overexpressing APP and nebula. We found Nebula reduction did not further reduce the amount of synaptotagmin reaching the terminal (Fig. 3B), albeit it did increase the number of APP-induced aggregates in the axon (Fig. 2B). This result indicates that either retrograde transport of synaptotagmin is altered, or the increase in aggregate number has not yet reached a critical threshold for further impairment. In addition, although no detectable synaptotagmin aggregate was seen in flies with Nebula reduction alone, a decrease in synaptotagmin staining was detected in the synapse (Figs. S5B and S5D). This result suggests that Nebula itself may be required for reliable axonal transport.
We also examined the effect of abnormal aggregate accumulations and reduced delivery of synaptic proteins on locomotor behavior. Overexpression of APP dramatically impaired larval movement (Fig. 3C and Movie S1). Nebula co-overexpression significantly rescued this locomotor defect, in further support of the hypothesis that Nebula upregulation exerts beneficial effects on synaptic functions by alleviating abnormal aggregate accumulations. Note that further reduction of Nebula in APP overexpression background did not significantly worsen the locomotor defect of APP overexpressing larvae, perhaps due to a threshold effect. Reducing Nebula alone was sufficient to induce a mild defect in locomotor activity (Fig. S5E), suggesting delivery of synaptic proteins to the synaptic terminals is crucial for normal synaptic function.
Similar to APP overexpression, upregulation of APPL decreased the delivery of synaptotagmin to the synapse. APPL and Nebula co-upregulation showed a higher level of synaptotagmin in the NMJ, confirming Nebula interacts genetically with APPL to rescue impaired in transport (Fig. S6A and S6B). We also found that similar to RNAi-nla larvae, Appl null mutant (Appld) displayed a slight decrease in the level of synaptotagmin at the synapse independent of aggregate accumulation (Figs. S6B and S6C). Reducing Nebula in neurons of Appld larvae with the RNAi-nla transgene driven by the pan-neuronal nSyb-GAL4 driver (Appld; RANi-nla/nSyb-GAL4) did not further enhance the phenotype, suggesting that the two proteins act in the same pathway to modulate axonal transport.
While monitoring synaptotagmin levels at the NMJ, we also noticed that APP overexpression triggered changes in synaptic morphology as previously reported [41], [42]. Fig. 4 shows presynaptic terminals stained with HRP to outline the presynaptic terminals, which revealed an increase in the total number of boutons and satellite boutons brought upon by APP overexpression. Nebula co-upregulation also rescued APP-induced synapse proliferation phenotype, but not the number of satellite boutons (Fig. 4B and 4D). Manipulating levels of Nebula alone without APP did not influence bouton number or morphology, suggesting that the satellite bouton phenotype is dependent on the presence APP in the synapse. Since reducing Nebula levels alone decreased the delivery of synaptotagmin to the synaptic terminal without altering synaptic morphology, axonal transport problems are not secondary consequences of altered synaptic morphology. A plausible mechanism by which Nebula suppresses the APP-induced over-proliferation phenotype is that Nebula co-upregulation restores the delivery of proteins required for normal synaptic growth such as Fasciclin II (FasII), a cell adhesion molecule shown to influence synaptic morphology [43], [44]. Previous reports suggest that changes in FasII levels differentially affect synaptic growth [42]–[44], and that increasing FasII levels presynaptically can significantly suppress the increase in bouton number observed in APPL overexpression synapses [42]. We therefore quantified FasII levels in the NMJ (Fig. S7). We found that overexpression of APP reduced the level of FasII in the NMJ, whereas APP and nebula co-overexpression restored it (Fig. S7). While APP upregulation may play other roles in synapse formation, these results together with previous reports imply that depletion of FasII in the presynaptic terminal could partially contribute to the hyper-growth phenotype. Furthermore, our data reveal that Nebula upregulation is effective in protecting against multiple phenotypes caused by APP overexpression, including age dependent photoreceptor neurodegeneration, vesicle accumulations in axons, and changes in synaptic morphology.
To directly evaluate the effect of Nebula on APP transport and to determine whether the observed axonal aggregates correspond to defective axonal transport, we performed live-imaging of human APP tagged with yellow fluorescent protein (APP-YFP). APP-YFP vesicles in larval motor axons displayed movement in both the anterograde and retrograde directions over the 2-minute imaging period as represented by kymographs depicting distance traveled and time in the x- and y-directions, respectively (Fig. 5A). Nebula co-overexpression had a mild, but significant, effect on APP-YFP movement. Nebula co-upregulation increased the percentage of anterograde moving vesicles and resulted in reduced number of stationary APP-YFP; knockdown of Nebula using RNAi increased the number of stationary APP-YFP (Figs. 5A and 5B). Quantification of the average speed of APP-YFP movement revealed that overexpression of nebula also increased the speed of APP-YFP movement in both the anterograde and retrograde directions (Fig. 5C). Together, these results suggest that Nebula upregulation enhances the transport of APP, consistent with the decreased aggregate accumulations of APP in axons and increased APP staining in the NMJ when Nebula is co-expressed (Figs. 2A and 3A).
To further confirm that Nebula facilitates synaptic vesicle movement in the presence of APP and to better assess the role of endogenous Nebula in regulating transport, we also monitored synaptotagmin movement in the motor axons of larvae expressing GFP-tagged synaptotagmin (GFP-SYT). We find the movement of GFP-SYT to be highly dynamic with anterograde, retrograde, and bi-directional movement (Fig. 6A). Overexpression of APP dramatically reduced the percentage of vesicles moving in both the anterograde and retrograde directions while nebula co-overexpression significantly facilitated synaptotagmin transport in both directions (Figs. 6A and 6B), albeit retrograde transport was more effectively restored by Nebula. Reducing Nebula using RNAi further diminished APP-induced synaptotagmin transport in both directions, confirming interaction between Nebula and APP. Reduction in the overall movement was also accompanied by a decrease in anterograde and retrograde velocity (Fig. 6C). Together, these results suggest that APP overexpression slows down the overall movement of vesicles, which may lead to accumulation of transported proteins. Nebula co-overexpression with APP partially restores the defect by increasing the movement and speed of transport in both the anterograde and retrograde directions.
To understand the role of endogenous Nebula in axonal transport, we examined the effect of Nebula manipulations on GFP-SYT movement in the absence of APP overexpression. We find that Nebula upregulation alone did not significantly influence transport; decreasing Nebula through RNAi was sufficient to reduce the number of moving synaptotagmin vesicle in both directions, as well as the speed of anterograde transport (Fig. 6). This result is consistent with the decrease in synaptotagmin staining in the NMJ seen in static images, and further confirms that Nebula is required for efficient transport of synaptic proteins.
To further determine if general axonal transport is affected by APP and Nebula upregulation, we also monitored mitochondrial transport. Proper distribution of mitochondria is vital for normal cell functions and defects in mitochondrial transport can adversely affect cell survival [45]–[47]. Time-lapse live imaging was performed in larvae with GFP targeted to mitochondria (mito-GFP) for the indicated genotypes (Fig. 7). APP upregulation severely impaired the movement of mitochondria in both the anterograde and retrograde directions both in terms of percent in motion and the speed of movement (Figs. 7B and 7C). Nevertheless, the APP-induced mitochondrial transport defect was partially restored by Nebula co-upregulation (Fig. 7 and Movie S2), similar to what was observed for synaptic vesicle transport. Manipulations in the level of Nebula did not significantly alter the overall mitochondrial movement, except that nebula overexpression alone seemed to enhance both the proportion and the speed of mitochondria transported in the retrograde direction. This result is consistent with our observation that Nebula co-upregulation was more effective in restoring retrograde GFP-SYT transport. Together, our results suggest that Nebula influences general axonal transport that extends beyond synaptic proteins.
Mitochondria are dynamic organelles whose distribution is tightly regulated to meet the energy demands within the polarized neuron [45], [48]. We find that despite the decrease in mitochondrial movement in flies overexpressing APP, the distribution and density of mitochondria within the proximal axon where imaging was performed did not vary across genotypes (Fig. S8A). These results imply that impaired synaptic vesicle transport is not likely caused by local depletion of mitochondria within the axon. Furthermore, mitochondria did not accumulate near the site of synaptotagmin aggregate formation in the axons (Fig. S8B), suggesting that mitochondria are either able to move past the stalled synaptic vesicle accumulations or that mitochondria travel on other non-blocked microtubule tracks.
Despite increasing evidence linking defective trafficking of presynaptic proteins, mitochondria, and signaling molecules to neuropathologies of AD, mechanisms for how APP overexpression affects axonal transport remain unclear. We first tested the possibility that APP upregulation impairs axonal transport by influencing overall microtubule integrity. To this end, we stained the axonal nerves and NMJs with antibodies against acetylated tubulin, β-tubulin, and Futsch (Fig. 8). Acetylated tubulin is a marker for stable microtubules [49]; Futsch is a microtubule binding protein homolog to human MAP1B and is involved in maintaining microtubule integrity at presynaptic terminals during NMJ growth [50]. Our data revealed that APP overexpression did not cause fragmentation of microtubules as revealed by both acetylated tubulin and β-tubulin staining in the axons (Fig. 8A), and filamentous acetylated tubulin staining in the synaptic terminals across all genotypes (Fig. 8B). Note that in Fig. 8B, we also highlighted the presynaptic boutons by HRP staining (red), since acetylated tubulin in the muscles are also detected in the background. Western blot analyses of dissected larval brains further confirmed that the overall level of acetylated tubulin is not altered by APP overexpression (Fig. 8C). Closer examination of Futsch staining also did not reveal differences in overall microtubule integrity (Fig. 8D). Together, these results suggest that APP overexpression does not cause axonal transport problems by influencing microtubule stability, which is consistent with a recent report that showed normal microtubule stability and acetylated tubulin level in larvae overexpressing APP-YFP [51].
Nebula encodes an inhibitor of calcineurin that is highly conserved across species [15], we therefore tested the hypothesis that calcineurin inhibition is an underlying mechanism for Nebula-mediated rescue of APP phenotypes. To this end, we genetically altered calcineurin activity in neurons using the UAS/GAL4 strategy. To elevate calcineurin activity, we expressed a constitutively active calcineurin (CaNAct) with its auto-inhibitory domain deleted (Figs. S9A and S9B). To reduce calcineurin activity, RNAi strategy against the calcineurin B gene (RNAi-CaNB), an obligatory subunit necessary for calcineurin activity, was used. We find that similar to Nebula upregulation, decreasing calcineurin using RNAi-CaNB in the presence of APP significantly reduced synaptotagmin aggregate accumulations and synaptic depletion, as well as restored larval locomotor behavior (Figs. S9C–E). Overexpression of CaNAct together with APP further exacerbated the APP-induced phenotypes (Figs. 9A and 9B), whereas co-overexpression of CaNAct and nebula diminished the ability of Nebula to protect against APP-induced transport defects. Similar to larvae with reduced levels of Nebula (RNAi-nla), larvae expressing CaNAct did not show aggregate accumulations in axons but displayed a reduced level of synaptotagmin staining in the synapse (Figs. S9D), indicating active calcineurin overexpression alone only has modest effect on axonal transport. As shown above, synaptotagmin aggregate accumulation in nerves and depletion in the synaptic terminals are reliable indicators of significant transport deficiencies; our results thus indicate that Nebula protects against APP-induced defects through inhibition of calcineurin. Furthermore, our data present for the first time that APP upregulation influences axonal transport through activation of calcineurin. This conclusion is further supported by direct measurement of calcineurin activity, in which we find that APP upregulation significantly elevated calcineurin activity but is further restored close to normal in flies overexpressing APP and nebula, or APP and RNAi-CaNB (Fig. 9C). Overexpression of APP, CaNAct, and nebula together showed an intermediate phenotype in both calcineurin activity and aggregate accumulations, suggesting that the severity of aggregate accumulation correlated with the level of calcineurin when APP is upregulated.
How does APP upregulation trigger calcineurin activation? Because calcineurin phosphatase activity is dependent on intracellular calcium concentration [52], we examined the possibility that APP overexpression elevates calcium levels. Using a genetically encoded fluorescent calcium sensor (Case12) previously shown to detect calcium with high sensitivity [53], [54], we compared Case12 signal across different genotypes. Fig. S10 shows that larval brain expressing Case12 displayed a significant increase in signal following application of calcimycin, a calcium ionophore, confirming that the Case12 construct can indeed detect increases in calcium. Overexpression of APP alone or overexpression of APP and nebula also caused a significant elevation in Case12 signal in the larval brain and the ventral ganglion (where the motor neuron cell bodies are located) as compared to the control (Figs. 9D and 9E). These data imply that an APP-mediated increase in calcium is triggering the increase in calcineurin activity. Furthermore, observations that co-overexpression of APP and nebula increased calcium while simultaneously restoring calcineurin activity indicate that Nebula is influencing axonal transport through calcineurin inhibition rather than acting at a step modulating calcium influx.
Mechanisms by which calcineurin regulates axonal transport are not well understood, but one potential pathway is through regulation of GSK-3β activity. Aberrant activation of GSK-3β has been associated with AD and calcineurin has been shown to activate GSK-3β through dephosphorylation of Ser9 of GSK-3β in vitro [55]–[58]. It was suggested that GSK-3β may negatively influence axonal transport by altering microtubule stability through hyperphosphorylation of tau, by inhibiting kinesin motor binding to the cargo through phosphorylation of the kinesin light chain (KLC), or by altering the kinesin motor activity [51], [59]–[61]. These previous findings led us to investigate the possibility that Nebula restores APP-dependent transport problems through calcineurin-mediated regulation of GSK-3β in vivo. The activity of GSK-3β is regulated by phosphorylation and dephosphorylation: dephosphorylation of Ser9 by a number of phosphatases including calcineurin is required to activate GSK-3β [56], [62], and phosphorylation at Tyr216 site is necessary to enhance GSK-3β activity [63], [64]. Interestingly, phosphorylation of GSK-3β at Ser9 can both inhibit GSK-3β activity and override the increase in activity even when phosphorylated at Tyr216 [65]. Because these phosphorylation sites are conserved between fly and human, we took advantage of phospho-specific antibodies to monitor GSK-3β activity. Western blot analyses using an antibody specific for phosphorylated Ser9 (pSer9) of GSK-3β revealed that APP upregulation indeed reduced the level of pSer9-GSK-3β while APP and Nebula co-upregulation partially restored the level to normal (Fig. 10A). This suggests APP upregulation leads to GSK-3β activation that is inhibited by Nebula upregulation.
To verify that GSK-3β activation is due to calcineurin activation, we reduced calcineurin activity in APP overexpressing flies using RNAi-CaNB (Fig. 10A). We find that APP and RNAi-CaNB co-overexpression in neurons, which was sufficient to restore calcineurin activity, completely prevented GSK-3β dephosphorylation at Ser9 site. This result indicates that APP-induced GSK-3β dephosphorylation at Ser9 is dependent on calcineurin activation in vivo. Note that we did not detect enhanced GSK-3β dephosphorylation when APP is expressed together with constitutively active calcineurin (CaNAct), suggesting that calcineurin may in part directly influence transport through GSK-3β-independent pathways.
Our data strongly implicate activation of calcineurin and subsequent GSK-3β induction to be a mechanism underlying APP-induced aggregate phenotype. Because activation of calcineurin alone did not result in synaptotagmin aggregate accumulation, we further hypothesized that APP upregulation also enhances GSK-3β activity through phosphorylation at Tyr216. Western blot analyses show that the level of phosphorylated GSK-3β at Tyr214 (conserved Tyr216 site in Drosophila) is indeed elevated in flies overexpressing APP or APP and nebula (Fig. 10B). Overexpression of CaNAct alone, however, failed to induce phosphorylation at Tyr214, suggesting that phosphorylation of Tyr214 is not affected by calcineurin and dependent on the presence of APP. Together, our data demonstrate that in addition to activating GSK-3β by relieving inhibition through calcineurin, APP upregulation further enhances GSK-3β activity through phosphorylation at Y214 in fly.
Active GSK-3β had been shown to phosphorylate KLC, leading to detachment of the cargo from the motor [59], [66]. Since synaptotagmin transport was severely inhibited by APP overexpression, and that synaptotagmin transport can depend on kinesin 3 [67], [68] and kinesin 1 (both KLC and kinesin 1 heavy chain) [69]–[73], we tested the possibility that APP overexpression perturbs KLC and synaptotagmin interaction via immunoprecipitation. APP overexpression indeed reduced synaptotagmin (cargo) and KLC interaction while overexpression of APP and nebula preserved this interaction (Fig. 10C). These results suggest that Nebula is likely to restore APP-induced axonal transport defects by correcting GSK-3β signaling and stabilizing cargo-motor interaction.
Having demonstrated that APP activates calcineurin signaling to regulate GSK-3β phosphorylation, we next examined if reducing GSK-3β can restore axonal transport. In the presence of APP upregulation, decreasing Shaggy (Sgg; fly homolog of GSK-3β) in flies with APP overexpression (sgg1;APP) resulted in significant suppression of the APP aggregate phenotype (Figs. 10D and 10E). This result is consistent with a recent report demonstrating mild enhancement of APP-YFP movement when GSK-3β is reduced [51]. Surprisingly, normal calcineurin activity was detected in these flies (1.00±0.16 fold of control for Sgg1;APP vs. 1.75±0.25 fold of control for APP). This result suggests the existence of feedback regulation of calcineurin activity and further implies that either a change in calcineurin activity or GSK-3β signaling could be responsible for the observed rescue. We therefore generated flies expressing APP and constitutively active calcineurin in sgg1 background (sgg1;APP/CaNAct). Note that we used the hypomorphic allele sgg1 because sgg null animals are lethal [74]. Consistent with GSK-3β being downstream of calcineurin, reducing Sgg diminished the effect of CaNAct in enhancing APP phenotype (Figs. 9B and 10E). We also expressed the constitutively active Sgg (sggS9A) together with APP, which surprisingly showed the same phenotype as APP overexpression. Calcineurin activity assay showed an unexpected decrease in calcineurin activity (0.74±0.06 fold of control) in these flies, suggesting that constitutive GSK-3β activation in the absence of calcineurin activation is sufficient to disrupt axonal transport potentially through phosphorylation of KLC. Interestingly, we find that overexpression of the constitutively active Sgg in neurons alone was sufficient to induce aggregate accumulation similar to flies with APP overexpression (Figs. 10D and 10E). Calcineurin activity assay revealed that these flies showed an increase in overall calcineurin activity (1.65±0.30 fold of control). This increase in calcineurin activity by active Sgg may be due to GSK dependent phosphorylation of Nebula, which has been shown to cause activation of calcineurin [75]. Since over-activation of calcineurin and GSK-3β pathway in the absence of APP upregulation fully replicated the aggregate accumulation phenotype, it suggests that abnormal activation of both the GSK-3β and calcineurin pathways are necessary for the severe axonal transport defect and aggregate accumulation phenotypes.
We have demonstrated a novel role for Nebula, the Drosophila ortholog of DSCR1, in ameliorating axonal transport impairments associated with the upregulation of APP. We find that Nebula upregulation significantly delayed photoreceptor neurodegeneration and dramatically decreased the axonal “traffic jam” phenotype caused by APP overexpression. Reducing Nebula independent of APP was sufficient to trigger defects in axonal transport, suggesting that Nebula is normally required for reliable delivery of synaptic cargos, likely through calcineurin dependent pathway. We demonstrate for the first time that APP overexpression causes calcineurin-dependent activation of GSK-3β kinase in vivo, thus implicating altered calcineurin signaling as a novel mechanism regulating axonal transport (Fig. S11). We find that co-upregulation of Nebula preserved the vesicular cargo to molecular motor interaction, ameliorated axonal transport defects, and protected against locomotor deficits. As impaired transport of essential organelles and synaptic vesicles caused by perturbation of APP is thought to precede synaptic failure and neurodegeneration in AD, our findings further suggest that DSCR1 upregulation may be a neuroprotective mechanism used by neurons to combat the effects of APP upregulation and delay progression of AD.
Although upregulation of APP had been shown to negatively influence axonal transport in mouse and fly models 28–31, mechanisms by which APP upregulation induces transport defects are poorly understood. Several hypotheses have been proposed, including titration of motor/adaptor by APP, impairments in mitochondrial bioenergetics, altered microtubule tracks, or aberrant activation of signaling pathways [76]. The motor/adaptor titration theory suggests that excessive APP-cargos titrates the available motors away from other organelles, thus resulting in defective transport of pre-synaptic vesicles [29]. Our finding that Nebula co-upregulation enhanced the movement and delivery of both synaptotagmin and APP to the synaptic terminal argues against this hypothesis. In addition, earlier finding suggest that Nebula upregulation alone impaired mitochondrial function and elevated ROS level [77], thus implying that Nebula is not likely to rescue APP-dependent phenotypes by selectively restoring mitochondrial bioenergetics. Furthermore, consistent with a recent report showing normal microtubule integrity in flies overexpressing either APP-YFP or activated GSK-3β [51], our data revealed normal gross microtubule structure in flies with APP overexpression. Together, these results suggest that changes in gross microtubule structure and stability is not a likely cause of APP-induced transport defects.
Instead, our results support the idea that Nebula facilitates axonal transport defects by correcting APP-mediated changes in phosphatase and kinase signaling pathways. First, we find that APP upregulation elevated intracellular calcium level and calcineurin activity, and that restoring calcineurin activity to normal suppressed the synaptotagmin aggregate accumulation in axons. The observed increase in calcium and calcineurin activity is consistent with reports of calcium dyshomeostasis and elevated calcineurin phosphatase activity found in AD brains [78]–[80], as well as reports demonstrating elevated neuronal calcium level due to APP overexpression and increased calcineurin activation in Tg2576 transgenic mice carrying the APPswe mutant allele [81], [82]. Second, APP upregulation resulted in calcineurin dependent dephosphorylation of GSK-3β at Ser9 site, a process thought to activate GSK-3β kinase [56]. APP upregulation also triggered calcineurin-independent phosphorylation at Tyr216 site, which has been shown to enhance GSK-3β activity [64], [65]. The kinase(s) that phosphorylates APP at Tyr216 is currently not well understood, it will be important to study how APP leads to Tyr216 phosphorylation in the future. Based on our results, we envision that APP overexpression ultimately leads to excessive calcineurin and GSK-3β activity, whereas nebula overexpression inhibits calcineurin to prevent activation of GSK-3β (Fig. S11). Our findings that nebula co-overexpression prevented GSK-3β activation and enhanced the transport of APP-YFP vesicles are consistent with a recent report by Weaver et al., in which they find decreasing GSK-3β in fly increased the speed of APP-YFP movement [51]. Furthermore, consistent with our result that APP upregulation triggers GSK-3β enhancement and severe axonal transport defect, Weaver et al. did not detect changes in GFP-synaptotagmin movement in the absence of APP upregulation.
Active GSK-3β has been shown to influence the transport of mitochondria and synaptic proteins including APP, although the exact mechanism may differ between different cargos and motors [51], [83], [84]. One mechanism proposed for GSK-3β-mediated regulation of axonal transport is through phosphorylation of KLC1, thereby disrupting axonal transport by decreasing the association of the anterograde molecular motor with its cargos [59]. Accordingly, we find that APP reduced KLC-synaptotagmin interaction while Nebula upregulation preserved it. Synaptotagmin transport in both the anterograde and retrograde directions were affected, consistent with previous reports showing that altering either the anterograde kinesin or retrograde dynein is sufficient affected transport in both directions [85], [86]. Our results also support work suggesting that synaptotagmin can be transported by the kinesin 1 motor complex in addition to the kinesin 3/imac motor [67]–[73]. As kinesin 1 is known to mediate the movement of both APP and mitochondria [37], [86]–[88], and that phosphorylation of KLC had been shown to inhibit mitochondrial transport [89], detachment of cargo-motor caused by GSK-3β mediated phosphorylation of KLC may lead to general axonal transport problems as reported here. However, GSK-3β activation may also perturb general axonal transport by influencing motor activity or binding of motors to the microtubule tract. Interestingly, increased levels of active GSK-3β and phosphorylated KLC and dynein intermediate chain (DIC), a component of the dynein retrograde complex, have been observed in the frontal complex of AD patients [90]. Genetic variability for KLC1 is thought to be a risk factor for early-onset of Alzheimer's disease [91]. There is also increasing evidence implicating GSK-3β in regulating transport by modulating kinesin activity and exacerbating neurodegeneration in AD through tau hyperphosphorylation [21], [51], [55]. It will be interesting to investigate if Nebula also modulates these processes in the future.
Although calcineurin had been shown to regulate many important cellular pathways, the link between altered calcineurin and axonal transport, especially in the context of AD, had not been established before. We show that calcineurin can regulate axonal transport through both GSK-3β independent and dependent pathways. This is supported by our observation that the severity of the aggregate phenotype was worse for flies expressing APP and active calcineurin than it was for flies expressing APP and active GSK-3β. These findings point to a role for calcineurin in influencing axonal transport directly, perhaps through dephosphorylation of motor or adaptor proteins. Our data also indicate that calcineurin in part modulates axonal transport through dephosphorylation of GSK-3β as discussed above; however, upregulation of APP is necessary for the induction of severe axonal transport problems, mainly by causing additional enhancement of GSK-3β signaling. GSK3 inhibition is widely discussed as a potential therapeutic intervention for AD, our results suggest that perhaps calcineurin is a more effective target for delaying degeneration by preserving axonal transport.
DSCR1 and APP are both located on chromosome 21 and upregulated in DS [4], [10]. Overexpression of DSCR1 alone had been contradictorily implicated in both conferring resistance to oxidative stress and in promoting apoptosis [8], [9], [16], [17]. Upregulation of Nebula/DSCR1 had also been shown to negatively impact learning and memory in fly and mouse models through altered calcineurin pathways [15], [92]. How could upregulation of DSCR1 be beneficial? We propose that DSCR1 upregulation in the presence of APP upregulation compensates for the altered calcineurin and GSK-3β signaling, shifting the delicate balance of kinase/phosphatase signaling pathways close to normal, therefore preserving axonal transport and delaying neurodegeneration. We also propose that axonal transport defects and synapse dysfunction caused by APP upregulation in our Drosophila model system occur prior to accumulation of amyloid plaques and severe neurodegeneration, similar to that described for a mouse model [30].
DS is characterized by the presence of AD neuropathologies early in life, but most DS individuals do not exhibit signs of dementia until decades later, indicating that there is a delayed progression of cognitive decline [2], [93]. The upregulation of DSCR1 may in fact activate compensatory cell signaling mechanisms that provide protection against APP-mediated oxidative stress, aberrant calcium, and altered calcineurin and GSK3-β activity.
Flies were cultured at 25°C on standard cornmeal, yeast, sugar, and agar medium under a 12 hour light and 12 hour dark cycle. The following fly lines were obtained from the Bloomington Drosophila Stock Center: Gmr-GAL4, UAS-APP695-N-myc (6700), sgg1/FM7a, UAS-sggS9A (Sgg constitutively active), UAS-nla-RNAi (27260), UAS-CaNB-RNAi (27307), UAS-syt.eGFP (6925), UAS-APP.YFP (32039), and UAS-mitoGFP. Elav-GAL4 stock was kindly provided by Dr. Feany (Harvard University), UAS-nlat1, and nla1 flies were reported previously [15]. UAS-ΔCaNAct construct (constitutively active calcineurin) was generated by deleting the autoinhibitory domain of the CaNA gene Pp2B-14D and subcloned into the pINDY6 vector similar to that described [94]. UAS-Case12 was generated by inserting Case12 (from Evrogen) into pINDY6 vector [53]. Transgenic flies were generated by standard germline transformation method [95].
Adult Drosophila of 0, 15, 30 and 45 days of age were collected, decapitated and had their proboscis removed. Heads were incubated in Mirsky's fixative for 30 minutes, washed with PBS, and post-fixed in 4% paraformaldehyde for 20 minutes. Fly heads were then transferred to 25% sucrose overnight at 4°C and were subsequently embedded in Tissue-Tek O.C.T Compound for cryostat sectioning (10 µm). Photoreceptor axons were immunostained with 24B10 (1∶10; Developmental Studies Hybridoma), Phosphorylated APP (1∶400; Sigma), and 4G8 (1∶500; Signet).
Flies were placed in 2 clear round bottom test tubes joined at the opening. After allowing 2 minutes for the flies to acclimate to the tubes, flies were lightly tapped and the percentage of flies that moved toward light in horizontal position within 30 seconds was counted.
Wandering 3rd instar larvae were dissected in cold calcium-free dissection buffer and fixed with 4% paraformaldehyde in PBS for 25 minutes at room temperature (RT). Samples were blocked in 5% normal goat serum in PBS+0.1% triton for 1 hour at RT and then incubated with primary antibodies overnight at 4°C. Antibodies included synaptotagmin (1∶1,000; gift from H. Bellen) and mAb 4G8 (1∶1,000; Signet), β-tubulin (1∶1000; DSHB), acetylated tubulin (1∶500, Abcam), Cy3-conjugated HRP (1∶200, Jackson ImmunoResearch). Alexa-conjugated secondary antibodies were applied at 1∶500 and samples mounted in Pro -long Gold Antifade reagent (Invitrogen).
Images of motor axons and synaptic terminals from NMJ 6/7 in segment A2 or A3 were captured in a z-series using Zeiss LSM5 scanning confocal. The number of aggregates was determined manually by counting the number of punctate staining with intensity above background and size greater than 0.2 µm2. For quantification of antibody staining intensities at the NMJ, dissected larvae were stained together using the same condition. Images were captured in a z-series and parameters were set to minimize saturation of pixel intensity. Intensity of Z-projected images was analyzed using ImageJ and fold changed calculated by comparing to the control.
Wandering 3rd instar larvae expressing APP-YFP or GFP-SYT in combination with other transgenes were dissected in calcium free dissection buffer: 128 mM NaCl, 1 mM EGTA, 4 mM MgCl2, 2 mM KCl, 5 mM HEPES, and 36 mM sucrose. Live imaging of GFP-SYT was done as described [96]. For imaging of mito-GFP, dissected larvae were bathed in HL-3 solution [97]. Time-lapse images were acquired at 5-s intervals using a Zeiss LSM5 confocal using minimum laser intensity to prevent photobleaching and damage to the tissues. Images were acquired for 5 minutes with a 63× lens and a zoom of 1.7. All live imaging experiments were completed within 15 minutes starting from the time of dissection in order to ensure health of the samples.
The Manual tracking Plugin in ImageJ was used to track individual vesicle and mitochondria movement. At least 10 frames (>50 s) were used to calculate the average speed of movement. Percentage of movement was determined by counting the percentage of moving vesicles over the imaging period. A vesicle is labeled as moving if it moved in three consecutive frames (over a 15-s period) over a distance of at least 0.1 µm. Direction of movement is determined by direction of net displacement of the vesicle at the start of imaging. Average speed was determined by tracking a vesicle for an uninterrupted run in either the anterograde or retrograde direction. The total distance of movement was divided by the total duration of movement in a specific direction. Student's t-test was used to determine statistical significance.
Deficits in larval locomotor behavior were assessed as described previously [98]. Briefly, larvae were washed with PBS and placed in 60 mm petri dish filled with 1% agarose. Using a moistened paint brush, 3rd instar larvae were collected and allowed to habituate for 30 seconds. The number 0.5 cm2 boxes entered was counted for a 60-s period.
Drosophila adults (1–2 days) were collected on dry ice. Heads were removed and homogenized in cold RIPA buffer. The brains of 3rd instar larvae were dissected and collected on dry ice. Equal amount of protein per genotype (10–20 µg) was run on SDS polyacrylamide gel and transferred to nitrocellulose membrane. Blocking for phosphorylated antibodies was performed using 5% BSA in PBS+0.1% tween (PBS-TW). Blocking for non-phosphorylated antibodies was done using 5% milk in PBS-TW for one hour at RT. Membranes were incubated with the following primary antibodies overnight at 4°C: N-APP (1∶5,000; Sigma), β-tubulin (1∶500; Developmental Studies Hybridoma Bank), Nebula (1∶7,000), Fasciclin II (1∶50 Developmental Studies Hybridoma), acetylated tubulin (1∶1,000, Cell Signaling), phospho-GSK3β Ser9 (1∶1000, Cell Signaling), phospho-GSK3β Tyr126 (1∶1000, Cell Signaling), and GSK3 α/β (1∶2,000, Cell Signaling). Secondary antibodies used were: anti-mouse Alexa 680 (Invitrogen), anti-rabbit Dylight 800 (Piercenet), anti-mouse coupled HRP or anti-rabbit coupled HRP. HRP signals were detected using ECL Reagents (GE Healthcare). Alexa 680 and Dylight 800 signals were detected using Odyssey Imaging system (LI-COR Biosciences). For reprobing, membranes were stripped using Reblot Plus strong antibody stripping solution (Millipore) and reprobed. NIH Image J software was used to measure signal intensity, and the fold change in specific protein level was normalized to a loading control and compared to the control flies.
Fly heads were collected over dry ice, decapitated, and homogenized in lysis buffer (10 mM Tris pH 7.5, 1 mM EDTA, 0.02% Sodium Azide). Calcineurin phosphatase activity was determined using the Ser/Threonine Phosphatase Assay Kit (Promega) following the manufacturer's protocol as done previously [15]. 5 µg of protein per genotype was used.
Flies heads were collected on dry ice by passing through molecular sieves and homogenized in lysis buffer (10 mM HEPES, 0.1 M NaCl, 1% NP-40, 2 mM EDTA, 50 mM NaF, 1 mM NA3VO4) plus Complete Mini protease inhibitor cocktail (Roche). Lysates were pre-cleared by incubating fly extract with magnetic A/G beads (Thermo Scientific) for 1 hour at 4°C. Pre-cleared extract was then used for IP using GFP antibody conjugated to magnetic beads (MBL International). Western blot analysis using an antibody against the kinesin light chain (1∶200; Novus Biologicals) was used to confirm interaction. To determine the efficiency of GFP pull down, an antibody against GFP (1∶1000, Abnova) was also used. To eliminate signal contamination from IgG, we used HRP conjugated TrueBlot anti-rabbit IgG (1∶1000, ebioscience) that is specific for native IgG as secondary antibody.
|
10.1371/journal.pcbi.1000818 | Fast- or Slow-inactivated State Preference of Na+ Channel Inhibitors: A Simulation and Experimental Study | Sodium channels are one of the most intensively studied drug targets. Sodium channel inhibitors (e.g., local anesthetics, anticonvulsants, antiarrhythmics and analgesics) exert their effect by stabilizing an inactivated conformation of the channels. Besides the fast-inactivated conformation, sodium channels have several distinct slow-inactivated conformational states. Stabilization of a slow-inactivated state has been proposed to be advantageous for certain therapeutic applications. Special voltage protocols are used to evoke slow inactivation of sodium channels. It is assumed that efficacy of a drug in these protocols indicates slow-inactivated state preference. We tested this assumption in simulations using four prototypical drug inhibitory mechanisms (fast or slow-inactivated state preference, with either fast or slow binding kinetics) and a kinetic model for sodium channels. Unexpectedly, we found that efficacy in these protocols (e.g., a shift of the “steady-state slow inactivation curve”), was not a reliable indicator of slow-inactivated state preference. Slowly associating fast-inactivated state-preferring drugs were indistinguishable from slow-inactivated state-preferring drugs. On the other hand, fast- and slow-inactivated state-preferring drugs tended to preferentially affect onset and recovery, respectively. The robustness of these observations was verified: i) by performing a Monte Carlo study on the effects of randomly modifying model parameters, ii) by testing the same drugs in a fundamentally different model and iii) by an analysis of the effect of systematically changing drug-specific parameters. In patch clamp electrophysiology experiments we tested five sodium channel inhibitor drugs on native sodium channels of cultured hippocampal neurons. For lidocaine, phenytoin and carbamazepine our data indicate a preference for the fast-inactivated state, while the results for fluoxetine and desipramine are inconclusive. We suggest that conclusions based on voltage protocols that are used to detect slow-inactivated state preference are unreliable and should be re-evaluated.
| Sodium channels are the key proteins for action potential firing in most excitable cells. Inhibitor drugs prevent excitation (local anesthetics), regulate excitability (antiarrhythmics), or prevent overexcitation (antiepileptic, antispastic and neuroprotective drugs) by binding to the channel and keeping it in one of the inactivated channel conformations. Sodium channels have one fast- and several slow-inactivated conformations (states). The specific stabilization of slow-inactivated states have been proposed to be advantageous in certain therapeutic applications. The question of whether individual drugs stabilize the fast or the slow-inactivated state is studied using specific voltage protocols. We tested the reliability of conclusions based on these protocols in simulation experiments using a model of sodium channels, and we found that fast- and slow-inactivated state-stabilizing drugs could not be differentiated. We suggested a method by which the state preference of at least a subset of individual drugs could be determined and tried the method in electrophysiology experiments with five individual drugs. Three of the drugs (lidocaine, phenytoin and carbamazepine) were classified as fast-inactivated state-stabilizers, while the state preference of fluoxetine and desipramine was found to be undeterminable by this method.
| Sodium channels are the key proteins in action potential firing for most excitable cells. They exhibit a complex, membrane potential-dependent gating behavior [1]. Even minor disturbances in the gating behavior can lead to hyperexcitability, which can be one of the causes of various disorders such as epilepsy, migraine, neuropathic and inflammatory pain, muscle spasms, and chronic neurodegenerative diseases. For several decades, sodium channel inhibitors (SCIs) have been successfully used to lower excitability as, for example, local anesthetics, anticonvulsants, antiarrhythmics, analgesics, antispastics and neuroprotective agents. Interestingly, the majority of antidepressants were also found to be potent SCIs. In a recent study [2] the highest incidence of SCI activity was found amongst this therapeutic class. We intend to test if the mechanism of action on sodium channels is similar to that of classic SCIs.
Thus far only a single drug binding site is established unequivocally on sodium channels, the “local anesthetic receptor”, located within the inner vestibule, its key residue being the phenylalanine located right below the selectivity filter, on domain 4 segment 6 [3]. However, the contribution of individual residues within the inner vestibule changes from drug to drug [4]–[6]. For certain drugs an alternative binding site have been proposed, which is supposed to be located within the outer pore [7], [8], but the exact position of the binding site(s) for specific SCIs (other than local anesthetics) is currently unsettled. For our case the exact location of the binding site is not relevant, we only need to suppose that the major mechanism of inhibition is preferential affinity to-, and stabilization of a specific inactivated state.
The major mechanism of SCIs is stabilization of an inactivated channel conformational state as a result of a preferential affinity for that state. The question of which inactivated state is preferred is under debate for many SCI drugs (e.g. [9]–[12], or [13]–[17]). Sodium channels are capable of fast inactivation (complete within a few milliseconds), and different forms of slow inactivation (time constants ranging from ∼100 ms to several minutes) [18]. Slow-inactivated state preference has been proposed as a therapeutic advantage [19]–[21]. Mutations of sodium channel genes which affect slow inactivation are associated with several diseases [22]. Slow inactivation determines sodium channel availability, and thereby contributes to overall membrane excitability, determining the propensity to generate repetitive firing, and the extent of action potential backpropagation. Slow inactivated state preference has been proposed as a potential therapeutic advantage in specific types of epilepsy, neuropathic pain and certain arrhythmias [19]–[22]. Furthermore, this mechanism of sodium channel inhibition has been proposed to modulate neuronal plasticity [22]. In recent years a number of novel slow-inactivated state-preferring drug candidates have been described, including the recently approved antiepileptic drug lacosamide (Vimpat) [19], [23]. This drug has been found to be effective in a model of treatment-resistant seizures, and of diabetic neuropathic pain, in which tests conventional anticonvulsants were found ineffective [19].
Special voltage protocols are used to evoke and study the slow-inactivated state. Availability of channels is studied after a prolonged depolarization (to induce slow inactivation), followed by a hyperpolarizing gap (to allow recovery from fast, but not slow inactivation). Because availability in such protocols is solely determined by the extent of slow inactivation, a drug that decreases availability is considered to be slow-inactivated state-preferring. However, gating rates (the rate of inactivation and rate of recovery from inactivation) are altered by drug binding. A fast-inactivated state-preferring drug stabilizes this state by delaying recovery. A delayed recovery does not necessarily indicate actual modification of the gating rate. For example if the bound drug prevents recovery from inactivation, then recovery will appear to be slowed because the drug needs first to dissociate [24], [25]. In our current study, however, we chose to use a model according to the modulated receptor hypothesis [26], [27], i.e., the change in affinity equals the actual modification of the gating rates. For this reason in our model increased affinity is synonymous with state stabilization. Altered gating rates have been experimentally demonstrated using gating charge measurements [28], [29]. Because of the altered gating, the rate of recovery from fast inactivation in the presence of the drug can easily overlap with the rate of recovery from slow-inactivated state. The rate of state-dependent association and dissociation of the drug should also be taken into account. As a result, interpretation of data obtained with these protocols is not straightforward (e.g. [9], [30]).
With the help of simulations, we intended to understand the interactions between binding and gating rates and wanted to test the major prototypical inhibitor mechanisms in commonly used protocols. We wanted to explore what could be deduced from these data, and wanted to find the right protocols that could help to determine the inhibition mechanisms.
Our data suggest that conclusions based on conventional protocols are not reliable. For example, the fact that one drug preferentially shifts the “steady-state slow inactivation curve” as compared to another drug does not necessarily mean that the drug prefers the slow-inactivated state. Figure 1 illustrates two (simulated) drugs investigated in “steady-state inactivation” protocols (protocols are discussed below). Both drugs shifted the “fast inactivation curve” (Figure 2, “FInact_V”) to the same degree, but Drug 1 caused a larger shift in the “slow inactivation curve” (Figure 2, “SInact_V”). In this special case, however, Drug 1 was defined to have a higher affinity to fast inactivated state, while Drug 2 had a higher affinity to slow inactivated state. We observed, on the other hand, that fast- and slow-inactivated state-preferring drugs tended to preferentially affect the onset of inactivation and recovery, respectively. Therefore we combined the information from these two protocols by plotting effectiveness in one protocol as a function of effectiveness in the other. We observed that data points for fast- and slow-inactivated state-preferring drugs were confined to definite areas of the effectiveness (inactivation) – effectiveness (recovery) plane. The two areas were found to overlap; therefore, explicit determination of the mechanism was not possible in all cases.
Using patch-clamp experiments, we tested three classic SCIs (lidocaine, phenytoin and carbamazepine) and two antidepressants (fluoxetine and desipramine). Properties of inhibition by classic SCIs were consistent with fast-inactivated state preference with fast binding kinetics. Inhibition by antidepressants was distinctly different. Whether the difference was caused by slow binding kinetics or slow-inactivated state preference could not be determined.
For simulations two different kinds of models were used: a phenomenological Hodgkin-Huxley type model and a state model similar to the one published by Kuo and Bean [31]. In both models, however, we introduced slow-inactivated states and drug-bound states with altered gating transition rates. For a detailed description of the models see Methods and Text S1. The Hodgkin-Huxley type model, which will be referred to as the “tetracube” model because of its topology (see Methods), was used for most simulations. The Kuo-Bean type model, referred to as the “multi-step-activation” (MSA) model, was only used for testing the robustness of our observations. In the models, both the degree of alteration of the transition rates and the state preference (the difference between affinities for different states) were given by a single factor CF (for fast-inactivated state-preferring drugs) or CS (for slow-inactivated state-preferring drugs). The kinetics of association and dissociation to the resting state are defined by the rate constants ka and kd, respectively. Association and dissociation rate constants to other states were calculated as described in Methods.
To compare simulated data with experimental results, we used similar voltage protocols in both the simulations and experiments (Figure 2). Throughout this study we used four protocols:
“FInact_V” is a standard “steady-state fast inactivation” protocol in which availability is assessed as a function of pre-pulse membrane potential. The pre-pulse duration was 0.1 s when we compared the effects of “FInact_V” and “SInact_V”. In electrophysiology experiments, because drugs with differing mechanisms of action and association kinetics had to be compared, a 2 s pre-pulse duration was used. Note, that although we use the widespread term “steady-state fast inactivation” protocol, the term is incorrect for two reasons. First, it is not necessarily “steady-state” in the sense that the pre-pulse duration may not be long enough for reaching equilibrium of either drug binding or channel gating (2 s is enough for the development of some degree of slow inactivation). Second, “availability” would be a better term than “inactivation”, because the protocol does not necessarily reflect only inactivation in the presence of a drug, since we cannot separate blocked open and inactivated channels; however, “fast availability” and “slow availability” protocols are improper terms.
“SInact_V” is a “steady-state slow inactivation” protocol in which occupancy of the slow-inactivated state is intended to be measured as a function of the membrane potential. It differs from the previous protocol in two respects: pre-pulse duration is longer (10 s), allowing more complete development of slow inactivation; and this protocol contains a 10 ms hyperpolarizing (−150 mV) gap between the pre-pulse and the test pulse. The hyperpolarizing gap serves to separate occupancy of the slow-inactivated state from that of fast-inactivated states: >95% of channels recover from the fast-inactivated state within this period. Despite the name, however, neither “FInact_V” nor “SInact_V” is able to measure drug effects on a pure population of fast or slow-inactivated channels. Fast inactivation practically reaches equilibrium at most membrane potentials within ∼10 ms. With longer durations of pre-pulses in the “FInact_V” protocols the ratio of slow-inactivated channels increases from ∼5% (0.1 s pre-pulse) to ∼40% (2 s pre-pulse). This is accompanied by a minor shift of the curve (ΔV1/2<4 mV). Drug effects can further change this distribution depending on binding kinetics and state preference. In “SInact_V” protocols most unavailable channels are in a slow-inactivated state in the absence of drugs. However, the presence of a drug may alter the distribution of channel states. The unavailable fraction does not consist of slow-inactivated channels only but also is “contaminated” with drug-bound fast-inactivated channels. The conventional name “steady-state” therefore is absolutely untrue for this protocol, as the extent and V1/2 of slow inactivation is strongly dependent on pre-pulse duration. We nevertheless need to use this terminology as we have discussed above.
“SInact_t” (“Slow inactivation onset as a function of time”) monitors the effect of prolonged depolarizations on sodium channel availability. In the absence of drugs, the onset of slow inactivation is monitored as a function of time (duration of depolarizing pulses). In the presence of a drug, it is not clear whether it reflects pure slow inactivation or a mixture of fast and slow inactivation (see below for a detailed explanation).
“Rec_t” (“Recovery from inactivation as a function of time”) monitors recovery after a 5 s depolarization to −20 mV as a function of hyperpolarizing gap duration (the gap is between the 5 s pre-pulse and the test pulse). In the absence of drugs, a 5 s depolarization causes both fast and slow inactivation (approximately 45–55%, respectively), and the protocol monitors recovery from both states. The time constants for recovery were 2.21 and 58.25 ms [32]. In the presence of drugs, measured recovery reflects the combination of dissociation and recovery from both inactivated states.
Concentration-response curves were simulated using single depolarizations to 0 mV from holding potentials of −150, −90 and −60 mV.
We plotted the nSOD values of the “Rec_t” protocol as a function of the nSOD values of the “SInact_t” protocol. (Figure 5). We investigated the effect of changing the following parameters: i) binding kinetics of drugs, ii) state preference factors (CF and CS), iii) drug concentration, iv) sodium channel model parameters, and v) hyperpolarizing gap duration in the “SInact_t” protocol.
Binding kinetics: We simulated 10 different pairs of rate constants spanning five orders of magnitude from 5*10−4 to 15 µM−1s−1 (ka) and from 0.1 to 3000 s−1 (kd). The ratio of ka and kd was kept constant ka/kd = 5*10−3, ensuring that the affinity of the drug toward the resting channel remained constant.
State preference factors: CF and CS were given the following values: 2, 5, 10, 20 or 50. Using the five CF and the five CS values, each with the ten pairs of ka and kd values, we simulated altogether 100 “drugs” in both “SInact_t” and “Rec_t” protocols. To correct for different potencies, the concentration of each drug was scaled: we used the concentration that caused 50% inhibition of single depolarizations at −90 mV holding potential (Table S5). Figure 5A shows the distribution of “Rec_t” nSOD vs.“SInact_t” nSOD values. As the binding kinetics were accelerated, data points for specific CF/CS values proceeded clockwise along a closed loop. The explanation is that binding kinetics have a range of optimal effectiveness; kinetics that are too slow do not allow for sufficient association during depolarizations, while kinetics that are too fast cause drug molecules to dissociate more during hyperpolarizations. Around the optimum conditions, effectiveness in the “SInact_t” protocol increases with an acceleration in the kinetics in parallel with a decrease of effectiveness in the “Rec_t” protocol.
When CF and CS values were changed without concentration correction, the absolute value of the change was proportional to the value of CF and CS, but the characteristic clockwise loop pattern was unchanged (Figure 5B).
Drug concentration: The effect of changing concentrations while keeping CF or CS constant (CF = 10 or CS = 10) is shown in Figure 5C. The concentration was decreased and increased tenfold. The effect increased with increasing concentration, while acceleration of the binding kinetics caused the points to move along the clockwise loop as described above.
When all simulation results were plotted on the nSOD(Rec_t) – nSOD(SInact_t) plane, we observed that fast- and slow-inactivated state-stabilizing drugs were confined to limited but overlapping areas of the plane (Figure 5D). Because of the clockwise progression of the points upon acceleration of the binding kinetics, the overlapping area contains mostly “FI_sb” and “SI_fb” type drugs.
Sodium channel model parameters: To test the influence of channel parameters, we plotted the results from Monte Carlo simulations of the four prototypical drugs on the nSOD(Rec_t) – nSOD(SInact_t) plane, and compared those with the areas based on Figure 5D. “FI” drugs were almost exclusively located within the “fast area,” while “SI” drugs were located within the “slow area,” practically irrespective of model parameters. The overlapping area was populated mostly by “FI_sb” and SI_fb” drugs, confirming the reliability of “fast” and “slow” areas (Figure 5E).
Hyperpolarizing gap duration of the “SInact_t” protocol: In simulations and experiments, we used a 10 ms gap duration, which is enough for a >90% recovery from the fast-inactivated state under control conditions. In the presence of a fast-inactivated state-stabilizing drug, recovery is slowed down. For this reason, in experiments where slow-inactivated state-stabilizing drugs are to be identified, gap duration is often chosen to be of a longer duration (up to 1 s) to ensure that the recovery from fast inactivation is complete. Our simulations indicated that “FI_sb” and “SI” type drugs nevertheless overlap in behavior no matter what hyperpolarizing gap duration is chosen (see Figure 3E). We tested the effect of setting the gap duration to 1 s (Figure 5F). “FI” and “SI” type drugs were no better separated with a 1 s than with the 10 ms gap duration.
In summary, localization on the nSOD(Rec_t) – nSOD(SInact_t) plane can reveal the state preference of a drug if it falls on one of the non-overlapping areas. However, many “SI_fb” and “FI_sb” type drugs are expected to fall in the overlapping section and, therefore, their state preference cannot be determined.
The following SCI drugs were used: the local anesthetic and antiarrhythmic lidocaine (300 µM), the anticonvulsants phenytoin (300 µM) and carbamazepine (300 µM), and the antidepressants fluoxetine (30 µM) and desipramine (30 µM). The concentrations were chosen to be similarly effective in causing a hyperpolarizing shift (−10 to −18 mV) of the “steady-state inactivation” curve (“FInact_V” – 2 s pre-pulse) (Figure 6A).
In the “SInact_t” protocol (Figure 6B), carbamazepine and phenytoin caused only a small acceleration in the process of inactivation. Fluoxetine and desipramine caused an obvious shift, similar to the one caused by the prototypical drugs “FI_sb,” “SI_fb” and “SI_sb.” Lidocaine strongly shifted the curve (especially in the early phase), which is typical of “FI_fb” type drugs. The reason for the small effect of carbamazepine was its fast dissociation kinetics. When the hyperpolarizing gap duration was changed from 10 ms at −150 mV to 5 ms at −120 mV (similar to the protocol used by Kuo et al. [12]), carbamazepine became as effective as lidocaine (Figure 6B inset).
In the “Rec_t” protocol (Figure 6C), fluoxetine and desipramine shifted the curve of recovery, similar to the prototypical drugs “FI_sb,” “SI_sb” and “SI_fb.” Carbamazepine, phenytoin and lidocaine only altered the early phase of the recovery curve, similar to the drug “FI_fb.”
We created the nSOD(Rec_t) – nSOD(SInact_t) plots for all five drugs (Figure 6D). The data points for fluoxetine and desipramine were in the overlapping area. The data points for carbamazepine, phenytoin and lidocaine fell into the non-overlapping area of fast inactivation stabilizing drugs.
Slow-inactivated state preference has been proposed to be a therapeutic advantage [19]–[21], and therefore different drugs have been tested for this property. The question of fast- or slow-inactivated state preference is a complex problem because of the interdependence of binding and gating equilibria. Multiple interconnected equilibria can be relatively easily handled by modeling; therefore, we used this approach to test hypotheses regarding state preference. Our current simulation data suggest that conclusions based on conventional protocols [19]–[21], [33]–[35] are not reliable.
A shift of the “steady-state slow inactivation curve” (“SInact_V” protocol), a shift of the “slow inactivation onset” curve (“SInact_t” protocol) and a shift of the recovery curve (“Rec_t” protocol) could all be caused by both fast- or slow-inactivated state stabilization. This conclusion was confirmed both by testing whether our observations were true for the entire parameter space and by applying a different type of model. We found that, with all combinations of parameters (within the reasonable range), our observations held true. Furthermore, both the phenomenological tetracube model and the MSA state model gave qualitatively similar results.
Nevertheless, the four prototypical mechanisms behaved appreciably differently. For this reason, we investigated the extent to which the two major mechanisms (“FI” and “SI”) could be distinguished using the combined information from different voltage protocols. Based on the nSOD(Rec_t) – nSOD(SInact_t) plots, we concluded that “FI” type drugs can be recognized, provided that their binding kinetics are fast enough. However, “FI” drugs with slower binding kinetics will overlap with “SI” drugs. Determination of the state preference would only be possible if we could measure the binding kinetics of individual drugs. However, distinguishing slow association from association to a slow-inactivated state is not trivial. In order to separate gating kinetics from binding kinetics, a rapid pulse application of the drug is necessary [32], [36]. Even in this case, association and dissociation rates cannot be correctly determined because the drug binding site on sodium channels is not extracellularly localized. Therefore, the onset rate of a drug effect may be determined by multiple processes: aqueous phase – membrane partitioning, outer to inner leaflet translocation, intramembrane diffusion and association, itself. Any one of these may be the rate limiting step, which obscures the microscopic association rate.
We investigated three well-known SCI drugs (lidocaine, phenytoin and carbamazepine) and two antidepressants (fluoxetine and desipramine). The uniquely high incidence of SCI activity among antidepressants [2], as well as their high affinity to sodium channels as compared to classic SCIs, suggests that the inhibition of sodium channels may contribute to their therapeutic effect. The therapeutic profile of antidepressants is different from that of classic SCIs (anticonvulsants, local anesthetics, antiarrhythmics), and we also intended to study whether the mechanism of inhibition was similar to that of classic SCIs.
The experimental behavior of the five drugs was remarkably similar to the behavior of prototypical drugs in simulations. We suggest that lidocaine, phenytoin and carbamazepine stabilize the fast-inactivated state, and that they have fast binding kinetics. Their nSOD(Rec_t) – nSOD(SInact_t) plot clearly fell into the “fast area.” Furthermore, their effect on the “Rec_t” curve was similar to the effect of “FI_fb.” Lidocaine behaved similarly to “FI_fb” in the “SInact_t” protocol as well. We hypothesized that the moderate effect of phenytoin and carbamazepine was due to their extra fast dissociation kinetics. This hypothesis was verified in the case of carbamazepine, which produced the characteristic “FI_fb” type effect on “SInact_t” curves upon minor modifications to the protocol.
The nSOD(Rec_t) – nSOD(SInact_t) plots of fluoxetine and desipramine fell into the overlapping area. Thus, their state preference could not be unambiguously determined. However, their properties of inhibition definitely differed from those of classic SCIs.
Patch clamp electrophysiology was done on native sodium channels in cultured hippocampal neurons. Cell culture preparation and electrophysiology were performed as published previously [32]. Cultured hippocampal neurons (prepared on the 17th day after gestation) were found to express mostly the Nav1.2 and Nav1.6 isoform, but Nav1.1, Nav1.3 and Nav1.7 isoforms were also detected in a some cells [37]. In spite of the differences in expression pattern biophysical properties of sodium currents were remarkably similar [32], [37], and potency of individual drugs showed no higher variance than in experiments using Nav1.2 expressing HEK 239 cells (data not shown). Error bars on the figures represent SEM, and the number of cells tested (n) was between 4 to 10.
All experimental procedures were approved by the Animal Care and Experimentation Committee of the Institute of Experimental Medicine, and as stated by the decision of the Animal Health and Food Control Department of the Ministry of Agriculture and Regional Development, were in accordance with 86/609/EEC/2 Directives of European Community.
The simulation was based on a set of differential equations with the occupancy of each state (i.e., the fraction of the ion channel population in that specific state) given by the following equation:(1)where Si(t) is the occupancy of a specific state at time t and Sj(t) is the occupancy of a neighboring state. Neighboring states are states where direct transitions are possible. n is the number of neighboring states, and kij and kji are the rate constants of transitions between neighboring states.
Differential equations were solved during simulations using a fourth-order Runge-Kutta method. We used either Berkeley Madonna v8.0.1 (http://www.berkeleymadonna.com/) or a program written in C++.
|
10.1371/journal.pgen.1000741 | On the Analysis of Genome-Wide Association Studies in Family-Based Designs: A Universal, Robust Analysis Approach and an Application to Four Genome-Wide Association Studies | For genome-wide association studies in family-based designs, we propose a new, universally applicable approach. The new test statistic exploits all available information about the association, while, by virtue of its design, it maintains the same robustness against population admixture as traditional family-based approaches that are based exclusively on the within-family information. The approach is suitable for the analysis of almost any trait type, e.g. binary, continuous, time-to-onset, multivariate, etc., and combinations of those. We use simulation studies to verify all theoretically derived properties of the approach, estimate its power, and compare it with other standard approaches. We illustrate the practical implications of the new analysis method by an application to a lung-function phenotype, forced expiratory volume in one second (FEV1) in 4 genome-wide association studies.
| In genome-wide association studies, the multiple testing problem and confounding due to population stratification have been intractable issues. Family-based designs have considered only the transmission of genotypes from founder to nonfounder to prevent sensitivity to the population stratification, which leads to the loss of information. Here we propose a novel analysis approach that combines mutually independent FBAT and screening statistics in a robust way. The proposed method is more powerful than any other, while it preserves the complete robustness of family-based association tests, which only achieves much smaller power level. Furthermore, the proposed method is virtually as powerful as population-based approaches/designs, even in the absence of population stratification. By nature of the proposed method, it is always robust as long as FBAT is valid, and the proposed method achieves the optimal efficiency if our linear model for screening test reasonably explains the observed data in terms of covariance structure and population admixture. We illustrate the practical relevance of the approach by an application in 4 genome-wide association studies.
| During the analysis phase of genome-wide association studies, one is confronted with numerous statistical challenges. One of them is the decision about the “right” balance between maximization of the statistical power and, at the same time, robustness against confounding. In family-based designs, the possible range of analysis options spans from a traditional family-based association analysis [1]–[4], e.g. TDT, PDT, FBAT, to the application of population-based analysis methods that have been adapted to family-data [1]–[3]. While, by definition, the first group of approaches is completely immune to population admixture and model misspecification of the phenotype, and can be applied to any phenotype that is permissible in the family-based association testing framework (FBAT [4]–[6]), the second category of approaches maximizes the statistical power by a population-based analysis. The phenotypes are modeled as a function of the genotype, and population-based methods such as genomic control [7],[8], STRUCTURE [9] and EIGENSTRAT [10], are applied to account for the effects of population admixture and stratification. Hybrid-approaches that combine elements of both population-based and family-based analysis methods, e.g. VanSteen algorithm [11] and Ionita weighting-schemes [12],[13] have been suggested to bridge between the 2 types of analysis strategies. Contrary to the other methods that combine family data and unrelated samples [14]–[17], such hybrid testing strategies maintain the 2 key features of the family-based association tests: The robustness against confounding due to population admixture and heterogeneity, and the analysis flexibility of the approach with respect to the choice of the target phenotype. Such 2-stage testing strategies utilize the information about the association at a population-level, the between-family component, to prioritize SNPs for the second step of the approach in which they are tested formally for association with a family-based test. The hybrid approaches can achieve power levels that are similar to approaches in which standard population-based methods are applied to family-data, but the optimal combination of the 2 sources of information (the between-family component and the within-family component) is not straightforward in the hybrid approaches.
In this communication, we propose a new family-based association test for genome-wide association studies that combines all sources of information about association, the between and the within-family information, into one single test statistic. The new test is robust against population-admixture even though both components, the between and the within-family components, are used to assess the evidence for association. The approach is applicable to all phenotypes or combinations of phenotypes that can be handled in the FBAT-approach, e.g. binary, continuous, time-to-onset, multivariate, etc [4]–[6],[18]. While the correct model specification for the phenotypes will increase the power of the proposed test statistic, misspecification of the phenotypic model does not affect the validity of the approach. Using extensive simulation studies, we verify the theoretically derived properties of the test statistic, assess its power and compare it with other standard approaches. An application to the Framing heart study (FHS) illustrates the value of the approach in practice. A new genetic locus for the lung-function phenotype, FEV1 (forced expiratory volume in the first second) is discovered and replicated in 3 independent, genome-wide association studies.
We assume that in a family-based association study, n family members have been genotyped at m loci with a genome-wide SNP-chip. For each marker locus, a family-based association test is constructed based on the offspring phenotype and the within-family information. The within-family information is defined as the difference between the observed, genetic marker score and the expected, genetic marker score, which is computed conditional upon both the parental genotypes/sufficient statistic [19] under the assumption of Mendelian transmissions. We denote the family-based association test for the ith marker locus by FBATi. Such an FBAT statistic can be the standard TDT, an FBAT for quantitative/qualitative traits, FBAT-GEE for multivariate traits, etc [4],[6],[18],[20],[21]. Similarly, for the ith marker, the between-family information can be used to assess the evidence for association at a population-based level by computing a VanSteen-type [11] “screening statistic” Ti. The screening statistic is computed based on the data for offspring phenotype and the parental genotypes/sufficient statistic. The statistic Ti can be a Wald test for the genetic effect size that is estimated based on the conditional mean model [22], or the estimated power of the family-based test FBATi [23], either of which is feasible. However, while the FBATi statistic is robust against population stratification, the screening statistic Ti is susceptible to confounding. For this reason, the VanSteen-type testing strategies have restrictively used the between-family information as weights for p-values of the FBAT-statistic, but never as a component in the test statistic itself.
In order to construct a family-based association test that incorporates both the within and the between-family information, the Z-statistics that correspond to the p-values of FBATi and Ti are computed. The statistic Zα* is the α quantile of standard normal distribution. pFBATi and pTi are the p-value of the FBAT-statistics and one sided p-value of the screening statistic where the direction of the one sided screening statistic is defined by the directionality of FBATi. Based on the statistical independence of FBATi and Ti [11] under the null-hypothesis, we can obtain an overall family-based association test statistic Zi by combining both Z-statistics in a weighted sum, where the parameters wFBAT and wT are standardized weights so that the overall family-based association test Zi has a normal distribution with mean 0 and variance 1, i.e. wFBAT2+wT2 = 1. In the literature, this approach of combining two test statistics is known as the Liptak-method [24]. However, the Liptak-method varies here from its standard application in that the 2 test statistics have to be combined so that confounding in the screening statistic Ti cannot affect the validity of the overall family-based association test statistic Zi. In the context of a genome-wide association study (GWAS), we are able to achieve this goal by using rank-based p-values for the screening statistic Ti instead of their asymptotic p-values.
The “screening statistics” Ti are sorted based on their evidence for association so that T(m) denotes the screening statistic with the least amount of evidence for association and T(1) the screening statistic with the largest amount of evidence for association. The rank-based p-value, (i – 0.5)/m, is then assigned to the ordered screening test statistics T(i). If there is a tie, then the average of the ranks will be used for the computation of the rank-based p-value for the ith marker. Since the null-hypothesis will be true for the vast majority of the SNPs in a GWAS, the rank-based p-values provide an alternative way to assess the significance of the population-based screening statistic Ti. The overall association test Zi is then computed based on the Z-score for the asymptotic p-value of the FBAT-statistic and the Z-score for the ranked-based p-value of the screening statistic Ti. In Text S1 we show that the overall association test Zi maintains the global significance level α, under any situations including population admixture and stratification. This can be understood intuitively as well. The smallest rank-based p-value is 0.5/m. Using the Bonferroni-correction to adjust for multiple testing, the individual, adjusted significance level is α/m which will always be smaller than the smallest rank-based p-value, 0.5/m, unless the pre-specified global significance level α is great than 0.5. This implies that the overall family-based association test can never achieve genome-wide significance just based on the rank-based p-values alone. The FBAT-statistic has to contribute evidence for the association as well in order for the overall family-based association test to reach genome-wide significance. Finally, we have to address the specification of the weights wFBAT and wT in the overall family-based association test statistic Zi. While any combination of weights wFBAT and wT will provide a valid test statistic Zi, the most powerful overall statistic Zi is approximately achieved when the ratio of the weights is equal to the ratio of the standardized effect sizes, the expected effect size of the regression coefficient divided by its (estimated) standard deviation. For quantitative traits in unascertained samples, one can show that optimal power levels are achieved for equal weights, i.e. wFBAT = wT. In general, the equal weighting scheme seems to provide good power levels for any disease mode of inheritance and for different trait types, e.g. binary traits, time-to-onset, etc. The theoretical derivation of optimal weighting schemes for such scenarios is ongoing research and will be published subsequently.
Furthermore, it is important to note that, instead of the Liptak-method, Fisher's method for combining p-values could have been used as well to construct an overall family-based association test which would have the same robustness properties as the overall-test based on the Liptak-method. However, simulation studies (data not shown) suggest that the highest power levels are consistently achieved with the Liptak method. We therefore omit the approach based on Fisher's method here.
In the first part of the simulation study, the type-1 error of the proposed family-based association test denoted as LIP was assessed in the absence and in the presence of population admixture, and we use the Wald test based on the conditional mean model [22] with between-family component for pTi in our all simulations. For various scenarios, we verified that the proposed overall family-based association test maintains the α-level.
For simplicity, we assume in the simulation studies that the random samples are given, i.e. no ascertainment, and that the parental genotypes are known. Assuming Hardy-Weinberg equilibrium, the parental genotypes are generated by drawing from Bernoulli distributions defined by the allele frequencies. The offspring genotypes are obtained by simulated Mendelian transmissions from the parents to the offspring. For the jth trio, the offspring phenotype Yj is simulated from a Normal distribution with mean aXj and variance 1, i.e. N(aXj, 1), where the parameter a represents the genetic effect size and the variable Xj denotes the offspring genotype. Under the null-hypothesis of no association, the genetic effect size parameter a will be set to 0.
For scenarios in which population admixture is present, we assume that the admixture is created by the presence of 2 subpopulations whose phenotypic means differ by 0.2. The allele frequencies for each marker in the two subpopulations are generated by the Balding-Nichols model [25]. That is, for each marker, the allele frequency in an ancestral population is generated from a uniform distribution between 0.1 and 0.9, U(0.1, 0.9). Then, the marker allele frequencies for the two subpopulations are independently sampled from the beta distributions (p(1−FST)/FST, (1−p)(1−FST)/FST) for the whole markers in each replicate of the simulated GWAS. A survey reported FST estimates with a median of 0.008 and 90th percentile of 0.028 among Europeans, and the corresponding values are 0.027 and 0.14 among Africans, and 0.043 and 0.12 among Asians [26]. The value for Wright's FST was assumed to be 0.05, 0.1, 0.2, or 0.3. Each trio was assigned to the one of the 2 subpopulations with 50% probability.
In the absence and presence of the population stratification (FST = 0.05, 0.1, 0.2, and 0.3), Table 1 shows the empirical type-1 error rates of the overall association test statistic Zi for a GWAS with 500,000 SNPs. The estimates for the empirical significance levels in Table 1 are based on 2,000 replicates. The empirical genome-wide significance level is calculated as the proportion of replicates for which the minimum p-values among the 500,000 markers is less than 0.05/500,000. We consider the proposed equal weights for wFBAT and wT, for genome-wide significance level 0.05 and Table 1 shows that the type-1 error rate is preserved well. For different significance levels, we calculate in Table 2 the empirical proportions of SNPs for which the overall family-based association test Zi is significant at the α-levels of 0.05, 0.01, 10−3, 10−4 and 10−5. The simulation studies are conducted in the absence and in the presence of population admixture. Table 2 does not provide any evidence for a departure of the empirical significance levels from the theoretical levels, both in the absence and presence of population substructure. These results confirm our theoretical conclusions that Zi is robust against population stratification and maintains correct type-1 error.
In the next set of simulation studies, we assess the effects of the local population stratification on the overall family-based association test. We generate local population stratification under the following assumptions: there are two subpopulations, G1 and G2 which distinguish themselves from each other in 2 marker regions. We assume that a subject can be from all possible 4 combinations at the 2 particular regions, e.g. (G1, G1), (G1, G2), (G2, G1) and (G2, G2). Both regions consist of 10K SNPs and 90K SNPs respectively and if subjects are from the same subpopulation in each genetic region, their assumed allele frequencies of the markers in the corresponding region are equal. For example, the allele frequencies of each marker in the marker region 1 are the same for samples in (G1, G1) and (G1, G2), but they are different for (G1, G1) and (G2, G2). In the simulation study, we generate the parental genotypes based on these allele frequency assumptions and obtain the offspring genotypes based on simulated Mendelian transmissions. Using the Balding-Nichols model we considered FST's of 0.001, 0.005, 0.01 and 0.05 in the simulation studies. The offspring's phenotype was generated under the null hypothesis, but we assumed that each sub-population strata had a different phenotypic mean: 0 for (G1, G1), 0.2 for (G1, G2), 0.4 for (G2, G1) and 0.6 for (G2, G2). Each replicate consists of 2,000 trios with equal number of trios for all 4 possible combinations. The data was analyzed with the proposed overall family-based association test and with standard linear regression after adjusting population admixture with EIGENSTRAT [10]. For EIGENSTRAT, we applied the principal component analysis to the mean of the paternal and maternal genotypes at each locus because parents of each offspring are from the same subpopulation, and then the residuals obtained from regressing offspring genotypes and phenotypes with eigenvectors respectively are used to calculate the generalized Armitage trend test [27]. Table 3 provides the empirical type-1 error for both analysis approaches based on 2,000 replicates. While EIGENSTRAT exhibits an inflated type-1 error, the proposed overall family test maintains the theoretical significance level.
For the analysis of quantitative traits, Table 4 provides the empirical power for 500K GWAS from 2000 replicates when there is no population stratification. Under the assumption of an additive disease model for a quantitative trait, the genetic effect, a, is given as a function of the heritability, h2, the minor allele frequency pDı and the phenotypic variance, σ2, by: a = σh/[2p(1−p)(1−h2) ]0.5. In the simulation study, we assume heritabilities of h2 = 0.001, 0.005, 0.01 and 0.015 for 2,000, 2,500 and 3,000 trios. The allele frequency of the disease locus, pDı, is 0.3 and the phenotypic variance is 1. We compare the achieved power levels of the proposed overall family-based association test, Zi, with the weighting approach by Ionita-Laza et al [12], the original VanSteen approach [11], the QTDT approach [28] and population-based analysis, i.e. using linear regression of the phenotype Y on the genotype X. Bonferroni correction is used to adjust for multiple testing in the population-based analysis, FBAT, QTDT and the proposed method. The results in Table 4 suggest that the proposed association test achieves power levels that represent a major improvement over the existing methods for family-based association tests (VanSteen [11] or Ionita-Laza [12]). Our approach reaches the same power levels as the population-based analysis. For the power comparisons that are shown in Figure 1, Figure 2, and Figure 3, the number of trios is assumed to be 1,000 in 500K GWAS and the empirical powers are calculated based on 10,000 replicates at an α-level of 0.001 for the all genetic models. The results confirm that the Liptak's method combining Ti and FBATi has similar power to the population-based method, and the choice of equal weights performs well. The simulation results in Table 4 also suggest that QTDT [28] approach achieves similar power levels as the standard FBAT approach, which is consistent with previously reported findings in the literature [29]. However, both standard FBAT and QTDT are still much less powerful than the proposed overall family-based association test. Table 5 shows the empirical power for a GWAS with 100,000 SNPs in the presence of population stratification. For the parameters of this simulation study, we assume FST = 0.001, 0.005, 0.01, and 0.05, and the additive mode of inheritance at the disease locus with values for the heritability of h2 = 0.005, 0.01 and 0.015. The disease allele frequency pDı in the ancestral population is assumed to be 0.3. The phenotypic data is simulated so that their phenotypic means for two subpopulations are 0 and 0.2 respectively. Each individual/trio is assigned to either subpopulation with probability 0.5. The parental genotypes are used to estimate the ancestry for EIGENSTRAT as before. Various methods have been suggested to adjust the population stratification in a population-based designs and we compare the proposed methods with the EIGENSTRAT approach [10]. In order to maximize the power of the proposed method, we apply the EIGENSTRAT approach to the population-based component Ti of our approach, i.e. principal component analysis based on the parental genotypes and the offspring's phenotype is integrated into the generalized Armitage test for Ti [27]. To keep the power comparisons unbiased, the population-based components of the approaches by VanSteen and Ionita-Laza are also adjusted for population admixture, using the EIGENSTRAT approach. The results in Table 5 show that the proposed test statistic Zi is considerably more powerful than population-based analysis adjusted with EIGENSTRAT. QTDT is slightly more powerful than FBAT, but it is much less powerful than LIP as is in Table 4. This suggests that EIGENSTRAT should be applied only to between-family component in family-based association studies. Our unpublished work showed that the proposed approach can be less powerful than the combination of population-based analysis and EIGENSTRAT if pTi is calculated from the conditional mean model [11],[22] without adjusting population stratification.
For the assessment of the severity of pulmonary diseases, the lung volume of air that a subject can blow out within one second after taking a deep breath is an important endo-phenotype. It is referred to as the forced expiratory volume in one second (FEV1). FEV1 is an important measure for lung function and we apply the proposed method to a family-based GWAS of FEV1. The proposed method is applied to 550K GWAS Framingham Heart Study (FHS) data set for FEV1, and then we confirm whether the selected SNPs are replicated in the British 1958 Birth Cohort (BBC), another population sample, as well as two samples of asthmatics in the the Childhood Asthma management program (CAMP) [30] and an Afro-Caribbean group of families from Barbados (ACG) [31]. In FHS, 9,274 subjects were genotyped and 10,816 subjects of those had at least one FEV1 measurement. Of the 8637 participants with genotyping and FEV1 measures, only those with a call rate of 97% or higher were included. We adjusted the covariates, age, sex and the quadratic term of height that are known to be associated with FEV1. For within-family components, the FBAT statistic for quantitative trait was applied. Markers were excluded from the analysis if the number of informative families was less than 20, or the minor allele frequency was less than 0.05. In total, 306,264 SNPs were used for analysis and, based on the number of SNPs, rank-based empirical p-values, pTi, and the genome-wide significance level was obtained with Bonferroni correction. When we let n and ninf be the total number of individuals and the number of informative trios respectively, ninf: (2n−ninf) are used for the weights of Zi because some of parental phenotypes are available.
Table 6 shows the p-values for the top 10 SNPs from the proposed method. In our analysis, the genome-wide significance level at 0.05 is 1.636×10−7 and our results show that only the first ranked SNP, rs805294, is significant at the genome-wide level 0.2 with Bonferroni correction. For rs805294, we also checked the significance in other data sets, BBC, CAMP [30] and ACG [31]. In CAMP, 1215 subjects in 422 families were genotyped and there are 488 informative trios for rs809254 and in ACG, there were only 33 informative trios (Table 7). In the BBC, 1372 unrelated subjects were genotyped with the Affymetrix chip and 1323 unrelated subjects genotyped with the Illumina chip. In CAMP and ACG, age, sex and the quadratic terms of heights were adjusted and in the BBC, age, sex, height, recent chest infection and nurse were adjusted. Table 7 also shows that rs805294 is significant and their directions are same for the considered studies except for the ACG sample. In particular, in the ACG study, the MAF of the SNP is different from other studies, which indicates a different local LD structure; The ACG sample is from an Afro-Caribbean population, contrary to the other studies which only include Caucasian study subjects. In addition, the ACG sample lacks statistical power for this particular SNP, i.e. there are only 33 informative trios in this sample. Thus, the inconsistent finding in the ACG study could be attributable to genetic heterogeneity, i.e. different local LD structure/flip-flop phenomena [32], or insufficient statistical power. For meta analysis, the sample sizes are used as weights for Liptak's method and we use 13∶13∶5∶1 = FHS∶BBC∶CAMP∶ACG as weights because the between-family information is used only for FHS. If the p-value from Illumina gene chip in BBC and the p-values from FHS, CAMP and ACG are combined, then the p-values by Liptak's method using proposed weights and Fisher's method are 1.534×10−8 and 1.081×10−7 respectively, and they become 4.625×10−9 and 3.554×10−8 if the p-values from one-tailed tests are used for BBC, CAMP and ACG with the same direction of FHS. If the p-value from the Affymetrix gene chip in BBC is combined with the other studies, then they are 3.787×10−8 (Liptak's method) and 1.890×10−7 (Fisher's method) for two-tailed tests, and 1.098×10−8 (Liptak's method) and 6.236×10−8 (Fisher's method) for one-tailed tests. As a result we can conclude that rs805294 is significantly associated with FEV1 at a genome-wide scale and the gene, LY6G6C, associated with rs805293 will be investigated in further studies.
Genome-wide association studies have become one of the most important tools for the identification of new disease loci in the human genome. However, even though advances in genotyping technology have enabled a new generation of genetic association studies that provide robust and replicable findings, population stratification/genetic heterogeneity and the multiple testing problems continue to be the major issues in the statistical analysis that have to be resolved in each study. While family-based association tests provide analysis results that are completely robust against confounding due to population-substructures, the analysis approach is not optimal in terms of statistical power. Numerous approaches have been suggested to minimize this disadvantage of family-based association tests but the previous approaches had to compromise either in terms of robustness or in terms of efficiency.
In this communication, we develop an approach that efficiently utilizes all available data, while maintaining complete robustness against confounding due to population substructure. The proposed methods combines the p-values of the family-based tests (the within-component) with the rank-based p-values for population-based analysis (the between component) to achieve optimal power levels. The use of rank-based p-values for the population-based component is similar in spirit to the genomic control approach. In principle, the genomic control functions as rescaling the variance inflated due to population stratification under the assumption of the constant FST. Rank-based p-value directly rescales the statistics based on their ranks, which always generates the uniformly distributed p-value and provides validity even for varying FST due to local population stratification etc.
Although our simulations are limited to independent unascertained samples and quantitative traits, the proposed work can be easily extended to ascertained samples, large pedigree, or different trait types, etc. By replacing the parental genotypes with the sufficient statistics by Rabinowitz&Laird [19], the FBAT-statistic and the screening-statistic can be adopted straight-forwardly to designs with extended pedigrees [23]. Similarly, parental phenotypes can be incorporated into the conditional mean model [23] or its non-parametric extensions [33] as additional outcome variables. The optimal weights can vary between the different scenarios and further theoretical investigation is currently ongoing, but limited initial simulation studies suggest that equal weights, while not always the most powerful choice in such situation, will always result in more powerful analysis than currently used methods.
|
10.1371/journal.pgen.1000115 | An Integrated Approach for the Analysis of Biological Pathways using Mixed Models | Gene class, ontology, or pathway testing analysis has become increasingly popular in microarray data analysis. Such approaches allow the integration of gene annotation databases, such as Gene Ontology and KEGG Pathway, to formally test for subtle but coordinated changes at a system level. Higher power in gene class testing is gained by combining weak signals from a number of individual genes in each pathway. We propose an alternative approach for gene-class testing based on mixed models, a class of statistical models that:
provides the ability to model and borrow strength across genes that are both up and down in a pathway,
operates within a well-established statistical framework amenable to direct control of false positive or false discovery rates,
exhibits improved power over widely used methods under normal location-based alternative hypotheses, and
handles complex experimental designs for which permutation resampling is difficult.
We compare the properties of this mixed models approach with nonparametric method GSEA and parametric method PAGE using a simulation study, and illustrate its application with a diabetes data set and a dose-response data set.
| In microarray data analysis, when statistical testing is applied to each gene individually, one is often left with too many significant genes that are difficult to interpret or too few genes after a multiple comparison adjustment. Gene-class, or pathway-level testing, integrates gene annotation data such as Gene Ontology and tests for coordinated changes at the system level. These approaches can both increase power for detecting differential expression and allow for better understanding of the underlying biological processes associated with variations in outcome. We propose an alternative pathway analysis method based on mixed models, and show this method provides useful inferences beyond those available in currently popular methods, with improved power and the ability to handle complex experimental designs.
| To help increase power to detect microarray differential expression and to better interpret findings, gene-class testing or pathway analysis has become increasingly popular [1]. These approaches allow the integration of gene annotation databases such as Gene Ontology [2] and KEGG Pathway [3] to formally test for subtle but coordinated changes at the system level. Improved power of gene-class testing is gained by combining weak signals from a number of individual genes in each pathway. In addition, pathway analysis has been effectively used to examine common features between data sets [4].
The most commonly used approach for pathway analysis, the enrichment or overrepresentation analysis, uses Fisher's exact test. This method starts with a list of differentially expressed genes based on an arbitrary cutoff of nominal p-values, and compares the number of significant genes in the pathway to the rest of the genes to determine if any gene-set is overrepresented in the significant gene list. The Fisher's exact test is implemented in a number of software packages such as GOTM [5], WebGestalt [6], GENMAPP [7], ChipInfo [8], ONTO-TOOLS [9], GOstat [10], DAVID [11], and JMP Genomics (http://www.jmp.com/genomics). Although straightforward to implement and interpret, this method loses information by using only the significant genes resulted from arbitrarily dichotomizing p-values at some threshold.
More recent approaches such as Gene Set Enrichment Analysis (GSEA) [12],[13] and its extensions use continuous distributions of evidence for differential expression and are based on a modified version of the Kolmogorov-Smirnov test that compares the distribution of test statistics in a pathway to the test statistics for the rest of the genes. However, as explained in [14], the specific alternative hypothesis for coordinated association between genes in a gene-set with phenotype is likely to be a location change from background distribution. The Kolmogorov-Smirnov test used by GSEA, which detects any changes in the distribution, is often not optimally powerful for detecting specific location changes. In addition, false positives may result when genes in a gene-set have different variances compared with genes outside the pathway. Methods that test for location changes include PAGE [15] and Functional Class Scoring [16]. PAGE uses normal distribution to approximate test statistics based on differences in means for gene-set genes and other genes; Functional Class Scoring method computes mean (-log(p-value)) from p-values for all genes in a gene-set, and compares this raw score to an empirically derived distribution of raw scores for randomly selected gene-sets of the same size using a statistical resampling approach.
Other examples of permutation- and bootstrap-based methods include SAFE [17], iGA [18] and GSA [19]. However, resampling-based methods rely on exchangeability that may be hard to achieve in complex experimental designs. For example, in designs with multiple random effects and/or time-series covariance structures, great care must be taken to achieve an appropriate resampling-based null distribution. In this paper, we propose an alternative, parametric approach for gene-class testing based on mixed linear models [20], which can readily accommodate complex designs under standard parametric assumptions.
Some parametric methods and their comparisons with the proposed method are in order. Wolfinger et al. [21] and Chu et al. [22] considered using mixed models for detecting differentially expressed genes for cDNA and Affymetrix microarrays. Ng et al. [23] proposed random effects models to cluster gene expression profiles, but their gene-sets are derived by statistical learning, not based on biological knowledge. Other parametric models include the random effect model of Goeman et al. [24] and the ANCOVA model of Mansmann [25] for testing whether a particular gene-set contains any gene associated with outcome. There is an important distinction between these models and our proposed model. Tian et al. [14] formulated two statistical hypothesis for testing coordinated association between a group of genes with a phenotype of interest: hypothesis Q1 - The genes in a gene-set show the same pattern of associations with the phenotype compared with the rest of the genes; and hypothesis Q2 - The gene-set does not contain any genes whose expression levels are associated with the phenotype of interest. Goeman et al. [24] and Mansmann et al. [25] both test Q2 whereas our proposed model tests Q1. The most similar parametric method with our proposed model that tests Q1 is PAGE [15] mentioned above; test statistics for both PAGE and the proposed method are based on differences in means for gene-set genes and other genes. Our method can be viewed as an extension of PAGE with the ability to account for design of experiment (e.g. covariate adjustment) and modeling dependency between genes with a more general covariance structure.
In Materials and Methods, we describe the proposed mixed model, including assumptions and interpretations. This model incorporates both fixed effects (e.g. type of tissues, cases vs. controls) and random effects which are assumed to be sampled from a normal distribution and naturally fall into a hierarchical empirical Bayes framework. The inclusion of random effects both facilitates inferences to be made to the underlying population represented by the observed samples and is a simple mechanism for modeling a covariance structure within groups of correlated observations. Another advantage is that mixed models provide a powerful, unified and flexible framework that allows one to conduct hypothesis testing for gene-sets and accounting for other design factors at the same time. With mixed models, between-arrays normalization, adjusting for covariates and gene-set testing are achieved in a single step; in contrast, other gene-class testing methods usually require separate data processing steps for normalization, assessing statistical significance of individual genes using a test statistics such as the t-score, and comparison of the test statistics for genes in the pathway with non-pathway genes. In the Results section, we first confirm the increased power over the nonparametric method GSEA and parametric method PAGE using simulations and then illustrate the method using two microarray datasets, a human diabetic muscle dataset [12] and a dose-response study [26]. In the Discussion section, we provide some concluding comments.
Given two groups of samples and an a priori defined set of genes from a particular pathway, we are interested in testing whether the differential expression between the groups are significantly different for genes in the pathway compared with the rest of the genes. For sake of concreteness we assume without loss of generality the two groups of samples are from patients with a disease phenotype (cases) and otherwise (controls).
We assume reliable numerical values are obtained from gene expression intensities and are on the log2 scale. In single colored arrays, the expression values for each gene are derived from each spot on the array; in two-colored arrays, the expression values for each gene can be the original intensities or the ratios of expression values for experimental sample compared to reference sample. When multiple probe sets for a gene are present, they can be mapped to some standard gene IDs such as the Ensembl Gene IDs (http://www.ensembl.org) and the median is used for further analysis. This is often done for computational efficiencies of larger arrays. In the following discussion, we assume there is one value for each gene, at the end of the Discussion section, we discuss extensions of basic mixed model to accommodate multiple gene expression values per gene.
Next, to homogenize variances for all the genes included in mixed model and to make their means comparable, we standardize values for each gene with control group mean and standard deviations. Specifically, the mean and standard deviation of each gene from control patients are calculated, and all the gene values are standardized by subtracting the control group mean and dividing by the control group standard deviation. The standardized gene expression values then represent the number of standard deviations away from the “normal” gene expression values. In a time course experiment, expression values at baseline can be used similarly as control group data to standardize all measurements in the time course.
Linear mixed models is a class of statistical models that handles data where observations are not independent, such as gene expression values from the same array. They include both fixed effects and random effects, and thus are called mixed effect models. The fixed effects model the systematic effects or the mean structure of data, and the random effects account for complex covariance structure of observations, such as those between genes. In addition, they also allow inferences to be made to the entire population of samples from which the observed samples arise.
Assuming after data pre-processing, there are one measurement per gene from each array, we propose the following basic linear mixed models for comparing differential expression pattern in the pathway (or gene-set) m and the rest of genes:Here, y represents log transformed gene expression values, j = 1 if gene g is from the pathway m and j = 0 otherwise; k = 1 for case values and k = 0 for control values. The parameters μjk model systematic effects or fixed effects affecting gene expression values, and correspond to a classical cell-means model [27]. The fixed effects portion of Model 1 is equivalent to a model with intercept, indicator variable Group (case or control), indicator variable Pathway m (yes or no), and Group×Pathway m interaction effects. Although Model 1 does not include gene-specific fixed effects, we account these through standardization of gene values (Data Preprocessing in Materials and Methods) which makes expression values from different genes comparable and homogenizes their variances.
While μjk are fixed unknown parameters to be estimated from data, the terms Arrayl and Pathwaym(g) for l-th array and m-th pathway are random variables, we use the subscript (g) to emphasize values for Pathway random effects vary according to genes. We discuss in detail the construction of these random effects and the specific covariance structure accounted by them in the Materials and Methods section. Finally, ε represents variations due to measurement error and we assume εgjklm∼N(0, σ2). Parameters from the mixed model are estimated using the method of restricted maximum likelihood (REML) along with appropriate standard errors.
The hypothesis we are testing is whether the amount of differential expression between cases and controls for gene-set genes are significantly different from the other genes. This is essentially the interaction effect between gene-set and group. In terms of Model 1, we want to test H0:(μ11−μ10)−(μ01−μ00) = 0. Here, μ11−μ10 represents differential expression for genes in the pathway and μ01−μ00 represents differential expression for the rest of the genes.
In feedback or reverse regulation, in response to an input signal, genes in a gene-set may shift in both directions, that is, a fraction of gene-set genes are up-regulated and another fraction of gene-set genes are down-regulated, then testing changes in the entire gene-set will not be effective as the changes in different directions will cancel each other out. Instead, we propose modeling reverse regulation withwhere i indicates direction of changes for gene g, i = 1 for up-regulated genes and i = 0 for down-regulated genes. With this model, we estimate <1?show=[to]?> where estimates amount of up-regulation and estimates amount of down-regulation.
Because the direction of change i for each gene is estimated from data, the hypothesis we are testing in this case is equivalent to H0:{[(μ11−μ10)−(μ01−μ00)|i = 1]−[(μ11−μ10)−(μ01−μ00)|i = 0]} = 0. Therefore, is the difference of two conditional random variables, its distribution is a skewed unimodal distribution and can not be approximated well using normal distribution. We propose a Box-Cox transformation [28] of the test statistics to account for this. Specifically, to test for significance of n (e.g. 500) gene-sets, we follow these steps:
We use the Monte Carlo simulation approach [29] to simulate gene expression values with the same covariance structure as those in real microarray data. First, we fit Model 2 to real microarray data and estimate covariance parameters corresponding to variance components for random effects and residual errors ε. Next, we simulate gene values with random effects and errors generated according to the estimated covariance parameters. Because the dependency between genes are captured approximately by random effects and covariance parameters in mixed models (Materials and Methods), the simulated gene expression values will have essentially the same covariance structure as gene values in real microarray data. Also, since no fixed effects were added, the simulated data do not depend on outcome and therefore correspond to null gene-sets values.
Once we obtain nominal p-values from steps described above, we next calculate adjusted p-values to control for False Discovery Rate (FDR). An adjusted p-value of 0.05 for a gene set indicates that among all significant gene sets selected at this threshold, 5 out 100 of them are expected to be false leads.
In Models 1 and 2, we assume normal distributions for the random effects: . Here, the Array random effects model effects due to sample variations and Pathway random effects represent variations associated with different biological processes defined by pathways. The random effects have the advantage of requiring only a single parameter (e.g. ), the variance component, to be estimated. In the simulation study we accommodate 50 pathways simultaneously. For real microarray dataset, one can also construct a separate pathway “other” to include all genes not belonging to any gene-sets to be tested.
Another important advantage of random effects is that they help capture the heterogeneous covariations across genes. In particular, the Array random effects account for covariance among all observations from the same array and Pathway random effects account for covariance among genes from the same pathway. Note that the random Pathway effects vary according to genes, to model different amount of dependencies between pairs of genes. We discuss the specific covariance structure accounted by these random effects and their constructions in details next.
The Array random effects are constructed as indicator variables for each array, that is, Arrayl = I{array l}. To construct the Pathway random effects, first, calculate t-statistics for each gene based on observed data. Let be gene expression values from control samples, and be gene expression values from case samples. Compute where X̅g and Y̅g are average gene values for control samples and case samples respectively. Next, we construct Pathwaym(g) = Tg×I{pathway m}, where I(pathway m) is indicator variable for a gene belonging to pathway m. Therefore, for genes within pathway m, Pathwaym(g) varies depending on Tg and it is 0 for genes outside pathway m.
Using matrix algebra, it can be shown that Array and Pathway random effects induce a covariance structure in the marginal model that accommodates different amount of dependencies between genes (see for example, [29], page 737). More specifically, let yglm be gene expression value for gene g from pathway m on array l, then where σ2 is residual variance associated with measurement errors andHere, tg denotes value of statistic T for gene g. In (B), for genes from the same pathway, the correlations between genes depend on directions and magnitudes of their differential expression changes. So two genes are highly positively correlated if there are large differential expression changes for both genes and their changes are in the same direction. In (C), assuming most of covariations between genes come from those genes within the same pathway and genes from different pathways but on the same array are only weakly correlated, we model a common covariance between these genes. In practice, we found assuming heterogenous covariances tend to be too strong for genes from different pathways and tests for gene-sets based on it lose too much power. Comparing (A) and (B), (C), genes from the same arrays and pathways are more correlated than those from different arrays or from different pathways. In (D), we assume genes from different arrays and different pathways are independent given the arrays are from independent patients.
We performed a simulation study to assess the sensitivity and specificity of a mixed model approach compared with GSEA and PAGE which also test hypothesis Q1 in Tian et al. [14], that is, the association of gene-set genes with outcome is similar with the association for the rest of the genes. For each scenario in Table 1, two sets of 50 microarray samples were simulated for treatment and control groups. Each sample consisted of 1500 values generated from the standard normal distribution as an approximation to log transformed gene expression values. These values were assigned to 50 gene-sets, each with 30 genes. Treatment effects were added to gene-set 1 according to the parameters p, up, μ where
Therefore, among all the genes in the gene-set, there were 30×p×up up-regulated genes and 30×p×(1−up) down-regulated genes. For example, for Scenario 1 in Table 1, there were 9 ( = 30×0.3) genes in gene-set 1 with treatment effect added, among them 5 (≈30×0.3×0.5) gene values were increased with 0.2 units and the remaining 4 genes were decreased with −0.2 units. In scenes 4–6 and 7–9, the total proportions of genes with treatment effects were changed to 0.5 and 0.8 respectively. In scenes 10–18, among treated genes, 80% of genes were moved up and 20% genes were moved down. These parameters were chosen to represent different degrees of feedback and reverse regulation. For each scenario, only the first gene-set was associated with treatment-control groups and the other gene-sets were null gene-sets by design of experiment.
The javaGSEA implementation was used for GSEA analysis and the algorithm described on page 10 of [15] was used for PAGE. SAS PROC MIXED [29] was used for mixed model analysis. For datasets with up_p = 0.5, GSEA algorithm was implemented with gene list sorting mode “abs”, so genes were sorted based on absolute values; the mixed model was implemented with Model 2. For each scenario with up_p = 0.5, was estimated by applying Box-Cox transformation (Linear Mixed Model in Materials and Methods) to t-statistics of the 49 null gene-sets. The results showed the estimated was 0.7 for all scenarios for the transformed t-statistics to achieve approximate normality.
To compare the performances of Mixed Model 1 with GSEA and PAGE, we generated 20 datasets for each set of parameters p, up, μ and computed the Area Under the receiver operating characteristic Curve (AUC) for each method. The receiver operating characteristic (ROC) curves show trade-off between sensitivity and 1-specificity as the significance cutoff is varied. The AUC assesses the overall discriminative ability of the methods at determining whether a given gene-set is associated with outcome over all possible cutoffs. In addition, we calculated the test sizes of each method (the proportions of p-values less than 0.05 for null gene-sets). Because under the null hypothesis we expect the p-values to follow a uniform distribution, a method with test size equal to or less than the significance cutoff (e.g. 0.05) is desirable.
In terms of AUC, when most genes are shifted in one direction (up_p = 0.8), the mixed model and PAGE performed similarly, and they both outperformed GSEA consistently across scenarios 10–18 (Table 1, Figure 1). These results show that improved power can be gained over GSEA, which tests for any differences in distributions, by using approaches such as the mixed model or PAGE that test for location changes. When genes are shifted in both directions equally (up_p = 0.5), the mixed model performed better than both GSEA and PAGE. The better performance of the mixed model vs. PAGE shows that combining signals for up-regulation and down-regulation by Mixed Model 2 is more effective in this setting because signals from genes shifted in different directions may be cancelled out. We note also that all methods maintained proper test sizes for all scenarios.
Mootha et al. [12] compared gene expression of skeletal muscle biopsy samples from human diabetes patients and patients with normal glucose tolerance. There were 17 control patients (group NGT) and 18 diabetic patients (group DM2) in this study and 149 curated gene-sets were tested for enrichment using GSEA. They found that genes involved in oxidative phosphorylation were coordinately down regulated in human diabetes. To compare the results of the mixed model approach with GSEA and to confirm that mixed models can also detect subtle but coordinated changes in gene expression within gene-sets, we reanalyzed this data set.
Table 2 tabulates analysis results for gene-sets selected by mixed models and the GSEA method. The results for GSEA were obtained from http://www-genome.wi.mit.edu/mpg/oxphos/. For the mixed model method, the nominal p-value were estimated by fitting Model 1 and testing the interaction term Type×Pathway. Because the Pathway random effects were also included in Model 1, they induce a more general covariance structure between genes, so mixed model analysis accounts for heterogeneous variances of different pathways and gene-gene correlations. False discovery rate (FDR) adjusted p-values were also calculated, an adjusted p-value of 0.05 for a pathway indicates that among all significant pathways selected at this threshold, 5 out 100 of them are expected to be false leads.
The results show that both the mixed model and GSEA selected the pathway “OXPHOS_HG_U133A_probes” as the most significantly changed pathway and ranked the pathways “human_mitoDB_6_2002_HG_U133A_pro”, “mitochondr_HG_U133A_probes” high on their significant pathways list. While mixed model selected 9 gene-sets at 5% FDR level, all FDR adjusted p-values for GSEA method were greater than 0.2 (the minimum was 0.447). As diabetes is primarily a chronic disorder of carbohydrate metabolism, additional pathways identified by the mixed model, such as the “Glycolysis/Gluconeogenesis” and “Starch and sucrose metabolism” make biological sense. Chronic diabetes has also been associated with changes in “Tyrosine metabolism” [30], another pathway identified by the mixed model.
We next applied the mixed model method to a dose-response microarray experiment. West et al. [26] conducted experiments to study the effect of HNE (4-hydroxy-2-nonenal) on RKO human colorectal carcinoma cells. It is postulated that HNE induces cellular dysfunction in a variety of disorders such as inflammation, cancer, neurodegenerative, cardiovascular disease [31],[32]. In this study, Affymetrix U133 Plus 2.0 chips were used with RKO cells to explore transcriptional changes induced following treatment for 6 or 24 hours with 5,20, or 60 µM HNE. Figure S1 shows the dose response relationships averaged over all genes for each gene set for each treatment duration.
Our main objective was to identify gene sets with significant monotone changes over doses and to assess whether the changes were similar for the two treatment durations. With permutation based methods such as GSEA, one needs to decide a priori whether to test for trends of gene expression over different doses of HNE for each treatment duration separately or to test for trends by pooling data from both treatment durations. In contrast, the mixed model framework provides a more efficient way to incorporate information from both treatment durations, and standard methods apply for testing polynomial trends of gene expressions over different doses of HNE and for testing trend by treatment duration interaction.
We next describe the analysis workflow. First, probe sets were mapped to Ensembl Gene IDs and median expression levels for multiple probe sets corresponding to the same gene were calculated. After this step, we were left with 17278 genes and they were tested for enrichment against gene sets generated based on the biological process categories in Gene Ontology. Genes in the human genome were mapped to GO categories according to Ensembl annotation (http://www.ensembl.org). We focused on GO categories with 10 to 200 genes by removing all the other categories. Note that this is the size of a gene set when all of the genes in the genome are considered. For genes on a specific array, the gene counts for a gene set will be slightly smaller. In order to reduce the redundancy in GO, we further removed all child-categories if corresponding parent-category was within the size limitation. After the above processes, 444 remaining gene sets were used for the enrichment analysis.
Next, we calculated means and standard deviations for each gene at dose 0 for each treatment duration separately and then used these values to standardize all gene expression values. That is, the values for each gene were standardized by subtracting the dose 0 means and dividing by dose 0 standard deviations. The standardized gene expression values then represented the number of standard deviation away from the “normal” gene expression at dose 0.
Finally, we applied the mixed model with fixed effects Dose, Treatment Duration, Dose×Treatment Duration to the gene expression values. Because the data were collected at different times, the variable Batch was also added to adjust for the effects of different batches. In addition, a random Array effect was included in the model to account for correlations of genes from the same array and to facilitate inference to an entire population of arrays, not only to those considered in this study. Contrasts of parameters from this model based orthogonal polynomial coefficients were then used to test for linear trend of expression values over doses and Duration×Linear trend effect. The orthogonal polynomial coefficients are linear transformations of the natural polynomial scores and they alleviate collinearity problems of natural polynomial scores. Adjusted p-values were then computed using the R multtest package [33] to control for False Discovery Rate (FDR) using the method of Benjamini and Hochberg [34].
Because we were mainly interested in gene sets directly responding to changes in HNE, our analysis focused on gene sets with significant linear trends of expression values corresponding to monotone changes over doses. At the adjusted p-value level of 0.01, we identified 5 and 1 responsive gene sets for 6 h and 24 h treatment, respectively (Figure 2). However, after testing for a Duration×Linear Trend interaction, and refitting gene sets for which the interaction was nonsignificant, we identified 40 responsive gene sets at the adjusted p-value level of 0.01 (Figure 3). Among them, 36 out of the 40 gene sets were not identified in the individual test. These 36 gene sets represented some important biological processes that are known to be responsive to HNE treatment, such as “mismatch repair”, “double-strand break repair”, and “response to inorganic substance” (Table S1). These results demonstrated that pooling data with similar trends from both treatment durations is helpful for improving statistical power and identifying biologically meaningful gene sets.
On the other hand, the interaction tests were also used to select gene sets showing different response trends for the 6 h and 24 h treatments. Among the 12 gene sets with significant interactions (p-value<0.01), 8 of them were responsive for 6 h treatment (adjusted p-value<0.05) but not for 24 h treatment (adjusted p-value>0.95, see Figure 3). These gene sets represented biological processes that responded to HNE in a quick manner, including “cytoplasmic sequestering of protein”, “negative regulation of transcription factor import”, and “cellular response to stimulus” etc. (Table S1). Down-regulation of the biological processes “cytoplasmic sequestering of protein” and “negative regulation of transcription factor import” at 6 h will lead to the release of transcription factors that are sequestered in the cytosol, which is consistent with the significant increase in overall transcription after 6 h of HNE treatment. One gene set, “pyrimidine deoxyribonucleotide metabolism”, showed a significant response for the 24 h treatment (adjusted p-value = 0) but not for 6 h treatment (adjusted p-value = 0.33). These results indicated that although both signaling and metabolic changes were involved in oxidative stress, metabolism response was slower than the signaling response, e.g. transcription factor import.
In this paper, we have proposed linear mixed models for the analysis of microarray data at the pathway-level. This flexible, unified and practical approach can be easily implemented in common statistical software packages. The proposed model makes three main improvements over popular methods for gene-set testing: improved power through testing location shift of gene-set genes, more refined modeling of covariance structure between genes through specification of random effects, and the ability to account for complicated experimental designs through inclusion of design factors and covariate effects.
As suggested by Tian et al. [14], power is lost when GSEA tests Q1 (genes in a gene-set show the same pattern of associations with the phenotype compared with the rest of the genes) but generates the null distribution of test statistic under hypothesis Q2 (all genes in gene-set are not associated with outcome) by permuting samples. In addition, the alternative hypothesis that is of interest for Q1 is more likely to be location shift for genes in the gene-set compared to background genes; the use of an omnibus test such as the Kolmogorov test by GSEA may result in further loss in power and produce false positives for tightly correlated gene-sets. Our proposed method provides a simple way to test for location shifts in Q1 while accounting for covariance structure between genes at the same time. It provides increased power while still maintaining control of the false positive rate.
The use of random effects to account for a general covariance structure that varies according to genes in the proposed models represent our efforts for improving covariance structure modeling of current parametric methods. False positives are likely to result when dependency between genes are not accounted for [15], or through the use of homogenous correlation between all genes on the same array [23]. Our proposed model, although may not be perfect, provides a way to capture the primary heterogeneous covariance structure between genes. As genes operate with complex covariation patterns, covariance structure modeling is a challenge for parametric methods and future study with further refined modeling of dependencies between genes will extend the power and potential of mixed models and other parametric methods.
On the other hand, the strength of parametric methods such as the proposed mixed models lie in their ability to account for complicated design information. When there are multiple sources of covariation in the data, permutation or resampling methods are often difficult to employ. In contrast, mixed Models 1 and 2 can be easily extended to handle a variety of more complex designs. For example, for two-color arrays and other arrays with multiple measurements per gene on each array, Model 1 can be augmented with additional random effects corresponding to spot or block effects. When arrays are processed in multiple batches, a batch effect can be added to the model to adjust for systematic effects from different batches. Similarly, other random effects from blocks and sites where the experiments were performed can also be incorporated into the models. In the A Dose Response Study section, although we have analyzed a dose response study, time-course experiments can also be analyzed in a similar way. For example, for a time-course study with two treatments and four time points, a mixed model with fixed effects Treatment, Time and Treatment×Time plus random effects can be constructed. In addition, these models can be further extended to accommodate design information such as matched case-control pairs. Littell et al. [29] provides a comprehensive set of examples covering a wide range of mixed models and related covariance structures. Tests for multiple interaction effects in these and numerous other mixed model settings can provide valuable sentinels for scientific discovery. |
10.1371/journal.pntd.0006239 | Zika virus epidemiology in Bolivia: A seroprevalence study in volunteer blood donors | Zika virus (ZIKV), was widely reported in Latin America and has been associated with neuropathologies, as microcephaly, but only few seroprevalence studies have been published to date. Our objective was to determine the seroprevalence amongst Bolivian blood donors and estimate the future potential circulation of the virus.
A ZIKV seroprevalence study was conducted between December 2016 and April 2017 in 814 asymptomatic Bolivian volunteer blood donors residing in various eco-environments corresponding to contrasting entomological activities. It was based on detection of IgG to ZIKV using NS1 ELISA screening, followed by a seroneutralisation test in case of positive or equivocal ELISA result.
Analysis revealed that ZIKV circulation occurred in tropical areas (Beni: 39%; Santa Cruz de la Sierra: 21.5%) but not in highlands (~0% in Cochabamba, La Paz, Tarija). It was modulated by Aedes aegypti activity and the virus spread was not limited by previous immunity to dengue. Cases were geo-localised in a wide range of urban areas in Santa Cruz and Trinidad. No differences in seroprevalence related to gender or age-groups could be identified. It is concluded that ZIKV has been intensely circulating in the Beni region and has still a significant potential for propagating in the area of Santa Cruz.
| Zika virus (ZIKV) is a virus of African origin, transmitted by Aedes mosquitoes, and related to dengue and yellow fever virus. It was originally believed to be responsible for a mild febrile illness in Africa and South-east Asia. However, in recent years, ZIKV has been responsible for outbreaks in the Pacific Islands before massively spreading in Latin America and the Caribbean. On this occasion, ZIKV has unexpectedly been associated with non-vector transmission (i.e., sexual and mother-to-foetus transmission) and with severe complications such as foetal abnormalities (e.g. microcephaly) and Guillain-Barré syndromes.
Little is known about the actual proportion of the populations infected by ZIKV in Latin America. Here, we report a seroprevalence data in this region, after studying 814 asymptomatic Bolivian volunteer blood donors residing in various eco-environments corresponding to contrasting entomological activities. We conclude that ZIKV has been circulating in Bolivian tropical areas but not in highlands, and that the epidemic has not been limited by previous immunity against dengue. Specific attention should be paid to the region of Santa Cruz, where the seroprevalence is still limited, but the density of Aedes aegypti populations makes plausible further spreading of the disease.
| Zika virus (ZIKV) is an arthropod-borne Flavivirus transmitted to humans mainly by Aedes mosquitoes. It has been responsible over the last decade for outbreaks in Pacific Islands [1] (Yap Island, 2007 [2]; French Polynesia, 2013 [3]; New Caledonia, Cook Islands and Easter Island, 2014 [4]; Vanuatu, Solomon Islands, Samoa and Fiji, 2015 [5]), in Latin America (from late 2013 [6] or early 2014 [7] in Brazil, then in a large number of other countries [8]), and in the Caribbean region (e.g., 2014 in Haiti [9], 2015 in Martinique Island [10]). The first autochthonous case reported in Bolivia was in January 2016 [11].
ZIKV is of African origin, but an Asian lineage emerged presumably in the first part of the XIXth century [6] and the viruses that spread in the Pacific and the Americas are descendants of this lineage. In addition to the classical picture of large arboviral outbreaks, significant public health burden was endured in Polynesia and America when severe and formerly undescribed foetal and neurological complications of the disease [12] as well as non-vectored routes of transmission [13] were reported.
Our capacity to estimate the future spread of ZIKV disease in South America significantly depends on our knowledge of the immune status against ZIKV in the populations exposed to the potential vectors (mostly Aedes aegypti mosquitoes). Few ZIKV seroprevalence results in the region have been made available yet, most probably due to the technical difficulty to distinguish antibodies to ZIKV from cross-reacting antibodies to the other flaviviruses, in particular dengue virus. Specific detection of antibodies to ZIKV can be improved when using demanding seroneutralisation methods for the primary detection of antibodies, or for confirmation of a more convenient screening assay such as an ELISA test.
In this context, we investigated the ZIKV serological status of blood donors from different regions of Bolivia and analysed results with reference to immunity of the same populations to dengue (DENV) and chikungunya virus (CHIKV), two arboviruses known to be transmitted by the same vector locally. DENV has been reported in Bolivia since 1931. In 1948, Bolivia was declared Ae. aegypti free, but the vector reappeared in the 1980s. Since then, DENV-1, DENV-2, DENV-3, lately DENV-4 and the co-circulation of serotypes have been reported, being massively endemic in the tropical regions of Bolivia [14, 15]. CHIKV, a member of the Alphavirus genus arrived in the Caribbean in late 2013 and then spread through Latin America the following years, causing explosive outbreaks in humans[16].
We conducted a study with the help of five Bolivian regional blood banks, representing a variety of eco-environments (Santa Cruz de la Sierra and Beni have tropical climate; Cochabamba, Tarija and La Paz have colder subtropical highland climates): 814 volunteer blood donors from Santa Cruz de la Sierra (n = 200), La Paz (n = 161), Cochabamba (n = 152), Tarija (n = 196), and Beni (n = 105) provided before blood donation their oral consent for detection of IgG to ZIKV. All donors accepting to participate in the study and providing consent were considered eligible. No specific sampling was performed and leftovers of blood samples collected and stored at -20°C after completion of laboratory analyses were used. The contribution of each site was evaluated locally according to the actual possibilities to recruit donors, process samples and provide the epidemiological data requested (see below). The sample size was in agreement with previous epidemiological studies of arbovirus seroprevalence in other countries and expected to provide a reliable picture of the global epidemiological situation [17, 10]. Sampling was performed in December 2016 in the Beni region, then from March to April 2017 in the other sites, according to local logistical possibilities.
Blood samples and personal data (date of donation, sex, age, birthplace, living place, occupation and neighbourhood) were irreversibly anonymised.
Adults blood donors approved to participate in the study by providing oral consent during the face-to-face questioning before blood gift. This procedure was considered the most suitable by local blood banks. The sampling and analysis protocol was approved by the ethics committee of the Medical College of Santa Cruz.
Samples were tested for the presence of IgG to ZIKV as previously described [10,17]. In brief, alike Netto and collaborators[18] we performed an initial screening with a recombinant NS1-based ELISA test (Euroimmun, Lübeck, Germany) [19, 20] which is the only test certified for serological diagnostics of ZIKV by the responsible Brazilian authority ANVISA (Agência Nacional de Vigilância Sanitária)[18] and a subsequent Virus Neutralisation Test (VNT) for samples with a positive or equivocal ELISA result (ratio ≥0.8). VNT was performed in a 96-well format based on cytopathic effect (CPE), using ZIKV strain H/PF/2013[21], Vero ATCC cells monolayers and serum dilutions from 1:10 to 1:320. All samples were tested in duplicate with positive and negative serum controls. Sera with titre ≥1:40 were considered positive, according to the recommendations of the French National Reference Centre for Arboviruses. This testing strategy was previously demonstrated to provide specificity and sensitivity values above 98.5% in volunteer blood donors of Martinique Island tested before, during, and after the 2016 ZIKV outbreak[22] and with heavy exposure to dengue[23].
To estimate the serological immune background against dengue and chikungunya (which in Bolivia are also transmitted by Aedes aegypti), we randomly selected approximately half of the samples on each site (Beni, n = 60; Santa Cruz de la Sierra, n = 108; Tarija, n = 111; La Paz, n = 93; Cochabamba, n = 77; i.e., 449 in total). They were tested for the presence of IgG to DENV and CHIKV (Euroimmun dengue ELISA IgG and chikungunya ELISA IgG assays) according to the manufacturer's recommendations.
Donors were assigned for analysis to 3 age-groups (18–30 years-old; 31-40yo; 41-60yo) and 6 ZIKV ELISA ratio groups (<0.8; 0.8–1.09; 1.1–2.49; 2.5–3.99; 4.0–5.49; ≥5.5).
Statistical analyses were performed using IBM-SPSS Statistics v 24.0.0.0 software. Statistical association between ZIKV seropositivity and age-group or gender was evaluated, as well as relationship between ZIKV seropositivity and DENV or CHIKV seropositivity (Chi square test, significant threshold: p = 0.05).
Serological results in the different sites for ZIKV are shown in Fig 1. ZIKV has been circulating in the two regions with a typical tropical climate (Beni and Santa Cruz de la Sierra), but not in highlands in which the entomological activity is limited. In the tropical regions, the usual period of circulation of Aedes borne viruses ranges from November to April. The high seroprevalence rate in Beni (39% after VNT, (95% CI [30%–48%])) is in agreement with the report of an outbreak of febrile cases with rash locally, with laboratory PCR confirmed cases before our study (performed in December 2016). Since cases were reported in Beni until the end of the rainy season, it is likely that the final seroprevalence rate is even higher nowadays in this region.
The significant rate in Santa Cruz (21.5% after VNT, (95% CI [16%-27%])) is in line with the report of PCR confirmed cases locally before the collection of samples (in March and April 2017). Detection of cases benefitted from the local implementation of the Cenetrop National Reference Laboratory, but no clear epidemic pattern was reported. Since the study was performed locally at the end of the rainy season, it is most probable that the seroprevalence rate remained locally at a lower level than in Beni.
Of note, the proportion of ELISA positives confirmed by VNT was much higher in Beni than in Santa Cruz (63.1% vs 35.2%), possibly reflecting the frequent and intense immune stimulation against ZIKV in the Beni population.
When examining the relationship between ZIKV ELISA and VNT results, it appears, as previously observed [10, 17], that the rate of VNT confirmation increases with the value of the ELISA ratio, reaching 65% for an ELISA ratio ≥4 and 95% for an ELISA ratio ≥5.5 (S1 Table).
Table 1 shows the distribution of seropositives per site with the background of serological results for dengue and chikungunya, which circulation in Bolivia has been widely documented. Immunity to dengue virus is the rule in the tropical regions (>90% of seropositives in Beni and Santa Cruz), but also significant in the region of Tarija (44.1%). It is more limited in Cochabamba and La Paz (ca ~10%). For chikungunya, approximately half of the population is seropositive in the tropical regions and less than 10% in highlands. The differences observed with dengue most probably reflect the lower number of epidemic waves of chikungunya that hit the country. Dengue was first reported in Bolivia in 1931 and since then has been circulating intensely with major epidemics in the tropical regions[15] and, over time, multiple imported cases and possibly transient local transmission in highlands (in particular the Tarija region). By contrast, autochthonous transmission of CHIKV was first reported in Bolivia as recently as 2015 following the introduction of the virus in the Americas[24], limiting herd immunity outside the areas of intense epidemic transmission. Altogether, the spreading pattern of ZIKV follows that of other Aedes aegypti-borne viruses in Bolivia, in particular that of the recently introduced CHIKV.
S2 Table details the distribution of seropositives for ZIKV, DENV and CHIKV according to sex and age groups. Differences are minimal between groups (none reaches a significance threshold of 0.05), suggesting that age and sex do not significantly impact exposure to these arboviral diseases in the population investigated. Importantly, there is in the population studied a strong relationship between seropositivity to ZIKV and to DENV (p<10−12), but also to CHIKV (p<10−15), obviously pointing to exposure to a common risk factor: the bite of the Aedes aegypti vector of the three diseases.
This study has classical limitations linked to the sampling procedure in blood donors (in particular data regarding individuals under the age of 15 years old and in pregnant women could not be obtained). However, many previous studies on arboviruses including dengue virus[23] chikungunya virus[25], Zika virus [10] have suggested that blood donors constitute a valuable population to identify the major epidemiological trends that underlie exposure to arbovirus transmission and spread. Amongst great advantages of studying blood donor's populations, one can mention the logistical capability to perform rapidly multisite studies in the absence of robust local research infrastructure, the absence of the need to perform specific sampling, and the access to comparable populations in multiple sites that allows comparison of prevalence values.
We conclude that this seroprevalence study confirms the circulation of Zika virus in Bolivia. Despite previous reports of the presence of A. aegypti in all 5 departments investigated [14, 26, 27, 28, 29], the potential for epidemic spread is deeply modulated by the variable entomological activity in the different locations, in relation with different eco-environments (and in particular different altitudes) and as reflected by contrasted exposure to dengue or chikungunya. In the tropical areas of Santa Cruz and Beni, cases were identified from the city centres to outlying district, reflecting the wide distribution of A. aegypti. According to previous information relating to the circulation of dengue in Bolivia and corroborated by our present seroprevalence data, the spread of Zika virus was not limited by previous herd immunity to dengue virus. However, it remains possible that prior immunity to DENV modified the epidemiological pattern of the virus global spread.
The same methodological protocol has been used previously to estimate ZIKV seroprevalence in Martinique Island [10] (Caribbean region) and Cameroon[17] (Central Africa). Clearly, the transmission pattern in Bolivia is very different from the (peri-)sylvatic transmission reported in Cameroon, and more closely related to the urban transmission by the (peri-)domestic A. aegypti in Martinique. With reference to the Martinique outbreak and seroprevalence study, a minimum seroprevalence rate around 50% seems to be required to provide herd immunity that can stop ZIKV circulation. Accordingly, the ~21% rate observed in Santa Cruz (in the last phase of the arbovirus circulation period) would be insufficient to give protective herd immunity in the presence of abundant potential vectors and intense entomological activity, and with sustained circulation and potential reintroduction of the virus in Latin America. It is therefore expected that ZIKV circulation should be limited in the near future in the Beni Region (data were collected in December and the virus could circulate for at least four additional months locally, therefore the final seroprevalence rate may be even higher), but, in contrast, ecological and epidemiological conditions are favourable for further circulation of the virus in Santa Cruz, with its consequent complications especially in pregnant women.
|
10.1371/journal.pgen.1006165 | Functional Crosstalk between the PP2A and SUMO Pathways Revealed by Analysis of STUbL Suppressor, razor 1-1 | Posttranslational modifications (PTMs) provide dynamic regulation of the cellular proteome, which is critical for both normal cell growth and for orchestrating rapid responses to environmental stresses, e.g. genotoxins. Key PTMs include ubiquitin, the Small Ubiquitin-like MOdifier SUMO, and phosphorylation. Recently, SUMO-targeted ubiquitin ligases (STUbLs) were found to integrate signaling through the SUMO and ubiquitin pathways. In general, STUbLs are recruited to target proteins decorated with poly-SUMO chains to ubiquitinate them and drive either their extraction from protein complexes, and/or their degradation at the proteasome. In fission yeast, reducing or preventing the formation of SUMO chains can circumvent the essential and DNA damage response functions of STUbL. This result indicates that whilst some STUbL "targets" have been identified, the crucial function of STUbL is to antagonize SUMO chain formation. Herein, by screening for additional STUbL suppressors, we reveal crosstalk between the serine/threonine phosphatase PP2A-Pab1B55 and the SUMO pathway. A hypomorphic Pab1B55 mutant not only suppresses STUbL dysfunction, but also mitigates the phenotypes associated with deletion of the SUMO protease Ulp2, or mutation of the STUbL cofactor Rad60. Together, our results reveal a novel role for PP2A-Pab1B55 in modulating SUMO pathway output, acting in parallel to known critical regulators of SUMOylation homeostasis. Given the broad evolutionary functional conservation of the PP2A and SUMO pathways, our results could be relevant to the ongoing attempts to therapeutically target these factors.
| Posttranslational modifiers (PTMs) orchestrate the proteins and processes that control genome stability and cell growth. Accordingly, deregulation of PTMs causes disease, but can also be harnessed therapeutically. Crosstalk between PTMs is widespread, and acts to increase specificity and selectivity in signal transduction. Such crosstalk exists between two major PTMs, SUMO and ubiquitin, wherein a SUMO-targeted ubiquitin ligase (STUbL) can additionally mark SUMO-modified proteins with ubiquitin. Thereby, STUbL generates a hybrid SUMO-ubiquitin signal that is recognized by selective effectors, which can extract proteins from complexes and/or direct their degradation at the proteasome. STUbL function is critical to maintain genome stability, and it also mediates the therapeutic effects of arsenic trioxide in leukemia treatment. Therefore, a full appreciation of STUbL regulation and integration with other PTMs is warranted. Unexpectedly, we find that reduced activity of PP2A, a major cellular phosphatase, compensates for STUbL inactivation. Our results indicate that PP2A-regulated phosphorylation reduces the SUMO chain output of the SUMO pathway, thus reducing cellular dependency on STUbL and the functionally related factors Ulp2 and Rad60. Our data not only reveal a striking level of plasticity in signaling through certain PTMs, but also highlight potential "escape" mechanisms for SUMO pathway-based therapies.
| Posttranslational modification (PTM) of the proteome drives most aspects of cell growth including cell cycle transitions, DNA replication, and DNA repair. Accordingly, deregulation of key PTMs such as phosphorylation, SUMOylation and ubiquitylation causes cell cycle defects, genome instability, and malignant transformation or cell death [1]. Crosstalk between PTMs in signal transduction is widespread [2], and has recently come to the fore in the SUMO and ubiquitin field. SUMO and ubiquitin are small protein PTMs that are covalently attached to target proteins via similar enzymatic cascades of E1 activating, E2 conjugating enzymes, and E3 ligases [3].
Both modifiers can form chains, with ubiquitin chains of different topologies supporting functions that range from proteolysis to protein recruitment [1, 3]. In contrast, physiological role(s) of SUMO chains are poorly defined, and blocking their formation has no discernible impact on fission yeast viability or genotoxin resistance [4]. In budding yeast, SUMO chain-deficient mutants exhibit reduced sporulation following meiosis, and an apparently pleiotropic impact on chromatin organization, transcription and genotoxin sensitivity [5, 6]. However, an earlier study on various SUMO chain mutants in budding yeast, with the exception of a drastic SUMO all K to R mutant, found no overt genotoxin sensitivities or growth defects [7]. Thus, any physiological requirement for SUMO chains is subtle.
In contrast to any positive roles, SUMO chains that accumulate in the absence of the desumoylating enzyme Ulp2 cause severe cell growth defects, genome instability, and genotoxin sensitivity [4, 7]. Accordingly, a SUMOKtoR mutant that reduces SUMO chain formation rescues the phenotypes of ulp2∆ fission and budding yeast [4, 7]. An accumulation of SUMO chains also causes the extreme genome instability and cell cycle phenotypes of fission yeast that lack the SUMO-targeted E3 ubiquitin ligase (STUbL) Slx8-Rfp1 [4, 8].
STUbLs bind SUMO chains through their amino-terminal tandem SUMO interaction motifs (SIMs) and ubiquitinate them using their carboxy-terminal RING domains [9–12]. Thereby, SUMO-ubiquitin hybrid chains are generated that can selectively recruit effector proteins with dual SUMO and ubiquitin binding domains. These effector proteins can either support further signaling e.g. Rap80 in the DNA damage response, or drive extraction of target proteins from complexes and/or their proteasomal degradation e.g. Cdc48-Ufd1-Npl4 [13]. Interestingly, inhibiting STUbL activity in both fission and budding yeast rescues ulp2∆ cell phenotypes to a similar extent as blocking SUMO chain formation, despite an overall increase in SUMO chains [14, 15]. Therefore, unscheduled STUbL activity on SUMO-chain modified proteins, rather than the SUMO chains themselves, is toxic to ulp2∆ cells.
From the above, it is clear that SUMO pathway homeostasis is critical to multiple processes that impact genome stability and cell growth. Moreover, the SUMO pathway and associated factors, like STUbL, are important therapeutic targets in cancer and other diseases (e.g. [16–18]). Therefore, the identification of activities that modulate SUMO pathway outputs is of fundamental importance. To this end, we exploited the lethality of the STUbL mutant slx8-29 to screen for spontaneous suppressors, which resulted in the identification of a number of what we called "razors" due to their STUbL suppression. Here, we report the identification and characterization of one such razor, rzr1-1, which reveals novel functional crosstalk between the protein phosphatase PP2A and the SUMO pathway.
Hypomorphic mutants of the fission yeast STUbL Slx8 (e.g. slx8-29) are hypersensitive to a broad spectrum of genotoxins [8, 19–22]. When grown at semi-permissive temperature in the presence of genotoxins, spontaneous suppressors of slx8-29 appear with a frequency of ~10−7. Here we present the identification and characterization of one such slx8-29 suppressor, rzr1-1.
Backcrossing the slx8-29 rzr1-1 strain to wild-type confirmed that rzr1-1 is an extragenic suppressor of slx8-29. Testing the strains derived from the backcross for sensitivity to hydroxyurea (HU) and camptothecin (CPT) revealed potent suppression of the HU and temperature sensitivity of slx8-29 by rzr1-1 (Fig 1A). Interestingly however, rzr1-1 cells are more sensitive to CPT than slx8-29, so do not afford a growth benefit in this condition.
Having confirmed suppression of slx8-29 by rzr1-1, we then asked if rzr1-1 could bypass the essential functions of STUbL. The RING finger proteins Rfp1 and Rfp2 form independent heterodimers with Slx8 to constitute STUbL activity [22]. Notably, whereas rfp1∆ rfp2∆ cells are inviable [22], rfp1∆ rfp2∆ rzr1-1 triple mutant cells are able to form small colonies (Fig 1B). Therefore, consistent with the rescue of slx8-29, the rzr1-1 mutation strongly reduces cellular dependence on STUbL activity.
Classical genetic approaches to identify rzr1-1 were unsuccessful, so we re-sequenced the genomes of the backcrossed slx8-29 rzr1-1, slx8-29, and rzr1-1 strains. This analysis revealed that rzr1-1 is a duplicated sequence in the gene encoding Pab1 (Fig 1C), the PR55 (Cdc55) regulatory B subunit of protein phosphatase 2A (PP2A, [23]). The insertion occurs in the highly conserved beta-propeller or WD40 domain of Pab1, which makes contact with the catalytic subunit of the PP2A holoenzyme, and is thought to contribute to substrate selection (Fig 1C, [24]). The duplication retains the reading frame of Pab1, and so generates a hypomorphic allele, which we will refer to as pab1-1 from here on, rather than a null. Indeed, whereas pab1-1 (rzr1-1) strongly suppresses slx8-29, a full deletion of Pab1 (pab1∆) causes poor growth with morphogenesis defects [23, 25] and provides only a marginal rescue of the temperature and HU sensitivity of slx8-29 (Fig 1D).
The pab1-1 mutation lies in the beta-propeller region of Pab1 that is highly conserved within the B55 family, and could thus affect its association with the PP2A holoenzyme and/or substrates (Fig 1C, [24]). We therefore tested if slx8-29 cells were sensitive to the dosage of wild-type Pab1. We replaced the endogenous promoter of Pab1 with a tetracycline-regulated promoter in cells that also express or not the TetR repressor [26], and assayed the effect on slx8-29 cells. As for pab1-1, we found that reduced expression of wild-type Pab1 also rescued the HU sensitivity of slx8-29 cells, especially in the presence of the TetR repressor (Fig 2A). In addition, overexpression of Pab1 from an attenuated nmt promoter (nmt41, [25, 27]) strongly impaired the growth of slx8-29 but not wild-type cells (Fig 2B).
Given that Pab1 dosage is inversely correlated with the growth of slx8-29, we tested if the pab1-1 mutation affects Pab1 protein levels. Western analysis indicates that the pab1-1 mutation indeed destabilizes Pab1 and furthermore, appears to weaken its association with the PP2A catalytic subunit Ppa2 (Fig 2C). Together, these data suggest that pab1-1 suppresses slx8-29 phenotypes through a general reduction of PP2A-Pab1 activity. Consistent with this interpretation, deleting the catalytic subunit Ppa2 or the phosphotyrosyl phosphatase PP2A activator Ypa2 [28], also strongly suppresses the temperature and HU sensitivity of slx8-29 cells (Fig 2D). Overall, our data reveal that PP2A-Pab1 activity drives the pathological effects of slx8-29, and that pab1-1 likely provides an optimum general reduction in PP2A-Pab1 activity that allows cells to tolerate compromised STUbL function.
As PP2A is involved in a plethora of cellular functions, we next tested if pab1-1 is a general suppressor of HU-induced replication stress and/or temperature sensitivity. Initially, we generated double mutants of pab1-1 with either chk1∆, which lacks the G2 DNA damage checkpoint [29], or rhp51∆ that compromises homologous recombination (HR) repair [30]. In contrast to the robust rescue of slx8-29, chk1∆ pab1-1 double mutant cells showed additive sensitivity to HU (Fig 3A). Moreover, rhp51∆ pab1-1 double mutant cells were synthetically sick in the absence of drug, and exhibited synergistic sensitivity to HU (Fig 3A). Next we tested if pab1-1 can rescue cds1∆ cells, which lack the replication checkpoint [31–33], or rqh1∆ and mus81∆ cells that exhibit severe HR defects under replication stress [34, 35]. Again, the double mutant strains were either synergistically sensitive to HU (e.g. cds1∆ pab1-1), or pab1-1 provided no discernable growth benefit (Fig 3A). In addition, pab1-1 does not rescue the temperature or HU sensitivity of an allele of the Cdc48/p97 cofactor Ufd1 (ufd1-1, Fig 3B [8]). Together, these data demonstrate that the rescue of slx8-29 by pab1-1 is not the result of a general increase in cellular resistance to genotoxins.
In fission yeast, Xenopus and mammalian cells, PP2A-Pab1B55 antagonizes CDK activity by regulating its activating phosphatase Cdc25 and inhibitory kinase Wee1 (see [36–38] & refs. therein). Consistent with this, pab1-1 cells and those expressing lower than wild-type levels of Pab1 enter mitosis precociously at a reduced cell length (Figs 4A and S1A [38]). Therefore, we asked if elevated CDK activity per se mediates the rescue of slx8-29 by pab1-1. To this end, we determined the genetic interactions between slx8-29 and mutations that render CDK hyperactive or ablate the G2/M DNA damage checkpoint.
First we tested if a temperature sensitive Wee1 allele, wee1-50, was able to suppress slx8-29. Wee1-50 cells, like pab1-1, enter mitosis at a reduced cell size due to hyperactive CDK [39]. However, wee1-50 was unable to rescue the temperature and HU sensitivity of slx8-29 (Fig 4B). As Wee1 could have functions beyond antagonizing CDK that obscure any rescue of slx8-29, we also tested if a hyperactive allele of CDK, cdc2-1w [40], was able to rescue slx8-29. Again, cdc2-1w could not restore growth to slx8-29 at its restrictive temperature in the presence of HU, conditions in which pab1-1 affords a near complete rescue (Fig 4C).
Finally, we tested if ablating the G2 DNA damage checkpoint, which normally inhibits CDK and the cell cycle following DNA damage, could rescue slx8-29. In the presence of HU at 35°C, conditions in which both pab1-1 and chk1∆ cells are fully viable, only pab1-1 was able to rescue of slx8-29 (Fig 4D). Moreover, pab1-1 was able to rescue slx8-29 cell growth in the presence of MMS, in addition to rescuing their temperature and HU sensitivity (Fig 4D).
Taken together, these data demonstrate that although advanced mitosis is one phenotypic consequence of pab1-1, as it is for mutants of the PP2A catalytic subunit Ppa2 [36], it is not the mechanism for slx8-29 rescue.
In fission yeast, reducing SUMO chain formation renders STUbL activity largely redundant [4]. For example, the SUMOK14,30R or SUMOD81R chain deficient mutants allow slx8-29 cells to grow at high temperature, and in the presence of HU [4]. Therefore, we considered the possibility that the pab1-1 mutation affects SUMO pathway activity and reduces SUMO chain production.
We compared SUMO conjugate levels in slx8-29 single mutant cells versus double mutants of slx8-29 with pab1-1 and the controls wee1-50 and chk1∆. The characteristic accumulation of high molecular weight (HMW) SUMO conjugates was observed in slx8-29 single and slx8-29 wee1-50 or slx8-29 chk1∆ double mutants (Fig 5A). Although difficult to quantify due to their "smeary" nature, in contrast to slx8-29 and the double mutant controls, HMW SUMO conjugates are reduced in slx8-29 pab1-1 cells (Fig 5A). Because each of the slx8-29 pab1-1, slx8-29 wee1-50 and slx8-29 chk1∆ double mutants continue cycling, whereas slx8-29 cells accumulate in G2, cell cycle position does not contribute significantly to the observed SUMO levels. Moreover, the observed HMW SUMO conjugate levels correlate well with the growth of each strain in the presence of HU i.e. reduced levels support slx8-29 growth (Fig 4). Interestingly, ectopic expression of Pab1 in wild-type or slx8-29 cells leads to a dose-dependent increase in HMW SUMO conjugates, which is more pronounced in slx8-29 cells (Fig 5B). Because increased Pab1 dosage is toxic in slx8-29 cells (Fig 2), there is again a good correlation between HMW SUMO conjugates and slx8-29 cell viability.
As slx8-29 is a temperature sensitive and not a null allele, pab1-1 could conceivably "reactivate" it, leading to the observed reduction in HMW SUMO conjugates. We tested this possibility using a specific STUbL substrate we recently identified. In Nup132 mutant cells, the SUMO E3 ligase Pli1 is degraded in an Slx8-dependent manner [21]. Consistent with this, Pli1 is stabilized in Nup132 mutant cells by the slx8-29 mutation (Fig 5C, [21]). Importantly, Pli1 stabilization by slx8-29 is not reversed by pab1-1 (nup132∆ slx8-29 pab1-1, Fig 5C). This result indicates that pab1-1 does not reactivate the slx8-29 mutant, but instead affects the accumulation of HMW SUMO conjugates either through the regulation of SUMO pathway factors, or activities that act in parallel to STUbL.
Our data thus far indicate that pab1-1 reduces the burden of SUMO chains in STUbL mutant cells, making them less dependent on this normally essential activity. We therefore tested the effect of pab1-1 on other SUMO pathway mutants that experience SUMO chain-mediated toxicity.
In both fission and budding yeast Ulp2 SUMO protease mutants, SUMO chains accumulate and cause genome instability [4, 7]. This is evidenced by the suppression of ulp2∆ phenotypes in the presence of a SUMO chain blocking mutant e.g. SUMOK14,30R [4, 7]. Strikingly, as for slx8-29, we found that pab1-1 strongly suppressed the HU sensitivity of ulp2∆ cells (Fig 6A). Moreover, western analysis of SUMO again revealed a small but detectable reduction in HMW species in ulp2∆ pab1-1 double mutant versus ulp2∆ single mutant cells (Fig 6B).
The SUMO mimetic and genome stability factor Rad60 physically interacts with STUbL, and is functionally integrated with the SUMO pathway [4, 22, 41–44]. Rad60 contains two SUMO-like domains (SLD1 and SLD2) each of which interacts with different components of the SUMO pathway [4, 22, 43]. SLD1 interacts with STUbL, whereas SLD2 interacts with the SUMO conjugating enzyme Ubc9.
The SLD2:Ubc9 complex exists in competition with the SUMO:Ubc9 complex, as both use the same interface on Ubc9 [4, 43]. Because the SUMO:Ubc9 complex promotes SUMO chain formation in vivo, SUMO chains likely form in a rad60E380R mutant that is unable to interact with Ubc9, due to an excess of the SUMO:Ubc9 complex [4]. Consistent with this, SUMOK14,30R partially suppresses rad60E380R HU sensitivity [4]. Interestingly, pab1-1 also partially suppresses rad60E380R sensitivity to HU and UV irradiation, and the HU sensitivity of another Rad60 mutant, rad60-4 (Fig 6C). Together, these data indicate that pab1-1 may mitigate SUMO-chain mediated toxicity in Ulp2 and Rad60 mutant cells, as it does in slx8-29 cells. Moreover, as for slx8-29, the phenotypes of rad60-4 and ulp2∆ cells are not rescued by the hyperactive CDK allele cdc2-1w.
Given that Rad60 is phosphorylated during replication stress, and that it both physically and functionally associates with STUbL [20, 22, 41, 43, 44], we asked if it or other SUMO pathway factors were potential targets of PP2A-Pab1. To this end, we analyzed the phosphorylation status of Rad60, Slx8, Ubc9, and Pli1 in wild-type or pab1-1 cells by comparing their migration on regular or phos-tag PAGE gels [45]. Amongst these proteins, only Rad60 phosphorylation was detectably enhanced by the pab1-1 mutation (Fig 6D). Strikingly, this Rad60 hyper-phosphorylation was abolished when putative casein kinase sites in Rad60, serine S32 and S34, were mutated to alanine (Fig 6D, [44]). As Rad60 controls SUMO pathway output [4, 43], this result raised the possibility that Rad60 phosphorylation might mediate the suppressive effects of pab1-1 on slx8-29, ulp2∆ and rad60 mutants. However, pab1-1 also rescues the HU sensitivity of rad60S32,34A, which is refractory to pab1-1 induced phosphorylation (Fig 6D and 6E). Therefore, although pab1-1 impacts Rad60 phosphorylation, it is not the critical or sole target of PP2A-Pab1.
Homologous recombination (HR)-dependent chromosome linkages form that prevent normal chromosome segregation in rad60 and slx8 mutant cells after replication stress [22, 41, 42]. We therefore analyzed chromosome segregation in slx8-29 pab1-1, slx8-29, rad60-4 pab1-1, rad60-4, and pab1-1 cells following release from HU-induced cell cycle arrest. After acute treatment, 4–5 hrs in 15 mM HU, slx8-29 and rad60-4 underwent abnormal mitoses in 32% and 31% of cells, respectively (Fig 7). In contrast, mitosis was aberrant in only 4% of pab1-1 or slx8-29 pab1-1 cells, and 14% of rad60-4 pab1-1 cells (Fig 7). These data are consistent with the partial rescue of the chronic HU sensitivity of slx8-29 and rad60-4 by pab1-1 (Fig 6C), and moreover, indicates that pab1-1 suppresses aberrant HR in these mutants after replication stress.
Ulp2 mutant cells also undergo recombination-dependent mitotic catastrophe following replication stress, which can be suppressed by reducing SUMO chain formation [4, 7, 46]. We therefore tested the impact of pab1-1 on chromosome segregation in ulp2∆ cells following release from acute HU treatment. As for rad60-4 and slx8-29, we observed increased mitotic abnormalities in ulp2∆ (~50%), which were reduced (~19%) in ulp2∆ pab1-1 cells (Fig 7). Again, this result mirrors the partial rescue of the HU sensitivity of ulp2∆ cells by pab1-1 (Fig 6A).
Sumoylation and phosphorylation are master regulators of cell growth and genome stability, with defects in the homeostasis of these key PTMs driving human disease. Accordingly, deregulation of PP2A, an abundant heterotrimeric serine/threonine phosphatase, is observed in many cancers [47]. Cancer associated mutations are found in each of the PP2A subunits: A (scaffold), B (regulatory/substrate targeting), and C (catalytic). Given its hub role in cell growth and tumorigenesis, PP2A is a therapeutic target of interest [47].
Likewise, the SUMO pathway and its regulators such as SENPsULPs and STUbL are therapeutically valuable in a number of cancers, including leukemia and those overexpressing the MYC oncogene [17, 18]. The widespread interest in PP2A and the SUMO pathway as therapeutic drug targets makes it important to identify potential compensatory mechanisms that could allow cells to escape such interventions.
Here, we reveal that a hypomorphic allele of a B subunit of fission yeast PP2A, Pab1B55, is a potent suppressor of the phenotypes of STUbL, Ulp2 and Rad60 mutants. Strikingly, in a separate screen for spontaneous suppressors of the HU sensitivity of rad60E380R (to be reported elsewhere), we noted that some had a short cell phenotype, similar to that of pab1-1. Sequencing revealed these suppressors to be truncating mutations in the gene encoding Ypa2, an activator of fission yeast PP2A [28]. Together with the fact that ypa2∆ also suppresses slx8-29 phenotypes (Fig 2D), this result underscores the intriguing functional relationship between STUbL, Ulp2, Rad60 and PP2A. Evolutionary conservation of these factors suggests that such compensation could occur in human cells, undermining targeted therapies.
An overt phenotype of pab1-1 cells, as for mutations in certain other PP2A subunits, is their entry into mitosis at a reduced cell size due to elevated CDK activity (Figs 4A and S1A; [36–38] & refs. therein)). Defects in the SUMO pathway have previously been linked to a failure of cells to reenter the cell cycle after completion of DNA repair e.g. [48, 49]. Therefore, we initially speculated that the suppressive effects of pab1-1 were due to elevated CDK activity, and eventual bypass of the checkpoint induced cell cycle arrest. Indeed, the sickness of pab1-1 when combined with DNA repair mutants such as rhp51∆ and rqh1∆ is consistent with an attenuated G2/M cell cycle checkpoint, as in chk1∆ cells [50]. Moreover, the synthetic sickness of pab1-1 and cds1∆ in the presence of HU mirrors that of cds1∆ chk1∆ double mutants, which fail to delay the cell cycle in response to replication inhibition (Fig 3A, [31]).
However, neither hyperactive CDK (e.g. cdc2-1w, wee1-50), nor ablating the damage checkpoint (chk1∆) suppresses STUbL dysfunction. This specificity is all the more surprising given that cdc2-1w shares several phenotypes with pab1-1, including small cell size and sensitivity to CPT [51]. Therefore, pab1-1 likely affects the phosphorylation status of another kinase target, which in turn reduces cellular dependency on STUbL activity.
Others and we have shown that reducing SUMO chain formation bypasses the need for the SUMO protease Ulp2 [4, 7]. Moreover, the functions of STUbL can be bypassed by blocking SUMO chain formation in fission yeast [4, 13]. Western analysis of STUbL and Ulp2 mutants revealed that the pab1-1 allele reduces the abundance of HMW SUMO conjugates in both backgrounds, whereas overproducing Pab1 causes an increase. This effect is not due to cell cycle position, indicating that PP2A-Pab1B55 either directly or indirectly modulates SUMO pathway homeostasis.
SUMO and its E3 ligases are subject to phosphorylation-dependent regulation, providing a potential avenue for direct PP2A-dependent regulation [52–54]. However, we were unable to detect PP2A-Pab1 regulated phosphorylation of any SUMO pathway factors, other than Rad60. In addition, Rad60 hyper-phosphorylation was excluded as a contributor to the suppression of slx8-29, ulp2∆ or rad60 mutants by pab1-1. This illustrates the difficulty in identifying one amongst a plethora of cellular targets of PP2A-Pab1 that mediates the effects of pab1-1 on the SUMO pathway. This is certainly a goal for the future, and may be assisted by more global approaches such as phosphoproteome analysis of wild-type and pab1-1 cells. In this regard, a recent study compared global protein phosphorylation levels between wild-type and pab1∆ cells [55]. Unfortunately, amongst the numerous phosphoproteins identified, none were obvious candidates for mediators of the pab1-1 phenotypes revealed here. This study also highlighted a limitation with such proteome-wide analyses, wherein the data are “swamped” by the most abundant PP2A-Pab1 targets such as metabolic enzymes. These may mask low abundance factors that are critical for the pab1-1 SUMO-related phenotypes.
We also considered the possibility that the PP2A-Pab1B55 enzyme is a target of STUbL, Ulp2 and Rad60, and that pab1-1 bypasses the requirement for such regulation (see below). Indeed, in budding yeast, STUbL activity prevents the kinetochore proximal accumulation of a distinct PP2A complex, PP2A-Rts1B56 [56]. How STUbL antagonizes PP2A-Rts1B56 accumulation was not determined, but it was suggested to be through an indirect mechanism involving activation of the tension-sensing pathway, leading to increased Rts1B56 recruitment via Sgo1 [56]. Nevertheless, the positive genetic interaction between STUbL and PP2A-Rts1B56 raises interesting questions about the crosstalk between the SUMO and PP2A pathways across species. For example, does Rts1B56 mutation suppress budding yeast Ulp2 and Esc2 (Rad60 orthologue) mutants? If so, this could indicate that the SUMO-related functions of PP2A rely on distinct complexes in different organisms i.e. PP2A-Rts1B56 in budding yeast or PP2A-Pab1B55 in fission yeast.
To test PP2A-Pab1B55 as a target of the SUMO pathway, we analyzed Pab1 protein levels, potential SUMOylation state, and subcellular localization in slx8-29 and ulp2∆ cells, but observed no significant differences (S1B and S1C Fig). In addition, two recent proteomic analyses of protein SUMOylation in fission yeast failed to detect modification of any of the abundant PP2A complex proteins, even in cells with compromised STUbL activity [57, 58]. Although we cannot completely exclude direct or indirect regulation of PP2A-Pab1B55 by the SUMO pathway, we favor a model in which the hyper-phosphorylation of PP2A-Pab1B55 targets in pab1-1 cells mitigates the phenotypes of the tested SUMO pathway mutants.
For example, in pab1-1 cells, a pathway parallel to STUbL that degrades HMW SUMO species could be engaged e.g. through a phosphodegron in key SUMO conjugates. However, because pab1-1 suppresses slx8-29, ulp2∆ and rad60 phenotypes, this would imply the existence of a single or an overlapping set of SUMO conjugates that cause the phenotypes of each mutant. Although possible, it seems more likely that PP2A-Pab1B55 reduces the SUMO chain output of the SUMO pathway, thus providing pan-suppression of STUbL, Ulp2 and Rad60 mutants.
Replication stress induces lethal HR-dependent chromosome missegregation in STUbL, Rad60 and Ulp2 mutant cells [20, 22, 42, 46]. We found that in keeping with the ability of pab1-1 to improve the growth of each mutant in the chronic presence of HU, pab1-1 reduces HU-induced mitotic chromosome missegregation in the mutants. As SUMO chains drive the increased mitotic recombination and chromosome missegregation in ulp2∆ budding yeast [46], our chromosome segregation analyses of slx8-29, rad60-4 and ulp2∆ are again consistent with a role for PP2A-Pab1B55 in controlling SUMO pathway homeostasis.
Overall, we have identified critical functional crosstalk between the SUMO pathway and the major serine/threonine phosphatase PP2A-Pab1B55. Interplay between these master regulators of cell growth and genome stability is intrinsically important, and in the future will likely provide a basis for mechanistically defining phosphorylation-dependent regulation of SUMO pathway output.
Standard methods for S. pombe were performed as described previously [59]. All strains (Table 1) are of genotype ura4-D18 leu1-32 unless otherwise stated.
Cells were grown at 25°C to logarithmic phase (optical density at 600 nm [OD600] of 0.6 to 0.8), spotted in 5-fold dilutions from a starting OD600 of 0.5 on plates supplemented with the relevant drug. The plates were then incubated at 25 to 35°C for 3 to 5 days.
For live imaging, cells were grown in liquid EMM medium supplemented with leucine, uracil, arginine, and histidine (LUAH) and sterilized by filtration to logarithmic phase. Wide-field images of live cells were acquired using a Nikon Eclipse microscope with a 100x Plan Apochromat DIC H oil immersion objective and a Photometrics Quantix charge-coupled device camera. Images were analyzed with NIH ImageJ software. For staining with 4',6-diamidino-2-phenylindole (DAPI), cells were fixed for 5 minutes in 70% ethanol (EtOH), washed in PBS, and resuspended in 250 ng/ml of DAPI prior to imaging.
To immunoprecipitate N-terminally FLAG-tagged wild type or mutant Pab1, cell pellets were resuspended in 0.4 ml of Sp-lysis buffer (50 mM Tris, pH 8, 150 mM NaCl, 1 mM EDTA, 10% glycerol, 0.1% Nonidet P-40) [8], supplemented with 2 mM PMSF, and Complete protease inhibitor tablet, EDTA-free (Roche Applied Science), and lysed by beating with silica–zirconia beads three times at 5.0 m/s for 20 s in a FastPrep-24. After 10 min clarification by centrifugation at 16,000 x g in a microfuge at 4°C, supernatant was quantified for protein concentration based on OD reading at 280 nm. A total of 1 mg of proteins were incubated with 20 μl of Protein G MagBeads (GenScript) that have been preequilibrated with 10 μg of mouse-anti-FLAG antibodies (M2, Sigma). After 1 h of binding at 4°C, the beads were washed extensively with Sp-lysis buffer, and eluted with 2x LDS Sample Loading buffer (Life Technologies). One fifth of the eluted proteins, along with 10 μg (1%) of the input, were separated by SDS-PAGE. Western blotting was carried out as previously described [8, 58]. The membrane was blocked in 1% w/v non-fat milk in phosphate buffer saline solution with 0.1% v/v Tween-20, probed with primary antibodies, followed by HRP or IRDye-conjugated secondary antibodies, and detected either using an ECL Dura system (Pierce) on film; or scanning on an ODYSSEY scanner (Li-Cor). TAP-tagged proteins were probed with Peroxidase-Antiperoxidase (PAP), then directly detected using ECL.
Protein phosphorylation was identified by comparing the migration of a protein species on 10% Tris-Glycine gel (Life Technologies) to that on 10% SuperSep Phos-tag (50 μmol/L, Wako) [45].
To extract genomic DNA, 10 ml of saturated cultures were collected by centrifugation and washed with 0.5 ml of H2O. The cell pellet was resuspended in 0.2 ml of extraction buffer (0.2% Triton X-100, 1% SDS, 100 mM NaCl, 10 mM Tris-HCl, pH 8.0) and transferred to a screw cap 1.5 ml tube in which 0.2 ml of PCIA (phenol:chloroform:isoamyl alcohol at a ratio of 25:24:1) has been added. The cells were lysed by bead-beating as described above. Afterward, the lysate was centrifuged for 5 min at 16,000 x g in a microfuge, and the upper layer was transferred to a fresh tube and extracted twice with chloroform. The aqueous layer was digested with 50 μg of RNase A for 30 min at 37°C, then extracted again with PCIA, followed with two chloroform extractions. The genomic DNA was precipitated with 3 M sodium acetate and 100% ethanol. The DNA pellet was washed with 70% ethanol, dried and dissolved in 80 μl of TE (10 mM Tris, pH 8.0, 1 mM EDTA).
For Next Generation Sequencing, 1 μg of genomic DNA was sheared on an S2 Covaris set at 10% duty cycle, intensity of 5, 200 cycles per burst for 120 seconds, to obtain fragments of 200–300 bp in sizes. The fragments were then end-repaired, A-tailed with Taq Polymerase, kinased, and ligated to standard TruSeq (Illumina) barcoded adapters following manufacturer recommended protocols. The library was then amplified by PCR for 6 cycles. The amplified libraries were gel purified to select for DNA products of between 200–250 bp for single read 1x100 sequencing. The 100bp reads were generated by the HISeq 2000 Analyzer (Illumina) located at the Scripps DNA Sequencing Facility. The Genome Analyzer Pipeline Software (currently Casava v1.8.2) was used to perform the early data analysis of a sequencing run, using tools including image analysis, base calling, and demultiplexing. The reads were aligned to the genome with Bowtie 0.12.9 software using parameter to keep only the best scoring singleton.
To overexpress the pab1 gene, the pab1 coding sequence was amplified from a cDNA library using the primer pair Omn330 and Omn331 (Omn330: 5’-AGATTCATATGGATGATATAGAAGACTCTTTGGATC-3’; Omn331: 5’-GACTAGGATCCTTAGAGCTTAGAGAAAACAAAAAGATTATTAG-3’), and cloned into pREP41 and pREP1 plasmids at the NdeI and BamHI sites, to generate pREP41-pab1 and pREP1-pab1.
|
10.1371/journal.pgen.1005553 | Metabolomic Quantitative Trait Loci (mQTL) Mapping Implicates the Ubiquitin Proteasome System in Cardiovascular Disease Pathogenesis | Levels of certain circulating short-chain dicarboxylacylcarnitine (SCDA), long-chain dicarboxylacylcarnitine (LCDA) and medium chain acylcarnitine (MCA) metabolites are heritable and predict cardiovascular disease (CVD) events. Little is known about the biological pathways that influence levels of most of these metabolites. Here, we analyzed genetics, epigenetics, and transcriptomics with metabolomics in samples from a large CVD cohort to identify novel genetic markers for CVD and to better understand the role of metabolites in CVD pathogenesis. Using genomewide association in the CATHGEN cohort (N = 1490), we observed associations of several metabolites with genetic loci. Our strongest findings were for SCDA metabolite levels with variants in genes that regulate components of endoplasmic reticulum (ER) stress (USP3, HERC1, STIM1, SEL1L, FBXO25, SUGT1) These findings were validated in a second cohort of CATHGEN subjects (N = 2022, combined p = 8.4x10-6–2.3x10-10). Importantly, variants in these genes independently predicted CVD events. Association of genomewide methylation profiles with SCDA metabolites identified two ER stress genes as differentially methylated (BRSK2 and HOOK2). Expression quantitative trait loci (eQTL) pathway analyses driven by gene variants and SCDA metabolites corroborated perturbations in ER stress and highlighted the ubiquitin proteasome system (UPS) arm. Moreover, culture of human kidney cells in the presence of levels of fatty acids found in individuals with cardiometabolic disease, induced accumulation of SCDA metabolites in parallel with increases in the ER stress marker BiP. Thus, our integrative strategy implicates the UPS arm of the ER stress pathway in CVD pathogenesis, and identifies novel genetic loci associated with CVD event risk.
| Cardiovascular disease is a strongly heritable trait. Despite application of the latest genomic technologies, the genetic architecture of disease risk remains poorly defined, and mechanisms underlying this susceptibility are incompletely understood. In this study, we performed genome-wide mapping of heart disease-related metabolites measured in the blood as the genetic traits of interest (instead of the disease itself), in a large cohort of 3512 patients at risk of heart disease from the CATHGEN study. Our goal was to discover new cardiovascular disease genes and thereby mechanisms of disease pathogenesis by understanding the genes that regulate levels of these metabolites. These analyses identified novel genetic variants associated with metabolite levels and with cardiovascular disease itself. Importantly, by utilizing an unbiased systems-based approach integrating genetics, gene expression, epigenetics and metabolomics, we uncovered a novel pathway of heart disease pathogenesis, that of endoplasmic reticulum (ER) stress, represented by elevated levels of circulating short-chain dicarboxylacylcarnitine (SCDA) metabolites.
| Despite the strong heritability of cardiovascular disease (CVD), its underlying genetic architecture remains incompletely characterized. Genomewide association studies (GWAS) have converged on association of CVD with a locus on chromosome 9p21 [1], but the variants confer modest risk and are of unclear functional significance. One limitation of GWAS studies for complex diseases is the search for association with disease as a binary endpoint, rather than with molecular markers that define risk. An alternative approach is to search for variations in the genome that associate with variation in complex traits. In fact, many diseases can be defined by an underlying quantitative scale, and these “intermediate” traits may have a stronger functional relationship to the causative gene, thereby providing a stronger signal for the disease process. Metabolite levels measured by the emerging tools of metabolomics may be particularly useful for such studies. Indeed, integration of GWAS with metabolomic profiles in population-based cohorts [2] has demonstrated that as much as 12% of variance in metabolite levels is determined by single nucleotide polymorphisms (SNPs). However, most studies of this type performed to date have not used disease-burdened cohorts, so clear linkages between genetic signals, intermediate phenotypes and disease remain to be discovered.
Metabolomic profiling has identified novel biomarkers for CVD risk [3–5]. For example, a cluster of heritable [6] short-chain dicarboxylacylcarnitine (SCDA) metabolites measured in plasma (comprised of the mono-carnitine esters of short-chain, alpha-, omega-diacids), a cluster of long-chain dicarboxylacylcarnitines (LCDA), and a cluster of medium-chain acylcarnitines (MCA) predict CVD events in cardiovascular cohorts [4, 5], in patients undergoing coronary artery bypass grafting [3], and add incremental risk prediction to robust clinical models inclusive of >20 variables [5]. Little is known about the biological pathways represented by these metabolites and how they may predispose to CVD. Thus, we hypothesized that integration of metabolomics with genetics, epigenetics, and transcriptomics could define novel mechanisms of CVD pathogenesis by identifying metabolic quantitative trait loci (mQTL) that are CVD risk factors. We performed a GWAS of metabolite levels in a large cardiovascular cohort referred for cardiac catheterization (CATHGEN, N = 1490) and validated our findings in a second cohort (CATHGEN, N = 2022). A proportion of study subjects (44%) did not have clinically significant atherosclerotic coronary artery disease at time of catheterization; regardless, all individuals were analyzed given that metabolites predict risk of CVD events even in individuals without coronary artery disease, and because these individuals are still at risk for these events. We found that genetic loci that strongly associate with SCDA levels also predict incident CVD events, and are linked to ER stress. Genes differentially methylated in subjects at the extremes of SCDA levels also report on ER stress. Gene expression quantitative trait loci (eQTL) pathway analysis identified ER stress as an expression module associated with disease risk, particularly highlighting the ubiquitin proteasome system (UPS) arm of ER stress. Thus, this multi-platform “omics” approach identified a molecular pathway (ER stress and dysregulation of the UPS) associated with a prevalent complex disease.
Table 1 displays baseline characteristics of the study population. PCA of metabolomic data identified 14 factors with metabolites in each factor clustering within biochemical pathways (S1 Table), and clustering similar to our previous studies [3–5, 7]. For this study, we performed GWAS using the top three PCA-derived factors: factor 1 (composed of MCA metabolites), factor 2 (composed of LCDA metabolites), and factor 3 (composed of SCDA metabolites), all of which we have previously identified as predicting CVD events (S2 Table) [3–5]. S1 Fig details the overall study flow.
Factor 1, factor 2 and factor 3 scores were used as the quantitative traits in GWAS analysis to identify mQTL. Q-Q plots suggested the presence of SNPs associated with levels of each of the three metabolite factors (S2, S3 and S4 Figs). Several SNPs were significantly associated with metabolite factor levels at genomewide significance (p≤10−6) in additive models in the discovery cohort (Fig 1A–1F) and confirmed (p≤0.05) in the validation cohort (Table 2). Specifically, eight SNPs were associated with factor 1 (MCA) levels in any race, but with only two of these SNPs showing more than nominal significance in the validation cohort (Table 2): rs10987728 (in cyclin dependent kinase 9 [CDK9]) and rs6738286 (intergenic between transition protein 1 [TNP1] and disrupted in renal carcinoma 3 [DIRC3]). Twelve SNPs were associated with factor 2 (LCDA) levels in any race (Table 2), with only two of them showing more than nominal significance in the validation cohort (rs12129555 just downstream from polymeric immunoglobulin receptor [PIGR] and rs17025690 in Usher syndrome 2A [USH2a]). Factor 3 (SCDA) showed the strongest mQTL with twelve SNPs being associated with SCDA levels in any race (Table 2), and four of these SNPs showing more than nominal significance in the validation cohort: rs2228513 in HERC1 HECT and RLD domain containing E3 ubiquitin protein ligase family member 1 (HERC1), rs10450989 in ubiquitin specific protease 3 (USP3), rs11771619 in round spermatid basic protein 1-like (RSBN1L), and rs1869075 (intergenic between F-box protein 25 [FBXO25] and glutamate rich 1 [ERICH1]). Effect sizes (β, i.e. per 1 unit change in factor levels) ranged from to -0.38 to 2.17 (factor 1), -0.19 to 1.16 (factor 2), and -0.43 to 1.72 (factor 3).
In meta-analyses combining the race-stratified results, eleven SNPs were associated with factor 1 (MCA) levels, with three of these SNPs showing more than nominal association (Table 3); one of these (rs10987728 in CDK9) was also identified from race-stratified results and two (rs16990949 in PDX1 C-terminal inhibiting factor 1 [PCIF1]) and rs543129 [intergenic between cutaneous T-cell lymphoma-associated antigen 1 (CTAGE1) and retinoblastoma binding protein 8 (RBBP8)]) were new mQTL identified in these race meta-analyses. Eight SNPs were associated with factor 2 (LCDA) levels (Table 3); one gene had been identified in race-stratified analyses (ZNF521) but showed stronger association in the validation cohort in these analyses, and rs352216 near frizzled class receptor 3 (FZD3) was a new mQTL. Factor 3 (SCDA) again had the largest number and strongest mQTL with fourteen SNPs associated with SCDA levels, with eight SNPs showing more than nominal significance in the validation cohort (Table 3). SNPs in USP3, HERC1 and OLFM4|SUGT1 (intergenic between olfactomedin 4 and SGT1, suppressor of G2 allele of SKP1 [S. cerevisiae]) had already been identified in race-stratified analyses; additional mQTL identified in these race meta-analyses included rs12589750 and rs3853422 (in or near stonin 2 [STON2] and sel-1 suppressor of lin-12-like (C. elegans) [SEL1L]), rs930491 and rs11827377 (both intergenic between ribonucleotide reductase M1 [RRM1] and stromal interaction molecule 1 [STIM1]), rs11242866 (between solute carrier family 22 (organic cation transporter), member 3 [SLC22A23] and PX domain containing 1 [PXDC1]), and rs4544127 (near FRAS1-related extracellular matrix protein 2 [FREM2] and stomatin-like protein 3 [STOML3]).
Thus, to summarize, the most robust results overall were for mQTL associated with SCDA metabolite levels (factor 3) including an mQTL composed of USP3 (rs10450989) and HERC1 (rs2228513); and a locus composed of STON2 (rs12589750) and SEL1L (rs3853422), with loci meeting genomewide significance in the discovery cohort (p≤10−6), strong significance in the validation cohort (p = 2.4x10-3–7.7x10-7, except rs3853422 which only showed borderline significance [p = 0.01]), and stronger association in the meta-analyses (p = 1.6x10-6–7.2x10-12). The next strongest overall results for SCDA mQTL (based on race-stratified or race-combined meta-analysis p-values) in descending order of significance were for RRM1|STIM1, OLFM4|SUGT1, SLC22A23|PXDC1, RSBN1L, FBXO25|ERICH1, and FREM2|STOML3. The next strongest results overall were for mQTL associated with LCDA (factor 2) levels with SNPs in PIGR, ZNF521, USH2A and FZD3 showing more than nominal significance in the validation cohort. Finally, mQTL associated with MCA (factor 1) levels included CDK9, DIRC3, CTAGE1|RBBP8, and PCIF1.
We have previously shown that all three metabolite factors predict risk of incident CVD events, however the results from those studies were most robust for the SCDA metabolites [5]. Given these prior results, and the strength and consistency of findings for the SCDA metabolite factor in these GWAS analyses, we chose to focus the remainder of our analyses on this factor. Fig 2 and S5 Fig display Locus Zoom plots for these eight mQTL most strongly associated with SCDA metabolite factor levels. Interestingly, the majority of these (i.e. HERC1, USP3, STIM1, SUGT1, FBXO25 and SEL1L) encode proteins reporting on endoplasmic reticulum (ER) stress.
SCDA mQTL were tested for association with incident CVD events using Cox proportional hazards time-to-event analyses in the combined discovery and validation datasets, using meta-analysis of race- and dataset-stratified results, unadjusted for multiple comparisons. Of these eight mQTL (15 SNPs) loci, four SNPs predicted mortality in additive models: HERC1 rs2228513 (p = 0.05 in race combined, p = 0.04 in whites only), RRM1 rs11826962 (p = 0.03), and FBOX025 rs1869075 (p = 2.5x10-4 for blacks only, not significant in race combined analyses), with USP3 rs10450989 showing a trend for association (p = 0.06 in race combined, p = 0.05 in whites only). FREM2|STOML3 rs4544127 showed a trend for association (p = 0.06). We observed for the HERC1 SNP a 33% event rate for non-carriers and a 36% event rate for carriers of at least one copy of the minor G allele (the same allele associated with higher SCDA levels, S3 Table). Adjustment for SCDA levels in these models resulted in attenuation of the association between mQTL and CVD event (S3 Table), suggesting that the relationship between these mQTL and CVD events is in part mediated through SCDA metabolite levels.
To ensure that the relationships between SNPs and SCDA levels were not confounded by renal disease, we further adjusted for glomerular filtration rate. This adjustment caused no or minimal attenuation of the association for our strongest SNPs (S3 Table). In multivariable models, we found minimal attenuation of the association between most SNPs and SCDA levels (S3 Table), suggesting that these SNPs have effects on SCDA levels unrelated to other comorbidities. There was attenuation of association of SNPs near RRM1|STIM1 and STON2|SEL1L after adjustment (although still significant at p<0.05, unadjusted for multiple comparisons), suggesting that these SNPs have effects on SCDA levels mediated through these clinical factors, in particular renal disease.
Visual comparison of the distribution of methylated probes revealed similar distributions in individuals with high and low SCDA levels (N = 46, combined methylation discovery and validation datasets, S6 Fig). After filtering based on Δβ values, the presence of multiple correlated probes in a gene, and adjustment for estimated cell type proportions, sex, age and race, probes in 28 genes showed differential methylation in SCDA extremes (i.e. |Δβ|≥0.10 in ≥2 probes within a gene). Differential methylation in three of these genes was confirmed in the validation set based on |Δβ|≥0.10 (BRSK2, Hook2 and LMTK3, Table 4). Two of these genes, including the most significant one, report on ER stress: Hook2 (four probes, Δβ 0.25–0.30) and BRSK2 (four probes, Δβ 0.11–0.20). Hook2 may be involved in pathways contributing to the ubiquitin proteasome system (UPS) arm of ER stress via its role in establishment and maintenance of pericentrosomal localization of aggresomes (complexes of misfolded proteins, chaperones and proteasomes) [8]. BRSK2 encodes brain selective kinase 2, a serine/threonine kinase of the AMPK family that acts as a checkpoint kinase in response to DNA damage induced by UV irradiation. BRSK2 protein levels are down-regulated in response to ER stress and ER stress promotes localization of BRSK2 to the ER [9]. Knockdown of endogenous BRSK2 expression enhances ER stress-mediated apoptosis in human pancreatic carcinoma and HeLa cells [9].
Blood RNA microarray data were generated for N = 1204 CATHGEN individuals. We began by examining cis effects for the identified SNPs; however, many of the top SNPs did not have available cis-transcripts after extensive QC. Rs9591507,rs17573278, rs894840, and rs9285184 (all in OLFM4|SUGT1), rs11771619 (RSBN1L), rs1869075 (FBXO25), and rs1886848 (SULF2) showed evidence of cis-regulation (S4 Table). HERC1 and USP3 are not well-represented on the microarray (one probe per gene); there was only a minimal trend toward association between the HERC1 and USP3 SNPs with HERC1 expression (p = 0.16 and 0.19, respectively) and no association with the USP3 transcript.
We then performed eQTL analyses to find evidence of trans-acting pathways (S4 Table). When analyzed as single transcripts, among the top ten transcripts associated with HERC1 rs2228513 and USP3 rs10450989 were USP39 (p = 0.0002 and p = 0.0004, respectively) and CYLD (p = 0.00015 and p = 0.0007), suggesting that these SNPs show functional relationships with expression of trans-acting pathways related to the UPS arm of ER stress. USP39 has a role in pre-mRNA splicing and is essential for recruitment of the U4/U6.U5 tri-snRNP to the prespliceosome. The tumor suppressor CYLD is a deubiquitinating enzyme, acts as a negative regulator of NF-kappa-B signaling, and plays a pro-inflammatory role in vascular smooth muscle cells [10]. Cis- and trans-eQTL analyses were not adjusted for multiple comparisons, as we were looking for focused functional effects for each SNP.
Using GSEA [11], we then identified KEGG pathways of transcripts associated with each SNP; nominal p-values are reported. The most significant pathway associated with HERC1 rs2228513 was “ubiquitin mediated proteolysis” (p = 0.01; p = 0.12 for USP3 rs10450989). The most significant pathway for rs10450989 was “RNA degradation” (p = 0.03). Pathways associating with the other SNPs reported on various cellular processes: rs930491 and rs11827377 (RRM1|STIM1) with RNA polymerase pathway (both p = 0.001); rs11826962 (RRM1|STIM1) with JAK-STAT signaling pathway (p<0.0002); rs17573278 (OLFM4|SUGT1) with Alzheimer’s disease pathway (p = 0.008); rs894840 (OLFM4|SUGT1) with glycosaminoglycan biosynthesis (p<0.0002); rs12589750 and rs3853422 (STON2|SEL1L) with ribosome pathway (p<0.0001 and p = 0.001, respectively) and FC Gamma R mediated phagocytosis pathway (p = 0.001 for both). The Alzheimer’s disease pathway includes components of ER stress and there is evidence that neuronal death in Alzheimer’s disease may arise from ER dysfunction. The ER is also thought to play an important structural role in phagocytosis.
Finally, we performed GSEA for the correlation between SCDA levels with genomewide RNA expression; nominal p-values are reported. The most significant KEGG pathways were oxidative phosphorylation (p<0.0002), Parkinson’s disease (p<0.0002), cardiac muscle contraction (p<0.0002), porphyrin and chlorophyll metabolism (p = 0.002), and the proteasome pathway (p = 0.008). The proteasome is an integral component of the UPS arm of ER stress, degrading cellular proteins that are modified by ubiquitin. Also, an integral part of the Parkinson’s disease pathway includes components of the UPS.
In this and prior studies [4–6], SCDA were measured using a flow-injection-MS/MS method that is ideal for rapid profiling of samples, but full resolution of isomeric species comprising each SCDA metabolite peak is not achieved. C6-DC represents a SCDA that loads heavily on the PCA-derived SCDA factor in our studies, which can be comprised of either the branched-chain methylglutaryl acylcarnitine or the straight chain adipoyl acylcarnitine isomers. To resolve these metabolites, we adapted a liquid chromatography (LC)-MS/MS method [12]. Peak identification was facilitated by in-house chemical synthesis of internal standards for the two targeted analytes [13]. Using this method, we re-analyzed 29 human plasma samples from our original studies [5] that contained the highest C6-DC levels. We found that in the majority of individuals (19 of 29), the clearly predominant C6-DC isomer was the branched-chain 3-methylglutaryl carnitine metabolite, and in in 23 of the 29 individuals levels of the branched chain isomer were higher than the straight chain isomer (S7 Fig). The correlation between the C6-DC measured by flow injection-MS/MS with each of these LC-MS/MS measured isomers further confirms that it is primarily the branched-chain isomer accounting for the signal (r2 = -0.06, p = 0.8 for straight chain isomer; r2 = 0.67, p = 1.8x10-4 for branched-chain isomer). Interestingly, one potential source of the branched-chain 3-methylglutaryl carnitine metabolite is the branched-chain amino acid leucine. Our previous studies have shown an association of branched-chain amino acid metabolites with coronary artery disease [4, 7].
The above findings linking ER stress to SCDA metabolites led us to question whether nutrient-induced accumulation of dicarboxylacylcarnitines would be accompanied by ER stress in cultured cells. Exposure of human HEK293 kidney cells to 500 uM fatty acids for 24 hours (a condition designed to mimic elevated fatty acid levels observed in human obesity) increased cellular production and efflux of several long, medium and short-chain dicarboxylacylcarnitines (Fig 3A and 3B). Interestingly, fatty acid-induced production of dicarboxylacylcarnitines was accompanied by elevated expression of the molecular chaperone protein BiP (Fig 3C), a well-recognized marker of ER stress. At low doses of the ER stress agent tunicamycin (lower than required to cause cytotoxicity), fatty acid exposure also augmented BiP expression (Fig 3C). Together, these results point to an intriguing connection between cellular carbon load, dicarboxylic acylcarnitines and proteotoxicity.
We have analyzed metabolomics, genetics, epigenetics and transcriptomics together to establish genomewide associations between a cluster of SCDA metabolites that predict CVD events and specific genetic loci. Our findings implicate the UPS arm of ER stress as a factor influencing SCDA levels and CVD event pathogenesis. Several previous studies have successfully mapped metabolites to genetic loci [2], but primarily have not triangulated such genetic variation with disease endpoints and functional studies. Key findings of the current study include: (1) SNPs and CpG probes in genes reporting on components of ER stress were associated with levels of SCDA metabolites previously shown to predict CVD events [3–5]; (2) several of these SNPs themselves also predicted CVD events; (3) some of the SNPs/genes were linked with SCDA metabolites and ER stress through eQTL analyses; (4) the isomeric composition of the peak containing the major SCDA metabolite C6-DC was clarified; and (5) in cultured cells, nutrient-induced accumulation of SCDA metabolites occurred in parallel with increases in the ER stress marker BiP. Subjects in the CATHGEN cohort have a high prevalence of obesity, hyperlipidemia and diabetes (Table 1). Thus, our in vitro experiment may be viewed as a mimetic of the metabolic environment to which CATHGEN subjects are commonly exposed.
Our strongest finding was for two SNPs (HERC1 rs2228513 and USP3 rs10450989) that are in LD (r2 = 0.99) despite being separated by 104 kB. Rs2228513 is a missense variant (serine to phenylalanine) that is predicted to be “probably damaging” by PolyPhen, but no functional evaluation has been reported. Rs10450989 is an intronic SNP. The HERC gene family encodes a group of large proteins that contain multiple structural domains including a C-terminal HECT domain found in a number of E3 ubiquitin protein ligases. HERC1 is involved in membrane trafficking and may also act as an E3 ubiquitin-protein ligase, a protein that accepts ubiquitin from an E2 ubiquitin-conjugating enzyme and then directly transfers the ubiquitin to targeted substrates. Rs2228513 corresponds to residue 3152, which does not map to a specific domain in the protein. Our eQTL results suggest that this SNP is associated with differential expression of genes within a pathway reporting on the UPS. USP3 encodes ubiquitin-specific protease 3 which mediates release of ubiquitin from degraded proteins by disassembly of the polyubiquitin chains in the ER. Deubiquitination has been implicated in cell cycle regulation, proteasome-dependent protein degradation, and DNA repair [14]. Interestingly, an intergenic SNP 58 kB upstream from USP3 (rs10519210) was the strongest SNP associated with heart failure in a GWAS from the CHARGE consortium [15]. Rs10519210 not associated with SCDA levels in our study (p = 0.16) and is not in LD with rs10450989 (r2 = 0.002). Our next strongest finding was for a locus in/near STON2 and SEL1L. Rs12589750 is an intronic SNP within STON2 and rs3853422 is intergenic between STON2 and SEL1L. SEL1L plays a role in the ER-associated protein degradation (ERAD) machinery, and is part of a complex necessary for the retrotranslocation of misfolded proteins from the ER lumen to the cytosol where they are then degraded by the proteasome in a ubiquitin-dependent manner. Dysfunctional protein degradation causes ER stress.
Other mQTL included SNPs near RRM1 and STIM1; STIM1 encodes a calcium sensor in the ER that translocates to the plasma membrane upon calcium store depletion to activate calcium release-activated calcium channels. STIM1 induction, redistribution and clustering are important during ER stress when calcium stores are depleted [16]. FBXO25 is one of 68 human F-box proteins that serve as specificity factors for a complex composed of s-phase-kinase associated protein 1 (Skp1) and cullin1 (SCF), that act as protein-ubiquitin ligases, targeting proteins for destruction across the UPS. FBXO25 is cardiac specific and acts as a ubiquitin E3 ligase for cardiac transcription factors [17]. Rs17573278 and rs9591507 are intergenic SNPs >400 kB downstream from OLFM4 and SUGT1. SUGT1 is required cell cycle transitions and encodes a novel subunit of the SCF ubiquitin ligase complex [18]. OLFM4 encodes an anti-apoptotic protein that promotes tumor growth. The functions of the other SCDA mQTL loci are unclear.
Given the strength of association of SCDA metabolites (factor 3) with CVD and their particular strength of association in the current GWAS analyses, we chose to focus our subsequent analyses on SCDA. However, we did also identify mQTL for LCDA and MCA, both of which have also been shown to predict CVD events. LCDA are metabolic intermediates of long chain fatty acid oxidation in the mitochondria or peroxisomes. The most significant mQTL for LCDA metabolite levels included PIGR, USH2a, ZNF521 and FZD3. PIGR is a member of the immunoglobulin superfamily and ZNF521 is involved in regulation of early B-cell factor, suggesting a potential relationship between LCDA levels and immune and/or inflammatory pathways as a link to CVD. MCA are byproducts of mitochondrial fatty acid oxidation. The most significant mQTL for MCA show no obvious potential biologic relationship to mitochondrial function and/or CVD. More epidemiologic and functional work is necessary to clarify these links.
Importantly, and unique to this study, we have observed an association of mQTL and disease phenotypes. The SNPs most significantly associated with SCDA levels (HERC1 and USP3) were also associated with CVD events, with a consistent direction of effect (G allele associating with higher SCDA levels and events). STIM1|SEL1L SNPS were not associated with CVD events despite their strong association with SCDA levels; this may be due to limited power related to the low MAF in racial subsets. Adjustment for SCDA levels in these models resulted in attenuation of the association between SNP and CVD event suggesting that the relationship between underlying mQTL and CVD events is in part or in full mediated through SCDA metabolites and not through a different biological pathway. In combination, these results suggest potential functional and pathway relationships between SCDA metabolites and CVD events.
We also integrated transcriptomics and whole genome methylation with SNP and metabolomic data sets. eQTL identified ER stress pathways, and specifically those reporting on the ubiquitin proteasome pathway, as associated with the SNPs linked to SCDA via GWAS, and with SCDA metabolites themselves. Whole genome methylation identified epigenetic regulation of genes in ER stress pathways to be associated with extreme SCDA levels. These results support the concept that these polymorphisms and ER stress underlie the relation between SCDA metabolites and CVD events. Finally, we clarified the biochemical structure of the metabolite most strongly accounting for the C6-DC SCDA peak; these results will enable more accurate identification of the source pathways for C6-DC and other SCDA in future studies.
Many SCDAs result from the catabolism of amino acids, ω-oxidation of fatty acids or perhaps represent products of microbial metabolism [19], but the reasons for their accumulation in plasma in at-risk subjects, and how they may be related to CVD pathogenesis remain uncertain. Based on the convergence of GWAS, transcriptomic, metabolomic and functional data presented herein, we hypothesize that genetic and epigenetic variation predisposes to increased susceptibility to ER stress through proteasome dysfunction (reflected by the observation of upregulation of expression of ER stress genes), with ER stress in turn contributing to increased production of SCDA metabolites. This pathway of increased ER stress then leads to increased risk of CVD events, with SCDA metabolites and the genetic variants themselves predicting increased risk by reporting on this pathway (Fig 4). Epigenetic variation could be the influence of environmental or lifestyle factors inducing methylation changes; in this working model, diet and lifestyle-induced dyslipidemia and hyperglycemia could result in methylation changes as a regulatory mechanism to handle nutrient overload, thus predisposing to dysregulated ER stress which then leads to subsequent CVD events.
The UPS arm of the ER is responsible for the removal of misfolded proteins but is sometimes insufficient, for example, in the setting of increased production of misfolded proteins. The associated proteasome functional insufficiency can lead to cellular dysfunction and cell death, with cardiomyocytes being particularly vulnerable due to limited regenerative capability [20]. The UPS has been hypothesized to be involved in atherosclerosis based on the recognized roles of inflammation, oxidative stress, and endothelial dysfunction in this condition, and the intertwined relationships between the UPS and those pathways [21]. Preclinical evidence of the role of the UPS in atherosclerosis includes studies showing that oxidized LDL inhibits proteasomal activity in macrophages leading to apoptosis [22], and data suggesting that the UPS may contribute to foam cell formation by suppression of apoptosis of lipid-bearing macrophages by aggregated LDL in in vitro models [23]. Studies of proteasome inhibition have shown conflicting data; Hermann et al. found aggravation of atherosclerosis [24] and myocardial dysfunction [25] in pigs treated with proteasome inhibition, whereas a recent study showed reversal of uremia-induced atherosclerosis with proteasome inhibition in rabbits [26].
Human studies suggesting the role of the UPS in atherosclerosis are limited. Very small studies have shown greater amounts of ubiquitin conjugates in carotid endarterectomy tissues with unstable as compared with stable plaque morphologies [27] and increased UPS activity in carotid tissue from patients with symptomatic compared with asymptomatic carotid disease [28]. While preclinical studies have suggested the role of UPS in atherosclerosis as secondary to oxidative stress or other pathophysiologies, our identification of genetic variants in UPS/ER stress genes using unbiased analyses in our human cohorts provides strong support for the direct etiologic role of the UPS in promoting long-term cardiovascular risk. Importantly, we note that while ER stress is a common pathway in several disorders, we believe that the convergence of results on the UPS highlights its unique relationship to SCDA metabolism.
Our findings could have significant translational implications beyond CVD. Our primary objective of discovery of novel genetic risk variants using an mQTL approach was successful; the unexpected finding of genetic variation predisposing to ER stress could have much broader importance to human disease. Indeed, the response to ER stress is a trait that is known to be heritable in humans [29], but the genetic architecture has not been characterized. Equally as important, our data suggest the presence of easily quantifiable circulating biomarkers of ER stress, traditionally measureable only in tissue through ER stress-responsive gene expression studies. Thus, these results could have more wide-reaching implications for ER stress research in humans. Our prior work solidified the role of SCDA metabolites as predictors of CVD events [4, 5]; the current study has implications for clinical translation using SCDA metabolites for improved risk stratification even beyond CVD given the central role of normal and dysfunctional ER stress in health and disease.
The strengths of this study are the use of a priori defined discovery and validation cohorts; integration of genetics, epigenetics, metabolomics, transcriptomics in large cohorts; and careful biochemical refinement of the most strongly associated SCDA metabolite. Importantly, this represents one of the first studies to successfully identify genetic variants through mapping of intermediate metabolomic traits that themselves associate with disease endpoints. Our prior work had consistently identified SCDA metabolites as incremental predictors of CVD events, but little was known about the biological pathways underlying that association; the genomewide, multiple platform molecular approach taken in our study facilitated identification of the UPS more rapidly than other scientific methods. This work also adds an important finding to the metabolomics literature, namely that SCDA metabolites may be reporting on increased or dysregulated ER stress and specifically to proteasome functional insufficiency or dysregulation.
There are limitations to the study; the study population was comprised of individuals referred due to a suspicion of CVD and thus represents a disease-prone population. However, we note that 44% of study participants did not have significant coronary artery disease, highlighting the importance of the detailed angiographic phenotype to ensure that coronary artery disease is not confounding the relationship between genetic factors and outcome. Further, the high burden of CVD risk factors mirrors that of the general population, enabling generalizability of the study findings. Some of the results were isolated to a racial subset because the identified SNPs were either monomorphic or extremely rare in other races, underscoring the potential importance of including non-Caucasian races in such studies. Race-stratified sequencing of these genomic regions may identify different variants in these genes present in other races that may also serve as SCDA and CVD genetic variants. We a priori chose a p-value ≤10−6 as genomewide significant based on the commonly used threshold at the time we embarked on this study, and as a balance between the overly conservative Bonferroni correction and presence of linkage disequilibrium across the genome. More contemporary GWAS platforms cover a greater number of SNPs and include imputed SNPs in analysis, thus p<10−8 is now often considered genomewide significant; most of the key SNPs in this study would meet that threshold in combined meta-analyses, but not in the discovery cohort alone. The significance level also did not account for testing of two genetic models and for race-stratified analyses, however, most of the identified mQTLs would remain significant even after accounting for such multiple testing (p<3.0x10-7). More importantly, the use of a validation cohort and convergence of diverse omic’ data on the UPS obviate concerns about type I error with the threshold used for this study. Finally, while our study overall analyzed metabolomics with genetics, epigenetics and transcriptomics, not all individuals were profiled with all platforms, such that we co-analyzed genetic, epigenetic and transcriptomic data with metabolomics data one pair at a time. The ultimate goal for an eventual true systems biology approach would integrate all molecular platforms to unravel molecular pathways. However, to our knowledge this is the largest study deploying four diverse platforms in conjunction with cardiovascular event outcomes to date, and our consistent findings across platforms support further mechanistic interrogation of the identified pathway.
Our results highlight the power of combined molecular analyses and mapping of intermediate disease-related biomarkers for identifying the genetic architecture underlying common complex diseases, and could lead to improved CVD event risk prediction models as well as further mechanistic investigations of the role of the ubiquitin proteasome system in CVD.
The overall objective of this study was to integrate metabolomic, genetic (genomewide association study [GWAS]), transcriptomics and epigenetic data in a large human cohort to identify the genetic architecture regulating metabolite levels (metabolites shown to be incrementally predictive of CVD events [4, 5]) and thereby identify novel CVD risk genes. The analytic process was as follows (S1 Fig): (1) a GWAS was conducted of metabolite factor levels in a discovery cohort (N = 1490) individuals from the CATHGEN biorepository; (2) SNPs meeting genomewide significance from the discovery cohort were validated in a second cohort (N = 2022) CATHGEN subjects; (3) to identify potential epigenetic variation regulating SCDA metabolite levels (factor 3), analyses of whole genome methylation profiling of CATHGEN individuals with extremes of SCDA metabolite levels was performed (N = 46); (4) to elucidate potential downstream biological pathways, these validated GWAS SNPs were then tested for association using genomewide transcriptomic data (i.e., eQTL, N = 1204 CATHGEN individuals); similar analyses were conducted using SCDA metabolite levels and transcriptomic data. These analyses identified the UPS arm of ER stress and functional in vitro studies of that pathway were then conducted.
Individuals were selected from the CATHGEN biorepository of patients referred for evaluation of ischemic heart disease recruited sequentially through the cardiac catheterization laboratories at Duke University (Durham, NC) [30]. After informed consent, blood was obtained from the femoral artery, immediately processed to separate plasma, and frozen at -80°C. Individuals were fasting for a minimum of six hours prior to collection. Patients with severe pulmonary hypertension or transplant were excluded. The discovery cohort for mQTL GWAS analysis of metabolite levels consisted of a coronary artery disease (CAD) case-control sample; CAD cases were defined as having one to three coronary arteries with clinically significant stenosis (i.e. >50%). Controls were defined as not having clinically significant CAD (i.e. zero coronary arteries with >50% stenosis) and being free of cardiovascular disease, peripheral vascular disease and with a normal ejection fraction (LVEF>40%), and were matched to cases on age, race and sex (745 cases and 745 matched controls). This CAD definition was also used as a covariable in multivariable models assessing the association between mQTL and metabolite levels. To ensure generalizability of the mQTL results, the validation cohort for the metabolite GWAS consisted of a sequential cohort of 2022 CATHGEN individuals [30], and was not constrained on CAD or other status. Significant mQTL were tested for association with incident CVD events (death at any time during follow-up). All CATHGEN participants provided informed, written consent for participation in the CATHGEN biorepository at the time of enrollment. The Duke Institutional Review Board (IRB) approved the CATHGEN biorepository and this substudy.
The Illumina Human Omni1-Quad Infinium Bead Chip (Illumina, San Diego, CA, USA) was used for genotyping in both the discovery and validation cohorts following the manufacturer’s protocol using 200 nanograms of DNA. Quantification of DNA samples prior to genotyping was performed using the Quant-iT PicoGreen dsDNA reagent in a 96-well plate format (Life Technologies, Grand Island, NY, USA). DNA quality was assessed using gel electrophoresis. All samples were scored on a zero to five scale and samples with a score <3 were not further used. Briefly, the samples were denatured and amplified overnight, followed by fragmentation, precipitation and resuspension. DNA was then hybridized to the Illumina BeadChip for 16–24 hours, washed to remove unhybridized DNA, and then labeled with nucleotides to extend the primers to the DNA sample. After the genotyping protocol, BeadChips were imaged using the Illumina iScan system. Genotypes were called using Illumina’s GenomeStudio V2010.2 software (version 1.7.4 Genotyping module). Any SNPs with <98% call frequency, minor allele frequency (MAF)<0.01 in all races, or out of Hardy-Weinberg equilibrium (p<10−6) were excluded, resulting in the following number of autosomal SNPs for analysis: 785,945 in whites; 881,891 in blacks; and 871,209 in the “other” race (primarily Native American). Samples with <98% call rates for all SNPs, gender mismatches, cryptic relatedness, or with outlying ethnicity (as determined by multidimensional scaling plots of a linkage disequilibrium-pruned set of SNPs) were excluded (172 samples).
Quantitative determination of levels of 63 metabolites (45 acylcarnitines, 15 amino acids, total ketones, β-hydroxybutyrate, and total non-esterified fatty acids [NEFA]) was performed in N = 3512 individuals from the CATHGEN study (N = 1490 for discovery cohort, N = 2022 for validation cohort), using methods as we have done previously [4–6]. Ketones (total and β-hydroxybutyrate) and NEFA were measured on a Beckman-Coulter DxC600 clinical chemistry analyzer, using reagents from Wako (Richmond, VA). For MS-profiled metabolites (acylcarnitines, amino acids), proteins were first removed by precipitation with methanol. Aliquoted supernatants were dried, and then esterified with hot, acidic methanol (acylcarnitines) or n-butanol (amino acids). Analysis was done using tandem flow injection MS with a Quattro Micro instrument (Waters Corporation, Milford, MA). Quantification of the “targeted” intermediary metabolites was facilitated by addition of mixtures of known quantities of stable-isotope internal standards. Given the use of internal standards permitting absolute quantification of the metabolites in micromolar concentrations, values below the lower limits of quantification (LOQ) were reported and analyzed as “0”. Metabolites with >25% of values below LOQ were not analyzed (two acylcarnitines: C6 and C7-DC).
RNA purification processing was done utilizing Qiagen PAXgene Blood RNA MDx Kits in frozen whole blood PAXgene tubes. Strict adherence to the PAXgene Blood RNA MDx Kit Handbook, Second Edition, July 2005 protocol was maintained throughout the purification process. The purification process failed on 384 samples (four batches of ninety-six samples each) during processing for unidentified reasons and the samples were not repeated. Biotinylated total RNA was generated using the Illumina TotalPrep RNA amplification kit (Life Technologies, Grand Island, NY, USA); 200 nanograms of RNA was used for the kit. The quality of the RNA was determined using the Bioanalyzer RNA Nano chip assay (Agilent, Santa Clara, CA, USA). Quantification of the RNA was determined using the Quant-iT RiboGreen RNA Assay Kit. Samples with RIN scores less than 6.0 were not carried forward. The Human HT-12v3 Expression BeadChip (Illumina, San Diego, CA) was used for quantitative RNA profiling and scanned on the Illumina iScan system according to manufacturer’s protocol. Biotinylated RNA (750 nanograms) was hybridized to the BeadChip and washed; Cy3-SA was then introduced to the hybridized samples and the BeadChips scanned on the Illumina iScan system according to manufacturer’s protocol. Quality control (QC) and background subtraction was performed using Illumina GenomeStudio tools. Probes with a detection p-value <0.05 and detected in >50% of samples were retained for analysis. Expression values were log2 transformed and quantile normalized using Robust Multichip Average (RMA) methods. Results were visually inspected for outliers and sample failures after plotting for variance components comprising eight distinct and standard QC variables at the plate, chip and individual level. A total of 12,800 probes passed the detection and QC filters and 1204 samples passed the QC and outlier filters.
Principal components analysis (PCA) with varimax rotation was used for data reduction of metabolomic data from the combined cohorts (S1 Table and S2 Table) using SAS v9.1 (Cary, NC). Factor 1 (composed of a cluster of medium-chain acylcarnitines [MCA]), factor 2 (composed of a cluster of long-chain dicarboxylacylcarnitines [LCDA]), and factor 3 (composed of a cluster of short-chain dicarboxylacylcarnitine [SCDA] metabolites [similar to our previous studies [4, 5]]), were used as the quantitative traits for GWAS. Eigenstrat was used to define principal components (PCs) in GWAS. Four eigenvectors were used as PCs in whites, two in blacks, and seven in the “other” race category. Race-stratified linear regression models for each SNP (additive and dominant), adjusted for age, sex, race-specific PCs and metabolite batch, were constructed using PLINK [31]. Race-stratified results were also combined with meta-analysis using METAL [32]. Genomic inflation factors (λ) were <1.0. Significant SNPs were defined as those showing genomewide significance (p<10−6) in the discovery cohort and nominal association (p<0.05, unadjusted for multiple comparisons) in the validation cohort. Significant SNPs were then: (1) analyzed using meta-analysis of the cohorts using METAL [32]; (2) tested for association with metabolite factor levels after adjustment for glomerular filtration rate and in multivariable models (adjusted for BMI, hypertension, CAD, diabetes, left ventricular ejection fraction, dyslipidemia, smoking and renal disease); and (3) tested for association with time-to-death using Cox-proportional hazards modeling in the combined cohorts. Expression quantitative trait loci (eQTL) analyses of SNPs and SCDA levels were conducted using linear regression adjusted for age, race, sex and batch. Gene Set Enrichment Analysis (GSEA) [11], using the Preranked tool, was used on the resultant p-values for each SNP or SCDA covariate effect on expression levels to identify enriched KEGG pathways. GWAS analyses were corrected for multiple comparisons based on the above defined genomewide significance; other analyses were not adjusted for multiple comparisons and nominal unadjusted p-values are reported, with a p≤0.05 considered statistically significant.
For the methylation studies, we analyzed blood samples from a discovery cohort composed of 11 individuals from the combined CATHGEN cohorts who had the highest SCDA factor levels and 12 individuals with the lowest levels; and a validation cohort of 12 individuals with the next highest SCDA factor levels and 11 individuals with the next lowest levels; all 46 individuals were selected from those with RNA expression microarray data also available. DNA was isolated from blood mononuclear cells and sodium bisulfite treated prior to being prepped for analysis on the Illumina HumanMethylation 450K BeadChip following the manufacturer’s guidelines, using the Zymo EZ DNA Methylation Kit using manufacture’s protocol (Zymo Research Corporation Irvine, California USA). The alternative incubation condition recommended if using the Illumina Infinium Methylation Assay was used (supplied in the manufacturer’s instruction manual appendix). Converted DNA was amplified, fragmented and hybridized to the Human Methylation27, RevB bead chip pool of allele-differentiating oligonucleotides.
We removed probes with detection p-value>.05 in >10% of samples, data based on fewer than three beads, and probes previously identified as cross-reactive with other genomic locations [33]. Samples were checked for gender mismatch using principal components analysis (PCA) of probes on chromosome X and assay controls were inspected to ensure good performance on all samples. After QC, the original group of 485K probes was reduced to 473K probes. Color bias correction and background adjustment were performed using lumi [34], followed by quantile normalization of methylated, unmethylated, type I and type II probes separately using wateRmelon [35]. Finally, we used Beta Mixture Quantile dilation (BMIQ) for intra-array normalization [36]. After preprocessing, overall methylation levels (β) were calculated as the ratio of methylated to total signal (i.e. β = M / (M + U)) where M is the methylated signal intensity for a probe, U is the unmethylated signal intensity, and β therefore ranges from 0 (unmethylated) to 1 (methylated). Δβ was calculated as the mean methylation difference between the high and low SCDA groups at each probe. To identify candidate regions of interest, we prioritized probes with |Δβ|>0.10 in the discovery set (N = 1287). After removing probes with a common SNP (MAF>.01) in the CpG or single-base extension site, we filtered to known genes containing at least two probes each with |Δβ|>0.10 within a 1 kB region (n = 97 probes in 28 genes). Finally, we restricted our probes with probes with |Δβ|>0.10 and the same direction of effect in both datasets (i.e. hypermethylation in high SCDA samples versus low, three genes). Although our primary criteria for follow-up were Δβ values and the presence of multiple correlated probes in a gene, we also tested for differential methylation using linear models and empirical Bayes methods as implemented in limma [37]. Our standard model adjusted for estimates of cell-type proportions present in each sample using the method of Houseman, et al. [38]; we also ran a sensitivity analysis that additionally included age, sex and race.
Adipoyl carnitine and 3-methylglutaryl carnitine were synthesized from carnitine chloride and the corresponding cyclic acid anhydride according to the method of Johnson [13]. Products were confirmed by mass spectrometry. The liquid chromatography (LC)-MS/MS method of Maeda et al. [12] was extensively modified. Acylcarnitines were derivatized to butyl esters. The analytical platform was converted to a UPLC format using an Acquity UPLC HSS T3 column and the ion pairing reagent was changed to triethyl ammonium acetate. The carnitines were eluted using a linear gradient using water as solvent A and 95/5 v/v acetonitrile/water as solvent B starting at 20% B.
|
10.1371/journal.pgen.1008061 | Asymmetrical localization of Nup107-160 subcomplex components within the nuclear pore complex in fission yeast | The nuclear pore complex (NPC) forms a gateway for nucleocytoplasmic transport. The outer ring protein complex of the NPC (the Nup107-160 subcomplex in humans) is a key component for building the NPC. Nup107-160 subcomplexes are believed to be symmetrically localized on the nuclear and cytoplasmic sides of the NPC. However, in S. pombe immunoelectron and fluorescence microscopic analyses revealed that the homologous components of the human Nup107-160 subcomplex had an asymmetrical localization: constituent proteins spNup132 and spNup107 were present only on the nuclear side (designated the spNup132 subcomplex), while spNup131, spNup120, spNup85, spNup96, spNup37, spEly5 and spSeh1 were localized only on the cytoplasmic side (designated the spNup120 subcomplex), suggesting the complex was split into two pieces at the interface between spNup96 and spNup107. This contrasts with the symmetrical localization reported in other organisms. Fusion of spNup96 (cytoplasmic localization) with spNup107 (nuclear localization) caused cytoplasmic relocalization of spNup107. In this strain, half of the spNup132 proteins, which interact with spNup107, changed their localization to the cytoplasmic side of the NPC, leading to defects in mitotic and meiotic progression similar to an spNup132 deletion strain. These observations suggest the asymmetrical localization of the outer ring spNup132 and spNup120 subcomplexes of the NPC is necessary for normal cell cycle progression in fission yeast.
| The nuclear pore complexes (NPCs) form gateways to transport intracellular molecules between the nucleus and the cytoplasm across the nuclear envelope. The Nup107-160 subcomplex, that forms nuclear and cytoplasmic outer rings, is a key complex responsible for building the NPC by symmetrical localization on the nuclear and cytoplasmic sides of the nuclear pore. This structural characteristic was found in various organisms including humans and budding yeasts, and therefore believed to be common among “all” eukaryotes. However, in this paper, we revealed an asymmetrical localization of the homologous components of the human Nup107-160 subcomplex in fission yeast by immunoelectron and fluorescence microscopic analyses: in this organism, the Nup107-160 subcomplex is split into two pieces, and each of the split pieces is differentially distributed to the nuclear and cytoplasmic side of the NPC: one piece is only in the nuclear side while the other piece is only in the cytoplasmic side. This contrasts with the symmetrical localization reported in other organisms. In addition, we confirmed that the asymmetrical configuration of the outer ring structure is necessary for cell cycle progression in fission yeast. This study provides notions of diverse structures and functions of NPCs evolved in eukaryotes.
| In eukaryotes, the nuclear envelope (NE) separates the nucleus from the cytoplasm. Molecular transport between the nucleus and cytoplasm across the NE occurs through nuclear pore complexes (NPCs). These complexes are cylindrical, eight-fold symmetrical structures that perforate the NE and are made of multiple sets of about 30 different protein species known as nucleoporins (Nups) [1–3]. Nups are classified into three groups: transmembrane Nups, FG repeat Nups, and scaffold Nups. Transmembrane Nups have transmembrane helices and anchor NPCs to the NE. FG repeat Nups contain phenylalanine-glycine (FG) rich repeats and are involved in molecular transport through the NPC cylinder structure. Scaffold Nups form two inner rings and two outer rings, which serve as the NPC structural core [4–7], and associate with the membrane through interactions with transmembrane Nups [3,8,9]. These NPC structures and most Nups are generally conserved among eukaryotes [1,2,10–13], although numerous species-dependent differences are found [14].
The Nup107-160 subcomplex is a key component of the outer rings and in most eukaryotes is composed of equal numbers of Nup107, Nup85, Nup96, Nup160, Nup133, Sec13, and Seh1, and depending on the species Nup37, Nup43, and ELYS are also included [15–20]. These Nups assemble to form, both in vitro and in vivo, the Y-shaped Nup107-160 subcomplex in Homo sapiens, the budding yeast Saccharomyces cerevisiae, and the thermophile Chaetomium thermophilum [5,20–25]. Nup85, Nup43, and Seh1 form one of the two short arms, while Nup160, Nup37, and ELYS form the other. The two arms are connected to Nup96 and Sec13, and Nup96 is connected to Nup107 and Nup133 to form the long stem (Nup96-Nup107-Nup133) of the Y-shaped molecule. Multiple copies of the Nup107-160 subcomplex form the outer rings on the nucleoplasmic and cytoplasmic sides of the NPC [4,5,26].
Like other eukaryotes, the fission yeast Schizosaccharomyces pombe has a set of conserved Nups [27–29] (hereafter, we use ‘sp’ to denote S. pombe proteins and 'sc' and 'hs' to indicate S. cerevisiae and H. sapiens proteins) (S1 Table). However, the S. pombe NPC has several unique features. It contains spEly5 (a potential homolog of metazoan ELYS) and spNup37, but lacks Sec13 and Nup43 [29], and it has two redundant Nup133 homologs (spNup131 and spNup132) in addition to two redundant scNic96/hsNup93 homologs (spNpp106 and spNup97). Most strikingly, the Nup107-160 subcomplex in S. pombe is composed of unequal numbers of spNup107, spNup120 (a homolog of hsNup160), spNup85, spNup96, spNup37, spEly5, spSeh1, spNup131 and spNup132 [29]; the relative numbers of spNup107and spNup131 are 4–8 in a single NPC, and that of spNup132 is approximately 48 whereas the other components are approximately16. Thus, a unique/different structural organization of the Nup107-160 subcomplex is suggested in S. pombe compared with S. cerevisiae and humans [29].
spNup131 (spNup133a) and spNup132 (spNup133b) have similar molecular structure and both are able to bind to spNup107 [27]. Despite their similar biochemical features, spNup131 and spNup132 are likely to have different functions because gene disruption strains show different phenotypes. The strain lacking spNup132 (nup132Δ) displays altered NPC distribution [27]; its growth is inhibited in the presence of thiabendazole (TBZ) (a microtubule-depolymerizing drug) or hydroxyurea (a DNA replication inhibitor) [28,29]; and it exhibits delayed chromosome segregation in meiosis and unusual spore formation [29,30]. The strain lacking spNup131 (nup131Δ) doesn't display any of these characteristics. In addition, telomere elongation and deficiency in SUMOylation have been reported in nup132Δ-specific phenotypes [31,32]. The causes of these functional differences remain unknown.
In the present study, we examined the positioning of each Nup within the NPC in S. pombe using immunoelectron microscopy and high-precision distance measurements using fluorescence microscopy and found asymmetrical positionings of the outer ring complex components in the nuclear and cytoplasmic sides of the NPC. In addition, genetical alteration of positioning of the key molecule Nup132 of the outer ring subcomplex resulted in the defects observed in nup132Δ.
The molecular architectures of spNup131 and spNup132 are similar to that found in other species, with an N-terminal β-propeller domain followed by a C-terminal α-helix stack domain [33] (S1A Fig). However, phylogenetic analysis indicates that spNup131 and spNup132 belong to evolutionarily distant clades, with the spNup131-containing clade branching from a common ancestor of yeast Nup133 homologs much earlier than that of spNup132 (S1B Fig). This suggests that despite the similarity in their domain architecture, spNup131 and spNup132 may have structural features that are distinct enough to confer different functions.
To understand the differences between spNup131 and spNup132, we examined the positioning of those Nups within the NPC by immunoelectron microscopy (IEM). The spNup131 or spNup132 gene was replaced with the respective gene N-terminally fused to GFP (GFP-spNup131 or GFP-spNup132). IEM was carried out using a specific antibody against GFP. The results showed that GFP-spNup131 is located at the cytoplasmic side of the NPC, while GFP-spNup132 is located at the nuclear side (Fig 1A). To confirm the accessibility of the nucleus to immunogold particles using this method, the nuclear centromere protein spMis6 [34] was co-stained (S2A–S2C Fig); only the cells positive for spMis6 were evaluated for staining of spNup131(S2B and S2C Fig). A montage picture of spNup131 with quantification shows that spNup131 was exclusively located in the cytoplasmic side of the NPC (left panels of Fig 1B). In contrast, the localization of spNup132 was exclusively in the nuclear side (right panels of Fig 1B), indicating that spNup131 and spNup132 have distinct localizations. To resolve potential artifacts of GFP-tagging at the N-terminus, we repeated these experiments using strains in which spNup131 and spNup132 were C-terminally fused to GFP. We obtained essentially the same results with N- and C-tagged proteins (Fig 1C), suggesting that spNup131 and spNup132 are differentially positioned at the cytoplasmic and nuclear sides, respectively, of the NPC.
To confirm the different localization of spNup131 and spNup132 in living cells by fluorescence microscopy (FM), we observed cells simultaneously expressing spNup131 and spNup132 fused with GFP and mCherry, respectively; we also tested cells expressing spNup131 and spNup132 fused with mCherry and GFP, respectively (Fig 1D). To determine their localizations within nuclear pores with high-precision, we applied an open-source program “Chromagnon” [35] (This software is available at https://github.com/macronucleus/Chromagnon [36]). After the chromatic correction, the average distance of spNup131 from spNup132 in each nucleus along its radial direction was measured (see Materials and Methods for details). As a result, the position of mCherry-spNup132 relative to GFP-spNup131 was -49.3 ± 1.8 nm (mean ± SEM), and the position of mCherry-spNup131 relative to GFP-spNup132 was 32.7 ± 1.4 nm, both indicating that spNup131 was located at the exterior position (distant from the nuclear center) compared with the location of spNup132 within the NPC (Fig 1E). These results support the interpretation of the IEM observation that spNup131and spNup132 are separately localized in the NPC: spNup131 is positioned at the cytoplasmic side and spNup132 is positioned at the nucleoplasmic side of the NPC.
To determine the function of spNup131 and spNup132, we identified its interacting proteins using affinity capture/mass spectrometry (S3A and S3B Fig). Several non-Nup proteins that interact with spNup131 and spNup132 were identified (S3C Fig).
Among the candidate proteins specifically interacting with spNup131, we selected spFar8 (also known as spCsc3) and examined its functional relationship with spNup131. spFar8 is one of the components of the striatin-interacting phosphatase and kinase (STRIPAK) complex [37,38] that regulates the functions of the spindle pole body (SPB; the yeast microtubule-organizing center) during mitosis [39]. GFP-fused spFar8 (spFar8-GFP) localized at the nuclear periphery during interphase, as previously reported [39] (see “wild type” in Fig 2A). IEM analysis revealed that spFar8-GFP localized at the cytoplasmic side of the nuclear pores (Fig 2B). This NPC localization of spFar8-GFP was greatly decreased in the nup131Δ cells (see “nup131Δ” in Fig 2A) but not in the nup132Δ cells (“nup132Δ” in Fig 2A). Previous studies have reported that in the nup132Δ background, NPCs cluster on the NE [27]. This NPC-clustering phenotype in the S. pombe nup132Δ cells is of low penetrance, when the cells are in exponential growth. In our experiment, NPCs did not cluster in the nup132Δ cells because the cells were growing exponentially. Western blotting analysis of the cell strains tested above showed no marked changes in the amount of spFar8 protein (Fig 2C). The mislocalization of spFar8 in the nup131Δ cells occurred when spNup132 was ectopically expressed, whereas normal localization was restored when spNup131 was expressed (Fig 2D and 2E). Localization of another STRIPAK complex protein, spFar11, to the nuclear periphery was also decreased in nup131Δ cells (S4 Fig). These results suggest that spNup131 plays a role in retaining the STRIPAK complex at the cytoplasmic side of the NPC in interphase cells. The genetic interaction of STRIPAK proteins with spNup131 shown by this experiment is consistent with their localization at the cytoplasmic side of the NPC indicated by IEM.
Among the candidate spNup132 interacting proteins identified by affinity capture/mass spectrometry (S3C Fig), we examined the functional relationship of spNup211 with spNup132. spNup211 is an S. pombe homolog of human Tpr and S. cerevisiae Mlp1 and Mlp2 [40]; Tpr homologs are known to localize on the nuclear side of the NPC [41,42]. IEM analysis of spNup211, whose C-terminal was tagged with GFP (spNup211-GFP), showed that spNup211 localized at the nuclear side of the nuclear pores as expected (Fig 3A). To examine its localization in relationship to spNup132, spNup211-GFP was observed by FM in wild type, nup131Δ, and nup132Δ cells. In wild type and nup131Δ cells, spNup211-GFP was localized at the nuclear periphery (see “wild type” and “nup131Δ” in Fig 3B). In contrast, in the nup132Δ cells, spNup211-GFP formed several bright foci at the nuclear periphery (“nup132Δ” in Fig 3B). We measured the maximum fluorescence intensity of spNup211-GFP in each nucleus in wild type, nup131Δ, and nup132Δ cells. The value was significantly higher in the nup132Δ cells (Fig 3C, left graph); the value was decreased when spNup132 was expressed (Fig 3C, left graph). On the other hand, spCut11-mCherry, as a control, was uniformly distributed at the nuclear periphery with similar maximum fluorescence intensities in all strains (Fig 3C, right graph). This result suggests that spNup132, but not spNup131, is the NPC component that contributes to spNup211 localization; however, we cannot exclude the contribution of additional factors. This fact is consistent with their localizations at the nucleoplasmic side of the NPC, and the lack of interaction between spNup211 and spNup131 is also consistent with their different localizations within the NPC.
Because the Nup133 homologs are integrated components of the Nup107-160 subcomplex, we next examined the positioning of the other Nup107-160 subcomplex components in S. pombe: spNup107 (scNup84/hsNup107), spNup120 (scNup120/hsNup160), spNup85 (scNup85/hsNup85), spNup96 (also known as spNup189C; scNup145C/hsNup96), spNup37 (hsNup37), spEly5 (hsELYS), and spSeh1 (scSeh1/hsSeh1) in cells expressing each of these Nups fused to GFP. A spNup98-spNup96 fusion protein is expressed as the nup189+ gene product, and spNup96 is generated by cleavage with the autopeptidase activity in the C-terminus of spNup98 [43].
IEM results showed that of the 7 Nups, spNup120, spNup85, spNup96, spNup37, spEly5, and spSeh1 were predominantly located on the cytoplasmic side of the NPC, whereas spNup107 was located on the nuclear side of the NPC (Fig 4A). The localization of spNup107 on the nuclear side was also confirmed using an N-terminus fusion protein (Fig 4A). Simultaneous detection of spMis6-GFP further confirmed the localization of spNup120, spNup85, spNup96, spNup37, spEly5 and spSeh1 on the cytoplasmic side of the NPC (Fig 4A). This result suggests that the Nup107-160 subcomplex is split into two pieces in S. pombe and that these two pieces are differentially located on the cytoplasmic and nuclear sides of the NPC. This result contrasts with the structure of the Nup107-160 complex reported in S. cerevisiae and humans, in which all of the components are connected to form the Y-shape structure. Thus, this result suggests an unexpected separation of the Nup107-160 subcomplex at the junction of spNup96 and spNup107 in S. pombe.
To confirm the different localizations of the Nup107-160 subcomplex components by FM, we examined the localization of these components in living cells. We chose spNup85 and spNup107 as representatives of cytoplasmic and nucleoplasmic components identified by IEM. We first determined the position of spNup85-GFP within NPCs by comparing it with that of mCherry-spNup131 and mCherry-spNup132 (Fig 4B). The position of mCherry-spNup131 relative to spNup85-GFP was 11.4 ± 1.9 nm (mean ± SEM) and the position of mCherry-spNup132 relative to spNup85-GFP was -26.2 ± 1.0 nm (Fig 4B; the relative position of Nup85 to Nup131 and Nup132 is indicated in the right panel). This result supports cytoplasmic positioning of spNup85 as indicated by IEM. Second, we determined the position of spNup107-GFP within NPCs using a similar method. The position of mCherry-spNup131 relative to spNup107-GFP was 34.0 ± 2.1 nm and the position of mCherry-spNup132 relative to spNup107-GFP was -4.4 ± 1.0 nm (Fig 4C; the relative position of Nup107 to Nup131 and Nup132 is indicated in the right panel). This result supports nuclear positioning of spNup107 as indicated by IEM.
Taken together, these results demonstrate that the Nup107-160 subcomplex components are differentially localized at the cytoplasmic and nucleoplasmic sides within the NPC in S. pombe (Fig 4D). Thus, we call hereafter the cytoplasmic components (spNup131, spNup120, spNup85, spNup96, spNup37, spEly5 and spSeh1) as the spNup120 subcomplex and the nucleoplasmic components (spNup132 and spNup107) as the spNup132 subcomplex.
Based on the localization analysis in this study and quantification of each Nup reported previously [29], we estimated the molecular weight of the cytoplasmic outer ring (composed of the spNup120 subcomplex) and nucleoplasmic outer ring (composed of the spNup132 subcomplex) as 7.4 MDa and 7.0 MDa, respectively: the cytoplasmic outer ring is estimated as 8-fold of a single unit (two each of spNup120, spNup85, spNup96, spEly5, and spNup37 and one each of spSeh1 and spNup131), and the nucleoplasmic outer ring is estimated as 8-fold of a single unit (six spNup132 and one spNup107). In this estimation, a single unit of both the cytoplasmic and nucleoplasmic outer rings consist of 7 α-solenoids and 6 β-propellers as elements, suggesting that similar ring structures may be consequently built in both sides. The values are a bit larger than the molecular weight (4.65 MDa) of one outer ring in S. cerevisiae, which consists of 8-fold of a single unit containing 5 α-solenoids and 4 β-propellers as elements [26].
To address the significance of the separated localization of the Nup107-160 subcomplex, we generated an S. pombe strain with spNup96 (one of the cytoplasmic outer ring Nup120 subcomplex components) artificially fused to spNup107 (one of the nuclear outer ring Nup132 subcomplex components) in a nup107Δ background (spNup96-spNup107-GFP). The strain was viable, and Western blot analysis confirmed expression of the fusion protein with the predicted molecular weight (Fig 5A). By IEM, the majority of the spNup96-spNup107-GFP fusion protein molecules were localized at the cytoplasmic side of the NPC (Fig 5B), showing a change in the location of spNup107 from the nuclear side to the cytoplasmic side. Under this condition, spNup132 was recruited to both the nuclear and cytoplasmic sides of the NPC (Fig 5C), suggesting that a significant fraction of the spNup132 molecules was recruited by spNup107. We assumed that the cytoplasmic recruitment of spNup132 molecules would result in a reduction of spNup132 at the nuclear side. It is known that the nup132Δ cells exhibit growth sensitivity to the microtubule-destabilizing drug, TBZ [28, 29], and also exhibit delayed meiotic division and abnormal spore formation [30]; however, the nup131Δ cells did not exhibit these phenotypes [29] (S6 Fig). Thus, we examined whether the spNup96-spNup107-GFP strain showed those deficiencies. We found that this strain exhibited growth sensitivity to TBZ (Fig 5D), delayed meiotic division (Fig 5E and 5F), and abnormal spore formation (Fig 5G), similarly to the defects found in nup132Δ cells. This result suggests that these defects are caused by reduction in the number of spNup132 located at the nuclear outer ring of the NPC, and also suggests that concomitant localization of spNup107 and a part of Nup132 with the spNup120 subcomplex at the cytoplasmic side of the NPC caused defects in normal progression of mitosis and meiosis. Alternatively, these defects might be caused by the altered dynamics of spNup96 and spNup107 at the NPC. Thus, split structure and asymmetrical localization of the Nup107-160 subcomplex are necessary for the normal progression of mitosis and meiosis in S. pombe.
Finally, we determined which domains of spNup131 and spNup132 were responsible for their different localizations. In this experiment, we expressed full length spNup131 or spNup132, fragments of spNup131 or spNup132, or chimeric proteins of spNup131 and spNup132 (Fig 6A). To exclude the possibility that endogenous spNup131 or spNup132 predominantly occupy preferred sites for spNup131 and spNup132 fragments or chimeric proteins, a nup131Δ nup132Δ double-deletion strain (lacking genes for both spNup131 and spNup132) was used. Protein expression was confirmed by Western blot analysis (S7A Fig). FM observation showed that GFP fused full-length spNup131 (spNup131FL) and spNup132 (spNup132FL) were localized at the nuclear periphery (Fig 6B). The N-terminal domains of both Nup131 and Nup132 (spNup131N and spNup132N) were not localized at the nuclear periphery, while the C-terminal domain of both proteins (spNup131C and spNup132C) were localized at the nuclear periphery, consistent with a previous report [44]. Interestingly, IEM analysis showed that spNup131C localized to the nuclear side of the NPC (Fig 6C) in contrast with the localization of full length spNup131 on the cytoplasmic side (compare the upper two panels in Fig 6C; also see Fig 1B). These results suggest that the C-terminal domains of both spNup131 and spNup132 have the ability to localize on the nuclear side of the NPC, probably through binding to Nup107, and also suggest that the N-terminal domains of spNup131 and spNup132 are involved in determining their differential localizations. To test this idea, we examined the localization of the chimeric proteins spNup131N+spNup132C and spNup132N+spNup131C: If the C-terminal domains were independent and functioning correctly, both chimeric proteins should be localized to the nuclear side of the NPC via interaction with Nup107, and if the N-terminal domains were independent and functioning correctly, the spNup131N+spNup132C protein should then be localized to the cytoplasmic side via Nup131N and the spNup132N+spNup131C protein should be retained on the nuclear side via Nup132N. The results showed that the spNup131N+spNup132C chimeric protein was predominantly diffused in the cytoplasm, with slight enrichment at the NE (Fig 6B) and that the other chimeric protein, spNup132N+spNup131C, also diffused to the cytoplasm but with no enrichment in the NE (Fig 6B). The relatively strong cytoplasmic staining of the chimeric proteins suggests that the N-terminal domains may function to prevent "abnormal" localization of spNup131 and spNup132. Overall, this result indicates that both N-terminal and C-terminal domains are required for the proper differential localization of spNup131 and spNup132. Consistent with this result, expression of these chimeric proteins failed to overcome the TBZ sensitivity in nup132Δ (S7B Fig), and to rescue the defects in meiosis in nup132Δ (S7C Fig).
Because the Nup107-160 subcomplex seems to have a unique organization in S. pombe, we wished to understand the overall structure of the NPC. For this purpose, we performed IEM of other Nups to reveal their positionings within the NPC. GFP-fused Nups were used and IEM was carried out using a specific antibody against GFP, unless specified otherwise (Fig 7). We first examined inner ring Nups known as the Nup93 subcomplex in vertebrates. spNup97 and spNpp106, redundant S. pombe homologs of scNic96/hsNup93, were both similarly positioned near the center of the NPC (Fig 7A); the redundancy of scNic96/hsNup93 homologs is unique to the Schizosaccharomyces genus. spNup184 (scNup188/hsNup188) and spNup186 (scNup192/hsNup205) were also positioned near the center (Fig 7A). spNup40 (scNup53/scNup59/hsNup35) and spNup155 (scNup157/scNup170/hsNup155) were also located near the center of the NPC, but they showed a slightly broader range of localization (Fig 7A).
The channel Nups spNup44 (scNup57/hsNup54) and spNup45 (scNup49/hsNup58) were positioned at the center of the pore (Fig 7B). spNup98 (scNup145n/hsNup98) was detected at the center of the pore, when using an anti-Nup98 antibody (clone 13C2) [45], which specifically recognizes GLFG repeats located at the N-terminal region of endogenous spNup98 (see “spNup98” in Fig 7B). On the other hand, spNup98 (scNup145n/hsNup98) that was fused with GFP at its C-terminus was detected on the cytoplasmic side of the nuclear pore, when using an anti-GFP antibody (see “spNup98-GFP” in Fig 7B). In H. sapiens and S. cerevisiae, the C-terminal region of the Nup98 homologs (hsNup98/scNup145n/scNup100/scNup116) interacts with Nup96 homologs (hsNup96/scNup145c) [46–49], and the C-terminal Nup98-APD (autoproteolytic and NPC-targeting domain) binds to Nup88 homologs (hsNup88/scNup82) and to Nup96 in a mutually exclusive manner [49]. Thus, the C-terminal region of spNup98 is located near the positions of spNup96 and spNup82, both of which are positioned at the cytoplasmic side of the NPC, while the N-terminal region is extended to the center of the pore. The S. pombe homolog of the conserved Nup Nsp1 (scNsp1/hsNup62) was localized frequently in the cytoplasmic side and infrequently in the nuclear side of the NPC (Fig 7B).
spNup82 (scNup82/hsNup88) and spNup146 (scNup159/hsNup214) were detected at the cytoplasmic side (Fig 7C) as expected from their homologous counterparts. spAmo1(scNup42/hsNlp1) was also detected at the cytoplasmic side (Fig 7C). The transmembrane Nups, spCut11 (scNdc1/hsNdc1), spPom152 (scPom152), and spPom34 (scPom34) were localized at the center of the pore and slightly biased toward the cytoplasm (Fig 7D). The conserved Nups, spNup60 (scNup60), spNup61 (scNup2/hsNup50), and spNup124 (scNup1/hsNup153) (Fig 7E) as well as spNup211(scMlp1/scMlp2/hsTpr) (Fig 3A) were detected at the nuclear side as expected from their homologous counterparts. We added a recently identified Nup spAlm1 [50] to the group of nuclear Nups, according to its nucleoplasmic localization (Fig 7E).
We summarized the positionings of the S. pombe Nups within the NPC in Fig 8. Transmembrane Nups (spCut11, spPom152, and spPom34); inner ring Nups (spNup97, spNpp106, spNup184, spNup186, spNup40, and spNup155); cytoplasmic ring Nups (spNup82, spNup146, spAmo1, spNsp1); central channel Nups (spNsp1, spNup98, spNup44, and spNup45); and nuclear basket Nups (spNup124, spNup60, spNup61, spNup211, and spAlm1) are positioned similarly to the positions of their orthologs in S. cerevisiae and human [1,4,5,26,51]. Thus, these NPC substructures, which include a central channel structure that is required for nucleocytoplasmic transport, seem to be common to other eukaryotes. In contrast, the outer ring Nup107-160 subcomplex (spNup131, spNup120, spNup85, spNup96, spNup37, spEly5 and spSeh1, spNup107 and spNup132) in S. pombe has a highly unusual asymmetrical organization: It splits into two pieces (spNup120 and spNup132 subcomplexes). The spNup120 subcomplex (spNup131, spNup120, spNup85, spNup96, spNup37, spEly5 and spSeh1) is located in the cytoplasmic side of the NPC while the spNup132 subcomplex (spNup107 and spNup132) is located in the nuclear side of the NPC. This asymmetrical organization of the Nup107-160 subcomplex may be required for building and maintaining the structural organization of the central channel and other NPC elements common to other eukaryotes, and consequently, the Nup107-160 subcomplex would be required for normal cell cycle progression in S. pombe.
IEM of S. pombe Nups suggest that the S. pombe-specific Nup107-160 subcomplex structure is uniquely split into two pieces that localize differentially to the cytoplasmic and nuclear sides of the NPC while preserving the conserved modular structures of the Nup107-160 subcomplexes of other eukaryotes. High-precision distance measurements using FM also support this result (Figs 1D, 1E, 4B, and 4C). The asymmetrical organization of the S. pombe Nup107-160 subcomplex contrasts with the localization of the Nup107-160 subcomplexes in H. sapiens, S. cerevisiae and Trypanosoma brucei, in which Nup107-160 complexes are found on both the cytoplasmic and nuclear sides of the NPC [1,5,13,26,52]. Our results suggest that the S. pombe NPC has a novel organization that has evolved in the Schizosaccharomyces genus, which commonly bears the Nup132 and Nup131 clades (S1B Fig).
In H. sapiens and S. cerevisiae, the Nups in the Nup107-160 subcomplex assemble to form Y-shaped structures [5,21,24,25,53]. A total of 32 Y-complexes form two concentric reticulated rings at both the nuclear and cytoplasmic sides of the human NPC [4,5]. This organization may be supported by the iso-stoichiometry of each Nup in the Nup107-160 subcomplex: the amounts of each Nup in the Nup107-160 subcomplex in human cells are nearly equal [54]. In contrast, in S. pombe, the amounts of each Nup in the Nup107-160 subcomplex are not equal but highly divergent, ranging from 1 copy to 5–6 copies/unit [29]. In addition, the Sec13 homolog in S. pombe does not localize in the NPC, and Nup43 is not conserved in the genome [29]. The different stoichiometry and composition of the S. pombe Nup107-160 subcomplex components are likely due to the unique fission yeast–specific separated structure of the Nup107-160 subcomplex revealed in this study.
The mechanisms underlying the difference in the positioning of spNup131 and spNup132 at the NPC are still uncharacterized. Affinity capture/mass spectrometry showed that both spNup131 (cytoplasmic side) and spNup132 (nuclear side) bind to spNup107 (nuclear side) and spNup85 (cytoplasmic side). Considering that the molecular structural features of the spNup131 and spNup132 proteins are similar to each other, it is possible that both Nup131 and Nup132 interact with spNup107 and Nup85 in vitro in the whole cell extract, where there is no steric hindrance. In contrast, under in vivo conditions, the steric hindrance of these proteins can be a more important factor in determining their position in the NPC structure. In fact, neither spNup131N+spNup132C nor spNup132N+spNup131C was properly positioned in the NPC (Fig 6B), supporting this notion. Although affinity capture/mass spectrometry showed the above-mentioned confusing results, this analysis also showed the interactor specific to each of spNup131 and spNup132. spNup146 (cytoplasmic side) appeared as a specific interactor to spNup131 (S3A Fig). On the other hand, spNup211 (nuclear side) appeared as a specific interactor to spNup132 (S3B Fig). These results suggest that spNup146 and spNup211 may be involved in the differential localization of spNup131 and spNup132, respectively. In fact, enforced relocalization of spNup107 to the cytoplasm side by expressing the spNup96-spNup107 fusion protein translocated half the fraction of spNup132 to the cytoplasmic side, while the other half remained in the nuclear side (Fig 5C), suggesting that at least one more factor, other than spNup107, is involved in the positioning of spNup132. spNup211 can be such a factor that positions spNup132 in the nuclear side of the NPC.
Two Nup133 homologs spNup131 and spNup132 play different roles with different positions at the NPC. spNup132 is required for normal kinetochore formation during meiosis in S. pombe [30]. Deletion of the nup132 gene but not the nup131 gene causes untimely kinetochore assembly and activates spindle assembly checkpoint machinery during the first meiotic chromosome segregation [30]. The deletion of the nup132 gene also increases the sensitivity to the microtubule-destabilizing drug TBZ, likely due to defects in the kinetochore structure in mitosis [29]. Similarly, in mammalian cells, kinetochore-related functions have been reported for the Nup107-160 subcomplex. A fraction of the Nup107-160 subcomplex is found at the kinetochores and spindle poles during mitosis [16,17,55,56], and depletion of the Nup107-160 subcomplex from kinetochores results in altered kinetochore protein recruitment [57–59]. While the molecular mechanism of underlying spNup132-mediated regulation of kinetochore proteins remains unknown, considering the similarity in kinetochore-related functions, spNup132, but not spNup131, is likely to be a functional homolog of mammalian Nup133.
On the other hand, spNup131 plays a different role at the cytoplasmic side of the NPC. IEM analysis in this study revealed that spNup131 is localized only on the cytoplasmic side of the NPC, and a genetic analysis showed interaction between spNup131 and spFar8. spFar8, which was proven to be located at the cytoplasmic side of the NPC in this study, is an S. pombe ortholog of the STRIPAK complex component Striatin [38]. In S. pombe, the STRIPAK complex regulates the septation initiation network through the conserved protein kinase Mob1 [37,60] and is required for asymmetric division of mother and daughter SPBs during mitosis [39]. The spNup131-dependent NPC localization of spFar8 revealed by this study implies that the NPC regulates STRIPAK localization in interphase cells. Human STRIPAK complexes have been proposed to play roles at the interface between the Golgi and the outer nuclear envelope [38]. Considering the physical and genetic interaction with spNup131, STRIPAK is likely to interact with the NPC on the cytoplasmic side of the NE. Although the role of the STRIPAK complex in interphase cells in S. pombe is not fully understood, the interaction between spNup131 and spFar8 may provide an important example linking the NPC to cytoplasmic structures.
This study demonstrates that the S. pombe Nup107-160 subcomplex has a novel separated structure and exhibits a localization pattern not reported in other organisms. Recent studies suggest that NPC structures are not necessarily the same among eukaryotes [26,61,62]. For example, the binucleated ciliate Tetrahymena thermophila, a single-cell organism, has two functionally distinct nuclei, the macronucleus (MAC) and micronucleus (MIC), in a single cell: The MAC and MIC differ in size, transcriptional activity, and nucleocytoplasmic transport specificity [63]. Interestingly, in T. thermophila, the MAC and the MIC NPCs are composed of differ amounts of the Nup107-160 subcomplex components [61]. The amount of the subcomplex in the MIC is about three times more than that in the MAC, suggesting that the Nup107-160 subcomplex forms different structures in the MAC from the MIC. In the green algae Chlamydomonas reinhardtii, a single cell organism, the NPCs have 24 Nup107-160 subcomplexes asymmetrically distributed within a single NPC: 16 at the nuclear side forming double outer rings and 8 at the cytoplasmic side forming one outer ring [62]. The asymmetrical distribution of the Nup107-160 subcomplex in C. reinhardtii may suggest a specific function of the complex in either cytoplasmic or nuclear side of the NPC although it is currently unclear. In some multicellular organisms, the expression level of Nups varies between cell types and during development [54,64–67]. It is also known that some mutations in Nups result in developmental defects in metazoans [68]. These findings suggest that the NPC composition in cell types and during development is biologically significant. Altered compositions might alter the NPC structure, at least in part; thus, unidentified NPC subcomplex structures may play roles in specific biological events such as differentiation and development. Thus, we speculate that novel NPC structures may be important to elucidate the biological functions of the NPC.
Strains used in this study are listed in S2 Table. YES or EMM2 culture medium was used for routine cultures [69]. For fluorescence microscopy, cells were grown in EMM2 liquid medium. ME medium was used to induce meiosis and spore formation. When necessary, TBZ (T5535-50G, Sigma-Aldrich Inc, Tokyo, Japan) was added to the YES medium to a final concentration of 10 μg/mL.
For immunoelectron microscopy, 1.5 × 108 cells expressing GFP-Nup131 were fixed in 1 mL of a mixture of 4% formaldehyde (18814–10, Polysciences, Inc, Warrington, PA, USA) and 0.01% glutaraldehyde (1909–10, Polysciences, Inc.) dissolved in 0.1 M phosphate buffer (PB) (pH 7.4) for 20 min at room temperature, treated with 0.5 mg/mL Zymolyase 100T (7665–55, Nacalai Tesque, Inc., Kyoto, Japan) in PB for 20–30 min at 30°C, and then permeabilized with 0.2% saponin (30502–42, Nacalai Tesque, Inc.) and 1% bovine serum albumin (BSA) in PB for 15 min. The GFP epitope tag was labeled with a primary antibody (600-401-215, rabbit polyclonal anti-GFP antibody, Rockland Immunochemicals, Limerick, PA, USA) diluted 1:400 in PB containing 1% BSA and 0.01% saponin, and a secondary antibody (7304, goat anti-rabbit Alexa 594 FluoroNanogold Fab’ fragment; Nanoprobes Inc., Yaphank, NY, USA) diluted 1:400. The same immunostaining conditions were used for the cells expressing each of the other GFP-fused Nups except for the dilution ratios of primary and secondary antibodies; the conditions used for each experiment are listed in S3 Table. For analysis of the spNup98 N-terminal region, we used a mouse monoclonal anti-Nup98 antibody (13C2) [43,45] (available from BioAcademia Inc, Japan, Cat. #70–345) diluted 1:100 and anti-mouse Alexa594 FluoroNanogold Fab’ fragment (7302, Nanoprobes) diluted 1:400. Cells then were fixed again with 1% glutaraldehyde in PB for 1 h at room temperature and treated with 100 mM lysine HCl in PB twice for 10 min each. The cells were stored at 4°C until use. Before use, the cells were incubated with 50 mM HEPES (pH 5.8) three times for 3 min each and with distilled water (DW) once, incubated with Silver enhancement reagent (a mixture of equal volumes of the following A, B, and C solutions: A, 0.2% silver acetate solution; B, 2.8% trisodium citrate-2H2O, 3% citric acid-H2O, and 0.5% hydroquinone; C, 300 mM HEPES, pH 8.2) at 25°C for 2–5 min. Cells were embedded in 2% low melting agarose dissolved in DW. Cells were post-fixed with 2% OsO4 in DW for 15 min and stained with 1% uranyl acetate in DW at room temperature. Cells were dehydrated using stepwise incubations in ethanol and acetone and finally embedded in epoxy resin Epon812. Solidified blocks containing cells were sectioned, and the ultra-thin sections were stained with uranyl acetate and lead citrate, the usual pretreatment for EM observations. Images were obtained using a JEM1400 transmission electron microscope (JEOL, Tokyo, Japan) at 120kV.
Nuclear pores containing more than two immunogold particles were chosen for localization analysis as described previously [1]. To confirm the accessibility of the nucleus to immunogold particles, the nuclear centromere protein spMis6-GFP was co-expressed with GFP-fused Nups in cells and stained with anti-GFP antibody for IEM. For quantification of the Nup signals, we chose only cell specimens with a positive spMis6-GFP signal. For quantitative representations, montage pictures were produced by stacking 20 NPC images with 5% opacity on Adobe Photoshop CS software.
We also quantified the distribution of GFP-fused proteins in the cytoplasmic, middle, and nucleoplasmic regions of the NPC by counting the number of gold particles in each region (S8 Fig).
nup107+ and nup37+ cDNAs were provided by National BioResource Project Japan (http://yeast.lab.nig.ac.jp/yeast/top.xhtml). cDNA fragments of other Nups were amplified from a cDNA library pTN-RC5 or pTN-FC9 (National BioResource Project Japan) using PCR.
To visualize nuclear pore localization of the different domains of Nup131 and Nup132, lys1+-integrating plasmids carrying GFPs65t-fused with respective Nup domains were introduced into cells with a nup131Δnup132Δ double-deletion mutant background. To construct the spNup96-spNup107 fusion Nup, a cDNA fragment encoding spNup107 or spNup107-GFP followed by a drug resistance marker gene was integrated after the chromosomal spNup96 coding region. After a diploid strain was obtained by crossing the spNup96-spNup107 fusion containing strain with the wild type strain, the nup107+ gene on the original chromosomal locus was deleted. To introduce the chromosomal fluorescent tag and gene disruption, a two-step PCR method was applied [70]. Nup-GFP fusion constructs were described previously [29]. The spMis6-GFP fusion was constructed as described previously [71] or using the two-step PCR method. mCherry-spAtb2 was visualized as described previously [30].
To express the full-length or domains of the spNup131 and spNup132 proteins, cDNA fragments were amplified by PCR. For overexpression, the PCR products were sub-cloned into the plasmid that carries the nmt1 promoter and terminator. For physiological expression, the PCR products were sub-cloned into the BglII site of the plasmid (pY53) that carries the nup132 promoter (-1000bp)-driven GFPs65t and the nmt1 terminator. Each sub-cloning was done by using the In-Fusion PCR cloning kit (Clontech Laboratories, Mountain View, CA, USA). Plasmids were integrated at the lys1 gene locus. Correct integrations were confirmed by PCR.
Images were obtained using a DeltaVision microscope system (GE Healthcare, Tokyo, Japan) equipped with a CoolSNAP HQ2 CCD camera (Photometrics, Tucson, AZ, USA) through an oil-immersion objective lens (PlanApoN60×OSC; NA, 1.4) (Olympus, Tokyo, Japan) as described previously [29]. Z-stack images were obtained and processed by deconvolution and denoising algorithm [72] when necessary. For time lapse microscopy, cells were observed every 5 minutes as described previously [30]. The projection images of z-stacks were made by softWoRx software equipped in the microscope system.
Cells to be analyzed were cultured in 150 μl of the EMM2 5S medium on an 8-well chambered coverglass, Lab-Tek II (Thermo Fisher Scientific Japan, Yokohama, Japan); cells used for references for chromatic correction were also cultured in a chamber of the same coverglass. Simultaneously acquired multicolor images were obtained with OMX SR (GE Healthcare) equipped with three sCMOS cameras. The three-dimensional (3D) images were deconvolved by constrained iterative deconvolution using the Priism suite (http://msg.ucsf.edu/IVE/) with a Wiener filter enhancement of 0.9 and 15 iterations. To correct for chromatic shifts of multicolor FM images, green-to-red photoconversion of GFP was used as the reference as previously reported [73,74]: Nup96-GFP was used in this study. For photoconversion, cells were illuminated with 405 nm light at about 393 W/cm2 for 4 seconds, then both green and red-converted GFP species were excited with 488 nm to obtain images of the same object in the green and red channels. Chromatic shifts were measured using such images, and the correction parameters were determined using the Chromagnon software (https://github.com/macronucleus/Chromagnon), by which a chromatic shift can be corrected with an accuracy of ~10 nm in 2D XY plane and ~15 nm in 3D XYZ space using test samples [35]. The correction parameters were applied to all images obtained from the same chambered coverglass as described in [35].
Next, individual nuclei in the images were identified and segmented by our software using functions from the “ndimage” module of the Scipy package (http://www.scipy.org). The centers of the nuclei were identified by least-square fitting with 3D ellipses using the 3D coordinates of the nucleoporin fluorescence above threshold. The images of the NE at the midsection were linearized with polar transformation. These images were then averaged across all angles to produce 1D intensity profiles along the radial direction of the nuclei. The intensity profiles were fitted to Gaussian profiles to determine a distance between the peaks of green and red for each of the individual nuclei. The mean distance was determined from more than 70 nuclei. Because not all nuclei were round, we rejected nuclei of elliptical or irregular shapes before the final calculation.
For western blot analysis, whole cell extracts were prepared from approximately 5×106 cells by trichloroacetic acid (TCA) precipitation as described previously [29]. To detect GFP-fused proteins, a rabbit polyclonal anti-GFP antibody (600-401-215, Rockland Immunochemicals Inc.) was used. To detect spNup98, a mouse monoclonal anti-Nup98 antibody (13C2) was used [45]. HRP-conjugated goat anti-rabbit or anti-mouse IgG (NA9340-1ml or NA9310-1ml, GE Healthcare) was used as a secondary antibody. Protein bands were detected by chemiluminescence using ChemiDoc MP imaging system (Bio-Rad).
Growing cells (about 5 × 109) were collected and washed with 10 mM HEPES buffer (pH 7.5) containing 1 mM phenylmethylsulfonyl fluoride (PMSF). The washed cells were divided into aliquots of 3 × 108 cells in 10 mM HEPES buffer (pH 7.5) containing 1 mM PMSF. Cells were again collected by centrifugation, and the cell pellet was kept frozen by liquid nitrogen until use. To make a cell extract, the cell pellet was thawed and suspended in 100 μL of lysis buffer (50 mM HEPES (pH 7.5), 150 mM NaCl, 1% Triton X-100, 1 mM EDTA, 2 mM PMSF) with a protease inhibitor cocktail (165–20281, Wako, Tokyo, Japan) and mashed by glass beads using Multi-beads shocker (Yasui Kikai Corporation, Osaka, Japan). We chose this extraction condition based on a previous study identifying the components of plant NPCs [11]. Because the presence of 1% Triton X-100 gave better results for extracting plant NPC components, we decided to use 1% Triton X-100 as a detergent, and we examined three different salt conditions (50 mM, 150 mM or 500 mM NaCl) in the presence of the detergent. As a result, 150 mM NaCl gave better results with low background. At least for extracting spNup131 and spNup132, this condition was better than that optimized for extracting NPCs in other organisms such as budding yeast (1.5 M ammonium acetate, 1% Triton X-100) [75] and trypanosoma (20 mM HEPES (pH7.4), 250 mM NaCitrate, 0.5% Triton X-100, 0.5% deoxy Big CHAP) [76]. After further addition of 400 μL of lysis buffer, the mashed cell pellet was transferred to new microtubes. The supernatant was collected after centrifugation at 15000 rpm for 15 min at 4°C and used as the whole-cell extract.
The whole-cell extract was incubated with a rabbit anti-GFP antibody (600-401-215, Rockland Immunochemicals). Antibody-conjugated proteins were collected by incubating with Protein A Sepharose beads (17528001, GE Healthcare). Beads were then washed 4–5 times with the lysis buffer described above. After elution in SDS-PAGE sample buffer, protein samples were loaded onto a 12% SDS-PAGE gel for liquid chromatography coupled to tandem MS (LC/MS/MS). Data analysis for LC/MS/MS was performed as described previously [77] using the Pombase protein dataset released on November 12th, 2015. To identify proteins interacting with spNup131 and spNup132, protein samples were prepared from two independent experiments and each preparation was analyzed by LC/MS/MS. Proteins detected as more than one unique spectrum were identified as interacting proteins with spNup131 and spNup132. Alternatively, when proteins detected as only one unique spectrum in the first experiment were repeatedly detected in the second experiment, they were also identified as interacting proteins.
|
10.1371/journal.ppat.1007771 | Blocking tombusvirus replication through the antiviral functions of DDX17-like RH30 DEAD-box helicase | Positive-stranded RNA viruses replicate inside cells and depend on many co-opted cellular factors to complete their infection cycles. To combat viruses, the hosts use conserved restriction factors, such as DEAD-box RNA helicases, which can function as viral RNA sensors or as effectors by blocking RNA virus replication. In this paper, we have established that the plant DDX17-like RH30 DEAD-box helicase conducts strong inhibitory function on tombusvirus replication when expressed in plants and yeast surrogate host. The helicase function of RH30 was required for restriction of tomato bushy stunt virus (TBSV) replication. Knock-down of RH30 levels in Nicotiana benthamiana led to increased TBSV accumulation and RH30 knockout lines of Arabidopsis supported higher level accumulation of turnip crinkle virus. We show that RH30 DEAD-box helicase interacts with p33 and p92pol replication proteins of TBSV, which facilitates targeting of RH30 from the nucleus to the large TBSV replication compartment consisting of aggregated peroxisomes. Enrichment of RH30 in the nucleus via fusion with a nuclear retention signal at the expense of the cytosolic pool of RH30 prevented the re-localization of RH30 into the replication compartment and canceled out the antiviral effect of RH30. In vitro replicase reconstitution assay was used to demonstrate that RH30 helicase blocks the assembly of viral replicase complex, the activation of the RNA-dependent RNA polymerase function of p92pol and binding of p33 replication protein to critical cis-acting element in the TBSV RNA. Altogether, these results firmly establish that the plant DDX17-like RH30 DEAD-box helicase is a potent, effector-type, restriction factor of tombusviruses and related viruses. The discovery of the antiviral role of RH30 DEAD-box helicase illustrates the likely ancient roles of RNA helicases in plant innate immunity.
| Positive-stranded RNA viruses are important and emerging pathogens that greatly depend on the host during infection. The host uses conserved innate and cell-intrinsic restriction factors as a first line of defense to combat viruses. Among the most intriguing host restriction factors are the family of DEAD-box RNA helicases, which can function as viral RNA sensors or directly as effectors by inhibiting RNA virus replication. RNA helicases are involved in cellular metabolism and perform RNA duplex unwinding and remodeling of RNA-protein complexes in cells. The authors demonstrate that the plant DDX17-like RH30 DEAD-box helicase acts as a strong restriction factor of tombusviruses by blocking multiple steps in the viral replication process. Overall, the findings presented open up a new avenue based on DEAD-box RNA helicases to improve the resistance of plants against viral infections.
| Positive-stranded (+)RNA viruses replicate inside cells and depend on many co-opted cellular factors to complete their infection cycle. These viruses build elaborate membranous viral replication compartments, consisting of viral replication proteins, viral RNAs and recruited host factors, in the cytosol of the infected cells. The hijacked host factors participate in all steps of RNA virus replication, including the assembly of membrane-bound viral replicase complexes (VRCs), viral RNA-dependent RNA polymerase (RdRp) activation and viral RNA synthesis. The growing list of co-opted host factors facilitating VRC assembly includes translation initiation and elongation factors, protein chaperones, RNA-modifying enzymes, SNARE and ESCRT proteins, actin network, and lipids [1–9]. Many (+)RNA viruses extensively rewire metabolic pathways, remodel subcellular membranes and take advantage of intracellular trafficking.
The host utilizes cellular proteins to sense viral pathogenicity factors and block virus replication with the help of cell-intrinsic restriction factors (CIRFs) as an early line of defense [2,10–12]. These CIRFs can be part of the innate immune responses and used for antiviral defense as sensors or effectors [13–16]. The identification and characterization of the many CIRFs against different viruses is still in the early stages.
Viral RNA replication is intensively studied with Tomato bushy stunt virus (TBSV), a tombusvirus infecting plants, based on yeast (Saccharomyces cerevisiae) surrogate host [17–19]. Expression of the two TBSV replication proteins, termed p33 and p92pol, and a replicon (rep)RNA leads to efficient viral replication. p92pol is the RdRp [20,21], whereas the more abundant p33 is an RNA chaperone. P33 functions in RNA template selection and recruitment and in the assembly of VRCs within the replication compartment [21–26].
TBSV, which does not code for its own helicase, usurps several yeast and plant ATP-dependent DEAD-box RNA helicases as host factors promoting TBSV RNA replication. The yeast DDX3-like Ded1p and the p68-like Dbp2p, and the plant DDX3-like RH20, DDX5-like RH5 and the eIF4AIII-like RH2 DEAD-box proteins were shown as pro-viral factors, which affect plus- and minus-strand synthesis, maintenance of viral genome integrity and RNA recombination in TBSV [27–29].
DEAD-box helicases are the largest family of RNA helicases and are known to be involved in cellular metabolism [30–32], and affect responses to abiotic stress and pathogen infections [33–35]. They function in unwinding of RNA duplexes, RNA folding, remodeling of RNA-protein complexes, and RNA clamping [36]. They have no unwinding polarity and can open up completely double-stranded RNA regions, however, unlike many other helicases, DEAD-box helicases do not unwind RNA duplexes based on translocation on the RNA strand. Instead, DEAD-box helicases directly load on duplexes and open up a limited number of base pairs. Strand separation within the duplexes is not coordinated with ATP hydrolysis, which is used for enzyme dissociation from the template. This unwinding mode is termed local strand separation [36,37]. DEAD-box helicases also affect RNA virus replication [38–41], and viral translation [42,43]. In case of plant viruses, turnip mosaic virus and brome mosaic virus have been described to co-opt cellular DEAD-box helicases for proviral function in translation or replication [39,42,44]. Altogether, cellular helicases are important co-opted host factors for several viruses, playing critical roles in virus-host interactions.
However, cellular RNA helicases also act as antiviral restriction factors, including functioning as viral RNA sensors (e.g., Dicer or RIG-I) or directly inhibiting RNA virus replication as effectors [45–47]. For example, DDX17 restricts Rift Valley fever virus [48], while DDX21 helicase inhibits influenza A virus and DDX3 blocks Dengue virus infections [49–52]. Thus, the emerging picture is that host helicases are important for the host to restrict RNA virus replication, but the mechanism of their activities or substrates are not well characterized.
In this work, we find that the plant DDX17-like RH30 DEAD-box helicase plays a strong restriction factor function against tombusviruses and related plant viruses. RH30 DEAD-box helicase is expressed in all plant organs, but its cellular function is not known yet [53]. We find that RH30 is re-localized from the nucleus to the sites of tombusvirus replication via interacting with the TBSV p33 and p92pol replication proteins. Several in vitro assays provide evidence that RH30 inhibits tombusvirus replication through blocking several steps in the replication process, including VRC assembly, viral RdRp activation and the specific interaction between p33 replication protein and the viral (+)RNA. RH30 knockout lines of Arabidopsis supported increased accumulation level for the related turnip crinkle virus, confirming the restriction factor function of RH30 against a group of plant viruses. This is the first identification and characterization of a plant helicase with an effector type restriction factor function against plant viruses. Since plant genomes codes for over 100 RNA helicases, it is likely that additional helicases have CIRF function against plant viruses.
To test if the host RH30 RNA helicase could affect tombusvirus replication, we expressed the Arabidopsis RH30 using agroinfiltration in Nicotiana benthamiana plants. Interestingly, expression of AtRH30 blocked TBSV replication by ~90% in the inoculated leaves (Fig 1A). The closely-related cucumber necrosis virus (CNV), which also targets the peroxisomal membranes for VRC formation, was also inhibited by ~4-fold through the expression of AtRH30 (Fig 1B). Replication of another tombusvirus, carnation Italian ringspot virus (CIRV), which builds the replication compartment using the outer membranes of mitochondria, was inhibited by ~9-fold by the transient expression of AtRH30 in N. benthamiana (Fig 1C).
To test if RH30 was also effective against TBSV when expressed in yeast cells, we launched the TBSV repRNA replication assay in wt yeast by co-expressing the viral components with RH30. After 24 h of incubation, TBSV repRNA analysis revealed strong inhibition of viral replication by RH30 expression (Fig 1F), suggesting that RH30 is a highly active inhibitor against TBSV replication even in a surrogate host.
To learn if the putative helicase function of RH30 is required for its cell intrinsic restriction factor (CIRF) function, we expressed a motif IV helicase core mutant of RH30(F416L) in N. bentamiana via agroinfiltration. Mutation of the highly conserved F residue within the helicase core domain (see S1 Fig) has been shown to greatly decrease both ATP binding/hydrolysis and strand displacement activities in Ded1 and other DEAD-box helicases [54]. Northern blot analysis revealed the lack of inhibition of TBSV replication, and only partial inhibition of CIRV replication by RH30(F416L) (Fig 1D and 1E, lanes 9–12). Thus, we suggest that the full helicase/ATPase function of RH30 is required for its CIRF function against tombusviruses.
VIGS-based silencing of the endogenous RH30 in N. benthamiana led to ~5-fold, ~3-fold and ~11-fold increased accumulation of TBSV, CNV and CIRV, respectively, in the inoculated leaves (Fig 2). The leaves of virus-infected and VIGS-treated plants showed severe necrotic symptoms earlier and died earlier than the control plants (i.e., TRV-cGFP treatment) in case of all three tombusvirus infections (Fig 2). On the other hand, the VIGS-treated plants became only slightly smaller than the TRV-cGFP treated control plants (Fig 2). Based on these and the RH30 over-expression data, RH30 DEAD-box helicase seems to act as a major restriction factor against tombusviruses in plants and yeast.
To identify the cellular compartment where RH30 DEAD-box helicase performs its CIRF function, first we used co-localization studies in N benthamiana protoplasts co-expressing GFP-RH30, p33-BFP (to mark the site of viral replication) and RFP-tagged H2B, which is a nuclear marker protein. We detected the re-localization of GFP-RH30 into the large p33 containing replication compartment from the nucleus during CNV replication (Fig 3A, top panel versus second panel). Both the p33-BFP and RFP-SKL (a peroxisomal matrix marker) showed the re-localization of GFP-RH30 into the large TBSV replication compartment, which consists of aggregated peroxisomes. Part of the ER is also recruited to the p33 and RH30 containing replication compartment (Fig 3A bottom panel), as shown previously [55,56].
Similar re-localization pattern of RH30 was observed in epidermal cells of whole plants infected with CNV (Fig 3B, top panel versus second panel). The expression of only p33-BFP was satisfactory to recruit the RH30 into the replication compartment (Fig 3B). RH30 was also re-targeted in CIRV-infected N. benthamiana cells into the p36 and p95pol containing replication compartment (Fig 3B, bottom panel), which consists of aggregated mitochondria [57,58]. Based on these experiments, we propose that the mostly nuclear localized RH30 helicase is capable of entering the tombusvirus replication compartment via interaction with the replication proteins. However, the formation of large tombusvirus-induced replication compartments seemed to be normal in the presence of RH30, indicating the lack of interference with the biogenesis of the replication compartment by RH30.
To test if the cytosolic localization of RH30 is required for its CIRF function, we fused RH30 with a nuclear retention signal (NRS) [59] to enrich RH30 in the nucleus at the expense of the cytosolic pool of RH30. Interestingly, unlike WT RH30, expression of NRS-RH30 did not result in inhibition of TBSV replication in N. benthamiana (Fig 4A). Confocal microscopy experiments confirmed that NRS-RH30-GFP is localized exclusively in the nucleus (Fig 4B). Infection of the N. benthamiana protoplasts with CNV did not result in the re-targeting of NRS-RH30-GFP from the nucleus to the replication compartment visualized via p33-BFP. The nuclear retention of NRS-RH30-GFP was also confirmed in N. benthamiana epidermal cells infected with CNV or mock inoculated (Fig 4C). Altogether, these experiments demonstrated that re-localization of RH30 helicase from the nucleus to the replication compartment is critical for its CIRF function in plants.
To learn about the tombusviral target of RH30 DEAD-box helicase, we co-expressed the His6-tagged RH30 with Flag-tagged p33 and Flag-p92pol replication proteins and the TBSV repRNA in yeast, followed by Flag-affinity purification of p33/p92pol from the detergent-solubilized membrane fraction of yeast, which is known to harbor the tombusvirus replicase [20,60]. Western blot analysis of the affinity-purified replicase revealed the effective co-purification of His6-RH30 (Fig 5A, lane 3), suggesting that RH30 targets the VRCs for its CIRF function. Interestingly, His6-RH30 was co-purified from yeast co-expressing either Flag-p33 or Flag-p92pol replication proteins (Fig 5A, lanes 1–2), suggesting that RH30 likely directly interacts with the tombusvirus replication proteins in a membranous compartment.
To show direct interaction between RH30 DEAD-box helicase and the TBSV p33 replication protein, we performed pull-down assay with MBP-tagged RH30 and GST-tagged p33 proteins from E. coli. We found that MBP-RH30 captured GST-p33 protein on the maltose-column (Fig 5B, lane 2), indicating direct interaction between the host RH30 and the viral p33 protein. In the pull-down assay, we used truncated TBSV p33 replication protein missing its N-terminal region including the membrane-binding region to aid its solubility in E. coli [61]. Interestingly, the helicase core mutant RH30(F416L) also bound to p33 replication protein as efficiently as the wt RH30 (Fig 5B, lane 3 versus 2). Altogether, these data suggest that the direct interaction between RH30 host protein and the replication protein of TBSV occurs within the viral protein C-terminal domain facing the cytosolic compartment.
To provide additional evidence that RH30 helicase interacts with the tombusvirus replication protein, we have conducted bimolecular fluorescence complementation (BiFC) experiments in N. benthamiana leaves. The BiFC experiments revealed interaction between RH30 and the TBSV p33 replication protein within the viral replication compartment, marked by the peroxisomal matrix marker RFP-SKL (Fig 5C). Altogether, these experiments revealed direct interaction between the cellular RH30 DEAD-box helicase and the TBSV p33 replication protein, which results in re-targeting of RH30 into the viral replication compartment.
To gain insight into the mechanism of CIRF function of RH30 helicase, we affinity-purified the recombinant RH30 and tested its activity in vitro in a TBSV replicase reconstitution assay, which is based on yeast cell-free extract [26,62]. Addition of RH30 to the replicase reconstitution assay led to inhibition of TBSV repRNA replication by ~10-fold (Fig 6A, lanes 9–10). The in vitro production of double-stranded repRNA replication intermediate was also inhibited by ~10-fold by RH30, indicating that RH30 likely inhibits an early step, such as the VRC assembly during TBSV replication.
We then used a step-wise TBSV replicase reconstitution assay [26,29], in which RH30 was added at different stages of VRC assembly (schematically shown in Fig 6B). RH30 showed significant inhibitory activity when added at the beginning of the TBSV replicase reconstitution assay (Fig 6B, lanes 3–4 versus 1–2). On the contrary, RH30 was ineffective, when added to TBSV replicase reconstitution assay after the VRC assembly step and prior to RNA synthesis (Fig 6B, lanes 7–8). These in vitro data support the model that the inhibitory role of RH30 is performed during or prior to the VRC assembly step, but RH30 is ineffective at the latter stages of TBSV replication.
We also utilized an in vitro RdRp activation assay based on the purified recombinant TBSV p92 RdRp, which is inactive and requires Hsp70 chaperone and the viral (+)RNA template to become an active polymerase [21]. Addition of the recombinant RH30 helicase strongly inhibited the polymerase activity of the p92 RdRp (Fig 6C), suggesting that RH30 blocks the critical RdRp activation step during tombusvirus replication.
Several RNA helicases are involved in regulation of cellular translation [63]. Therefore, we tested if RH30 affected the translation of tombusvirus genomic RNA, which is uncapped and lacks poly(A) tail [64]. CIRV genomic RNA was used in this in vitro assay based on wheat germ extract [65]. Addition of recombinant RH30 to the in vitro translation assay inhibited slightly the production of p36 replication protein from the gRNA when RH30 was used in high amount (Fig 6D). The highest amount of RH30 also had minor inhibition on translation of the control Tdh2 mRNA (Fig 6D). Thus, RH30 is unlikely to specifically affect the translation of the tombusvirus RNAs during infection.
Since the canonical function of RNA helicases to bind RNA substrates and unwind base-paired structures [36], we tested if RH30 DEAD-box helicase could perform these functions with the TBSV RNA in vitro. First, we used gel-mobility shift assay with purified recombinant RH30, which showed that RH30 bound to both the (+) and (-)repRNA (Fig 7A and 7B). Since each of the four regions in the TBSV repRNA contains well-defined cis-acting elements, we performed template competition assay with the four regions separately in the presence of recombinant RH30 helicase. This assay defined that the best competitors for binding to RH30 was RII(+) and RII(-), whereas RI(+), RIV(+) and RI(-), RIV(-) also become competitive when added in high amounts (Fig 7C). Because RII(+) contains a critical cis-acting stem-loop element, termed RII(+)SL, which is involved in p33-mediated recruitment of the TBSV (+)RNA template [24], and the activation of the p92 RdRp [21], we tested if the purified RH30 could bind to this stem-loop element in vitro. Interestingly, RH30 bound to RII(+)SL in the absence of added ATP (Fig 7D). However, the presence of extra ATP enhanced the binding of RH30 to RII(+)SL, suggesting that RH30 binds to RNAs in an ATP-dependent fashion, similar to other DEAD-box helicases [36,54,66]. The control p33 (an N-terminally-truncated, soluble version) bound to RII(+)SL more efficiently and in an ATP-independent manner (Fig 7D), as also shown previously [24]. This highlights the possibility that RH30 and p33 replication protein compete with each other in binding to this critical cis-acting element.
To test the RNA helicase function of RH30, we performed strand separation assays, where parts of the TBSV repRNA was double-stranded as shown schematically in Fig 7E and 7F. The RNA helicase activity of RH30 in the presence of ATP was found to efficiently separate the partial dsRNA templates, involving RI and RII sequences (Fig 7E and 7F). RH30 was much less efficient to separate the partial dsRNA templates in the absence of ATP or when we added its helicase core mutant RH30(F416L) (Fig 7E, lanes 6–9; 7F, lanes 5–8). It is possible that the residual strand-separation activity of RH30(F416L) might come from its RNA binding and RNA chaperone activity with the TBSV RNA substrates. Additional biochemical assays will be needed to test if the partial activity of RH30 in the absence of added ATP is due to the possibly copurified residual ATP bound to RH30.
To test if RH30(F416L) helicase core mutant still has antiviral activity, we performed a TBSV replicase reconstitution assay with yeast cell-free extract [26,62]. Addition of RH30(F416L) to the replicase reconstitution assay led to minor inhibition of TBSV repRNA replication (Fig 7G, lanes 1–2). Thus, mutation within the helicase core region of RH30 affected its antiviral activity on TBSV replication in vitro.
To further characterize the restriction function of RH30 during tombusvirus replication, we tested if RH30 helicase could inhibit the selective binding of p33 replication protein to the viral RNA template in vitro. To this end, we biotin-labeled RII(+) sequence of the TBSV RNA, which represents RII(+)-SL RNA recognition element required for template recruitment into replication by p33 replication protein [24]. Moreover, RII(+)-SL RNA is also essential part of an assembly platform for the replicase complex [67]. The biotin-labeled RII(+) RNA was then pre-incubated with purified RH30 (Fig 8A). Then, purified p33C (the soluble C-terminal region, including the RNA-binding and p33:p33/p92 interaction region of p33 replication protein) was added, which can bind specifically to RII(+)-SL if the hairpin structure with the C•C mismatch in the internal loop was formed [24]. After a short incubation, the biotin-labeled RII(+) RNA was captured on streptavidin-coated magnetic beads. After thorough washing of the streptavidin beads, the proteins bound to the RNA were eluted. Western blot analysis with anti-p33 antibody revealed that RH30 in the presence of ATP inhibited the binding of p33C to RII(+)-SL by 50% (Fig 8A, lane 2 versus lane 3) when compared with the control containing the MBP protein that does not bind to RII(+)-SL [24]. RH30 was less inhibitory of the p33C—RII(+)-SL interaction in the absence of ATP (Fig 8A, lane 4). We also performed the experiments when RH30 and p33C were incubated with biotin-labeled RII(+) RNA simultaneously. Western-blot analysis showed that RH30 was still inhibitory of p33C binding to RII(+)-SL (Fig 8B), but less effectively than above when RH30 was pre-incubated with the RII(+) RNA. These in vitro results suggest that one of the mechanisms by which RH30 helicase inhibits tombusvirus replication is to inhibit the binding of p33 to the critical RII(+)-SL RNA recognition element required for template recruitment into replication. This inhibition is likely due to local unwinding RII(+)-SL, because the presence of ATP enhanced the inhibitory effect of RH30.
In another set of experiments, we first incubated biotin-labeled RII(+) RNA with p33C, followed by capturing the RNA-p33 complex with streptavidin-coated magnetic beads and then, the addition of RH30 helicase to the beads (Fig 8C). Here we tested the released p33C from the beads in the eluted fraction by Western blotting. Interestingly, increasing the amounts of RH30 added in the presence of ATP led to the release of p33C from the RII(+) RNA (Fig 8C, lane 3–4), whereas RH30 was less efficient in replacing p33C in the absence of ATP (lanes 1–2). Based on these in vitro data, we suggest that RH30 helicase could replace the RNA-bound p33C by likely remodeling the RNA-p33 complex in an ATP-dependent manner.
We also tested the localization of RH30 helicase in comparison with the viral repRNA in N. benthamiana. The TBSV repRNA carried six copies of the RFP-tagged coat protein recognition sequence from bacteriophage MS2 in either plus or minus polarity [68]. CNV served as a helper virus in these experiments. Interestingly, RH30 was co-localized with both (-)repRNA and (+)repRNA, which were present in the replication compartment decorated by the TBSV p33-BFP (Fig 9A and 9B). The RFP signal within the replication compartment was usually weaker when RH30 helicase was expressed, likely due to the inhibitory effect of RH30 on tombusvirus replication. Similar outcome was observed when the viral dsRNA replication intermediate, detected via dsRNA probes [69], was co-localized with RH30 helicase within the replication compartment (Fig 10). These data demonstrate that RH30 helicase relocates to the replication sites where tombusvirus RNA synthesis takes place.
To learn if RH30 has restriction function against additional plant viruses, we tested the effect of RH30 expression on TCV carmovirus and red clover necrosis mosaic virus (RCNMV) dianthovirus, both of which belong to the Tombusviridae family. Expression of AtRH30 in N. benthamiana plants led to complete block of TCV gRNA accumulation and ~4-fold reduction in RCNMV RNA1 accumulation (Fig 11A and 11B). On the contrary, two separate transgenic RH30 knock-out lines of Arabidopsis thaliana supported increased levels of TCV gRNA accumulation by up to 2-fold (Fig 11C).
The Arabidopsis-TCV system was also used to estimate if TCV infection could induce RH30 gene transcription. RT-PCR analysis revealed induction of RH30 mRNA transcription in TCV-infected versus mock-inoculated plants (Fig 11D). All these data are in agreement that RH30 is a strong restriction factor against tombusviruses and related viruses in plants.
To learn if RH30 also has restriction function against an unrelated plant virus, we over-expressed AtRH30 in N. benthamiana and measured the accumulation of the unrelated tobacco mosaic tobamovirus (TMV). We observed a ~3-fold reduction in TMV RNA accumulation in N. benthamiana leaves expressing the WT RH30, but not in those leaves expressing the helicase core mutant of RH30(F416L) (Fig 11E). Expression of WT RH30, but not that of the RH30(F416L) helicase core mutant, also inhibited the accumulation of the insect-infecting Nodamura virus (NoV) by ~3-fold in yeast (S2A Fig). Interestingly, the accumulation of Flock House virus (FHV), an alphanodavirus, which is related to NoV, was only slightly inhibited by the expression of WT RH30 in yeast (S2B Fig). Based on these observations, we suggest that the plant RH30 DEAD-box helicase has a broad-range CIRF activity against several RNA viruses.
DEAD-box RNA helicases are the most numerous among RNA helicases [33,37]. They are involved in all facets of RNA processes in cells. RNA viruses and retroviruses also usurp several DEAD-box helicases to facilitate their replication and other viral processes during infection [70,71]. However, the host also deploys DEAD-box helicases to inhibit RNA virus replication [70,72]. Accordingly, in this work we present several pieces of evidence that the DDX17-like RH30 DEAD-box helicase restricts tombusvirus replication, including the peroxisomal replicating TBSV and CNV and the mitochondrial-replicating CIRV in yeast and plants, and the more distantly related TCV and RCNMV and the unrelated TMV in plants. On the contrary, knock-down of RH30 enhances the replication of these three tombusviruses in N. benthamiana or the related TCV in RH30 knock-out lines of Arabidopsis. On the other hand, the helicase core mutant RH30 can only partially inhibit tombusvirus replication in plants or in vitro, suggesting that the helicase function of RH30 is needed for its full antiviral activity.
How can RH30 restrict TBSV replication? We show that the antiviral RH30 helicase binds to p33 and p92pol replication proteins based on co-purification experiments of the viral replicase complex, a pull down assay, and BiFC in N. benthamiana. We propose that the interaction of RH30 helicase with the viral replication proteins might be important for the targeting of RH30 into the viral replication compartment (Fig 12). Accordingly, RH30 is recruited into the viral replication compartment from the cytosol and the nucleus based on live imaging in plant cells (Fig 3). The targeting of RH30 into the replication compartment is critical for its antiviral function, because fusion of a nuclear retention signal with RH30, which leads to its enrichment in the nucleus at the expense of the cytosolic pool of RH30, in turn, cancelled out the antiviral effect of RH30. Yeast CFE-based replicase reconstitution assays showed that RH30 acts in the early steps of replication, since both (-) and (+)RNA synthesis was inhibited by RH30 (Fig 6). Moreover, the in vitro RdRp activation assay demonstrated that RH30 inhibited the TBSV RdRp activation step during the replication process as well (Fig 6C). In contrast, the CFE-based TBSV replication was not inhibited by RH30 after replicase assembly was completed (see step 2, Fig 6B). These data suggest that RH30 DEAD-box helicase must act at the earliest steps in the replication process to inhibit TBSV replication.
RH30 also binds to the viral RNA, including the 5’ UTR (i.e., RI) and RII internal sequence present within the p92pol coding region (Fig 7). Using in vitro interaction and replication assays between RNA-p33 replication protein, we show that RH30 inhibits several steps in tombusvirus replication. These include the RH30-based inhibition of (i) the specific recognition of the critical RII(+)-SL cis-acting element in the viral (+)RNA by p33 replication protein, which is absolutely required for template recruitment into VRCs, (ii) the activation of the viral p92 RdRp, and (iii) the assembly of the VRCs [21,26,73]. Moreover, RH30 helicase could disassemble viral RNA-p33 complexes by likely remodeling the RNA structure in an ATP-dependent manner (Fig 8). However, RH30-mediated disassembly of viral RNA-p33 complexes is unlikely to occur after VRC assembly is completed, because RH30 helicase was not an effective restriction factor when added at a late step of TBSV replication (step 2, Fig 6B). We propose that the membrane-bound TBSV VRCs are protecting the viral RNA-p33 complexes by restricting accessibility of the VRC complex to RH30 DEAD-box helicase. Accordingly, we have shown before that the fully-assembled TBSV VRCs are resistant to cellular ribonucleases [74]. Therefore, RH30 helicase might only be able to disassemble viral RNA-p33 complexes before the vesicle-like spherule formation, which is the characteristic structure of the TBSV VRCs in yeast and plants [75]. Altogether, the in vitro assays provide plentiful data on the direct inhibitory effect of RH30 helicase on TBSV replication, indicating that RH30 functions as an effector-type, not signaling-type, DEAD-box helicase, which detect viral RNA and send signals to downstream components of the innate immunity network [72]. Future experiments will address if RH30 might have additional mechanisms to restrict tombusvirus replication.
A recently emerging concept in innate immunity is the significant roles of DEAD-box helicases expressed by host cells that greatly reduce virus replication and facilitate combating viruses and making the induced and passive innate immune responses more potent. Many of the identified yeast DEAD-box helicases with restriction functions against TBSV are conserved in plants and mammals. Altogether, the genome-wide screens performed with animal viruses have shown that helicases are the largest group of host proteins affecting RNA virus replication. For example, in case of HIV, the involvement of several cellular helicases has been demonstrated, including DDX17 and DDX3 [71,76,77]. Yet, the functions of the cellular helicases during virus replication are currently understudied.
The emerging pricture in plant-virus interactions, similar to animal-virus interactions, is the diverse roles of various host RNA helicases. Different plant viruses have been shown to co-opt plant RNA helicases for pro-viral functions. These include RH8 and RH9 for potyvirus replication and RH20, RH2 and RH5 for TBSV replication [27–29,39,44,78]. However, this paper shows evidence that a plant DEAD-box helicase, RH30, can also be utilized by host plants for antiviral functions. Thus, in addition to the previously identified Dicer-like RNA helicases [16,79–81], additional plant RNA helicases might function as CIRFs by recognizing plant virus RNAs. The DDX17-like RH30 DEAD-box helicase characterized here opens up the possibility that among the more than 100 helicases of plants, there are additional ones with antiviral functions, serving as effector-type or sensor-like RNA helicases. The discovery of the antiviral role of RH30 helicase illustrates the likely ancient roles of RNA helicases in plant innate immunity. In summary, we have demonstrated that the plant DDX17-like RH30 DEAD-box helicase acts as a major restriction factor against tombusvirus replication when expressed in plants and yeast surrogate host. We show that RH30 DEAD-box helicase is targeted to the large TBSV replication compartment. In addition, we find that RH30 blocks the assembly of viral replicase complex, the activation of the RNA-dependent RNA polymerase function of p92pol and binding of p33 replication protein to critical cis-acting element in the TBSV RNA (Fig 12). Altogether, the plant DDX17-like RH30 DEAD-box helicase is a potent, effector-type, restriction factor of tombus- and related viruses.
Biotinylated RII RNA of DI-72(+) was synthesized by in vitro T7 transcription in the presence of 7.5 μl of 10 mM ATP, CTP, GTP and 5 mM UTP as well as 0.35 μl of 10 mM biotin16-UTP (Roche) in a total of 50 μl reaction volume. The interaction assay was performed with 3.8 μM of recombinant MBP-RH30 and 1.9 μM of MBP-p33C along with 0.1 μg of biotinylated RNA, 0.1 μl of tRNA (1 mg/ml), 2 U RNase inhibitor, and 1 mM ATP in the presence of biotin-RNA binding buffer (100 mM Tris [pH 7.9], 10% glycerol, 100 mM KCl, 5 mM MgCl2, 0.1% NP-40) in a 10 μl reaction mixture. Non-biotinylated RII of DI-72(+) RNA or absence of ATP was used as controls.
Assay #1: Recombinant MBP-RH30 was incubated first with biontinylated RII(+) RNA at 25°C for 15 min. Then, the recombinant MBP-p33C was added to the reaction and incubated for another 15 min. Assay #2: Recombinant MBP-RH30 and MBP-p33C were co-incubated simultaneously with biontinylated RII(+) RNA at 25°C for 30 min. The reaction mixtures were incubated with 20 μl of Promega Streptavidin MagneSphere Paramagnetic Particles (VWR) at room temperature for 20 min. The particles were collected in a magnetic stand and washed with binding buffer for five times. The protein-RNA complexes were then eluted with 20 μl of SDS loading dye containing β-mercaptoethanol by boiling for 15 min. The eluted samples were analyzed by Western blot with anti-p33 antibody.
Assay #3: For the detection of p33 released from protein-biotinylated RNA complex, 1.9 μM of recombinant MBP-p33C was incubated with 0.1 μg of biontinylated RII of DI-72(+) RNA at 25°C for 15 min, followed by the addition of 20 μl of Promega Streptavidin MagneSphere Paramagnetic Particles for another 30 min incubation at room temperature. After collection of the beads and washing with biotin-RNA binding buffer for five times, the particles were incubated with either 0.95 or 3.8 μM of MBP-RH30 or MBP (used as control) in the presence of biotin-RNA binding buffer containing 1 mM ATP at 25°C for 15 min. The supernatant of the mixture was collected after collecting the particles in a magnetic stand and was analyzed by Western blot with anti-p33 antibody.
The conditions for the EMSA experiments were described previously [24]. Briefly, the EMSA assay was performed with 0.1 pmol of 32P-labeled RNA probes along with different concentrations (0.4, 1.9, and 5.7 μM) of purified recombinant MBP-fusion proteins or MBP in the presence of RNA binding buffer (10 mM HEPES [pH7.4], 50 mM NaCl, 1 mM DTT, 1 mM EDTA, 5% Glycerol, 2.5 mM MgCl2), 2 U of RNase inhibitor, as well as 0.1 μg of tRNA in a total of 10 μl reaction volume. Two different amounts (2 and 4 pmol) of unlabeled RNAs together with 5.7 μM of either MBP-RH30 or MBP were used for template competition.
To study if purified proteins could unwind partial dsRNA duplex, the dsRNA strand-separation assay was performed as described [28]. Firstly, the unlabeled single-stranded DI-72 (-) or DI-72 (+) RNAs were synthesized via T7 polymerase- based in vitro transcription. The 32P-labeled single-stranded RI(-) or RII(+) RNAs were synthesized by T7-based in vitro transcription using 32P-labeled UTP. To prepare partial dsRNA duplexes, consisting of either RI(-)/DI-72 (+) or RII(+)/DI-72 (-) (see Fig 7E and 7F), 2 pmol of 32P -labeled RI(-) or RII(+) were annealed to 6 pmol of unlabeled DI-72(+) or DI-72 (-) in STE buffer (10 mM TRIS [pH 8.0], 1 mM EDTA, and 100 mM NaCl) by slowly cooling down the samples (in a total volume of 20 μl) from 94°C to 25°C in 30 min. To test if the purified recombinant proteins could separate the partial dsRNA duplex, 1.9 and 5.7 μM purified MBP fusion proteins or MBP as a negative control were added separately to the partial dsRNA duplex in the RNA binding buffer (10 mM HEPES [pH7.4], 50 mM NaCl, 1 mM DTT, 1 mM EDTA, 5% Glycerol, 2.5 mM MgCl2) along with 1mM ATP, followed by incubation at 25°C for 25 min. The reaction mixtures were then treated with Proteinase K (2 μg/per reaction) at 37°C for 20 min, followed by loading onto 5% nondenaturing polyacrylamide gel with 200V for 1 h. Additional methods can be found in S1 Text and the primers used are listed in S1 Table.
|
10.1371/journal.pgen.1006625 | ANLN truncation causes a familial fatal acute respiratory distress syndrome in Dalmatian dogs | Acute respiratory distress syndrome (ARDS) is the leading cause of death in critical care medicine. The syndrome is typified by an exaggerated inflammatory response within the lungs. ARDS has been reported in many species, including dogs. We have previously reported a fatal familial juvenile respiratory disease accompanied by occasional unilateral renal aplasia and hydrocephalus, in Dalmatian dogs. The condition with a suggested recessive mode of inheritance resembles acute exacerbation of usual interstitial pneumonia in man. We combined SNP-based homozygosity mapping of two ARDS-affected Dalmatian dogs and whole genome sequencing of one affected dog to identify a case-specific homozygous nonsense variant, c.31C>T; p.R11* in the ANLN gene. Subsequent analysis of the variant in a total cohort of 188 Dalmatians, including seven cases, indicated complete segregation of the variant with the disease and confirmed an autosomal recessive mode of inheritance. Low carrier frequency of 1.7% was observed in a population cohort. The early nonsense variant results in a nearly complete truncation of the ANLN protein and immunohistochemical analysis of the affected lung tissue demonstrated the lack of the membranous and cytoplasmic staining of ANLN protein in the metaplastic bronchial epithelium. The ANLN gene encodes an anillin actin binding protein with a suggested regulatory role in the integrity of intercellular junctions. Our study suggests that defective ANLN results in abnormal cellular organization of the bronchiolar epithelium, which in turn predisposes to acute respiratory distress. ANLN has been previously linked to a dominant focal segmental glomerulosclerosis in human without pulmonary defects. However, the lack of similar renal manifestations in the affected Dalmatians suggest a novel ANLN-related pulmonary function and disease association.
| Acute respiratory distress syndrome (ARDS) is characterized by life-threatening impairment of pulmonary gas exchange and leads to substantial mortality in man. Spontaneous ARDS has also been described in dogs including a familial fatal ARDS-like syndrome in young Dalmatian dogs. The main clinical signs include progressive tachypnea and dyspnea leading to a severe respiratory distress and euthanasia. The prominent clinicopathological findings involve pulmonary lesions, although some affected puppies also presented with renal aplasia and hydrocephalus. This study finds the genetic cause of the disease by identifying a recessive nonsense variant in the ANLN gene. The ANLN gene encodes an anillin actin binding protein which has an important role in the integrity of the epithelial cell organization. The functional defect of ANLN due to early truncation and absence from the bronchiolar epithelium is consistent with the observed histopathology with hyper- and metaplasia of the bronchiolar epithelium and clinical ARDS. Our study reveals a novel pulmonary disease association for ANLN, provides new insights to pathophysiology and has enabled the development of a genetic test for breeding purposes.
| Acute respiratory distress syndrome (ARDS) is a multifactorial syndrome characterized by rapid-onset respiratory failure resulting from pulmonary inflammation [1]. ARDS is common and leads to substantial mortality in man [1]. Two forms of idiopathic interstitial pneumonia are found in human ARDS: a diffuse alveolar damage (DAD) in acute interstitial pneumonia (AIP) and an acute exacerbation of usual interstitial pneumonia in idiopathic pulmonary fibrosis [2].
The molecular mechanisms leading to human ARDS remain largely unknown. Candidate gene studies suggest the involvement of inflammatory mediators such as interleukins IL-6, IL-8 and IL-32 [3,4], pre-B-cell colony-enhancing factor (PBEF) [5,6], and angiotensin-converting enzyme (ACE) [7]. In addition, the nuclear factor erythroid-derived 2–like 2 (NFE2L2) transcription factor has been identified as a potential mediator of acute lung injury in a mouse model [8].
Spontaneous ARDS has also been described in dogs [9]. We previously described a familial fatal ARDS-like syndrome in young Dalmatian dogs with the main clinical signs including progressive tachypnea and dyspnea leading to severe respiratory distress and euthanasia [10]. The clinicopathological findings were restricted to pulmonary lesions in the majority of the affected Dalmatians, although some of the puppies presented with concurrent unilateral renal aplasia and hydrocephalus [10]. Pulmonary manifestations included multiple foci of marked atypical hyperplasia and squamous metaplasia of the bronchiolar epithelium, patchy ongoing fibrosis with myofibroblastic metaplasia, smooth muscle hyperplasia and occasional honeycombing of alveoli and hyperplasia of type II pneumocytes (PCII) along with acute alveolar edema [11].
Exclusion of specific causes like exposure to toxins and viruses and overrepresentation of the cases in the Dalmatian breed suggested a recessive genetic defect [10], which we aimed to reveal in this study. We describe the identification of a fully penetrant recessive nonsense variant in a novel candidate gene, ANLN. The ANLN gene encodes an anillin actin binding protein which has an important role in the integrity of the epithelial cell organization. The functional defect of ANLN due to early truncation is consistent with the observed histopathology with hyper- and metaplasia of the bronchiolar epithelium, consecutive DAD and clinical ARDS.
To identify the genetic cause of ARDS in Dalmatians, we performed a combined analysis of homozygosity mapping and whole genome sequencing (WGS). The study cohort of eleven Dalmatians including two affected littermates, one healthy obligate carrier, one healthy sibling, one healthy grandparent and six other healthy dogs were genotyped using Illumina’s CanineHD SNP array. Genotype data of two cases was used for homozygosity mapping, which revealed 49 shared homozygous regions (S1 Table). Whole genome sequencing with mean coverage of 16x was performed on one affected dog. The filtering of variants from WGS data under recessive model of inheritance against WGS and exome variant data of 136 unaffected dogs (S2 Table) uncovered 16,195 case-specific variants of which 98 were exonic (Table 1). Only 15 out of the 98 coding variants were found in the homozygosity regions of which eight variants were either non-synonymous (n = 4), frameshift (n = 3) or nonsense (n = 1) (Table 2). Manual inspection of the two PSMD6 deletions with Integrated Genome Viewer (IGV) revealed that the variants were also present in several control genomes (S2 Table), excluding the gene as a candidate for the disease. The other six exonic variants in ANLN, CD302, GANAB, MUC5B, OR16D05 and ORO8C02 genes (Table 2) were genotyped in a cohort of twelve dogs, including seven cases and five closely related unaffected dogs (parent, grandparent and three healthy siblings). Variants in the CD302, GANAB, MUC5B, OR16D05 and ORO8C02 genes were excluded as they did not agree with a recessive segregation pattern. A complete segregation was found only for the c.31C>T variant in the ANLN gene. All seven affected dogs were homozygous, both the parent and the grandparent were heterozygous, and healthy littermates were either heterozygous (1/3) or wild-type (2/3) (Fig 1A and 1B).
The ANLN c.31C>T variant is located within a 2 Mb region of continuous homozygosity on chromosome 14 at 46.78–48.71 Mb (Fig 1A). The 2 Mb haplotype is part of the larger 14.6 Mb homozygosity region identified by homozygosity mapping (S1 Table). The entire 14.6 Mb homozygosity region had an average 15x coverage (>90% bases with at least 10x coverage) and did not contain other case-specific coding homozygous variants. We manually assessed the haplotypes surrounding the ANLN variant in eleven dogs that were genotyped with the SNP array (Fig 1A). The 2 Mb haplotype was homozygous in both cases and in the parent and the grandparent. However, only the two cases were homozygous for the ANLN c.31C>T variant while the parent and the grandparent were heterozygous, suggesting a recent origin of the risk variant in the pedigree.
Further validation of the ANLN c.31C>T variant in 176 randomly selected unaffected Dalmatian dogs revealed a low 1.7% carrier frequency (3/176). No new genetically affected dogs were found while three new non-Finnish carriers were identified. The pedigree information of these three dogs was not available and therefore the relationship to the original Finnish discovery population remains unknown. A combined analysis in the entire cohorts of 188 dogs (12 +176 Dalmatians) gave a highly significant association between the T allele and the disease (p = 3.075x10-58).
Breed-specificity was studied by screening the ANLN variant in 31 Pointers, which is considered as the closest relative to Dalmatians. None of the Pointers carried the variant.
The identified ANLN c.31C>T nonsense variant is predicted to result in an early truncation of the normal 1121 amino acid ANLN protein after the first ten residues (Fig 1C). This severe truncation very likely completely abrogates the ANLN function.
The effect of the truncation could not be assessed at transcriptional level, as fresh RNA samples from the affected dogs were not available. Therefore, the ANLN expression was analyzed by immunohistochemistry, using antibody recognizing residues 1074–1124 of the protein, expected to be lacking in the affected dogs. Various organs from an age- and breed-matched control dog were included, and the staining was compared to that in the lungs and other tissues available from four affected Dalmatians. In addition, the expression pattern in the lungs of the affected dogs was compared to a canine lung affected by diffuse alveolar damage (DAD) to include the assessment of PCII cells, which are important for normal regeneration process in the alveolus [9]. The kidneys from seven affected Dalmatians were histologically re-evaluated, since loss of ANLN function has been linked to human focal segmental glomerulosclerosis (FSGS) [12]. Glomerular collagen was highlighted by Masson-trichrome staining and basal membranes by periodic acid-Schiff staining (PAS) in order to reveal even subtle fibrosis.
We found a specific positive membranous ANLN-signal in the epithelial cells lining the terminal bronchioles in the control sample (Fig 2A) and in the control lung affected by DAD. In addition, proliferating PCII cells showed a strong cytoplasmic positivity in the canine lung affected by DAD (Fig 2C). In contrast, the membranous bronchiolar epithelial ANLN-signal as well as the cytoplasmic PCII staining were absent in the lung specimens of the affected Dalmatians (Fig 2B and 2D). Few basal bronchial and alveolar interstitial cells showed nuclear positivity in all groups. In other organs, ANLN expression was neither detected in the control dogs, nor in the tested affected dogs. Glomerular tufts appeared histologically normal, with no excessive collagen deposition (Fig 2E) and slender glomerular basal membranes (Fig 2F). These results support the conclusion that ANLN is mainly expressed in the canine lung, co-localizes with the histological lesion and is lacking in the affected Dalmatians. Specific ANLN expression was not detected in organs other than lung in the canine control samples. Lesions similar to those described in human FSGS were not present in the kidneys of the affected dogs.
This study reveals the primary cause of ARDS in Dalmatian dogs by identifying a recessive nonsense variant (c.31C>T, p.R11*) in the ANLN gene. Several findings support the causality of the ANLN defect in ARDS. First, the nonsense variant was the only case-specific coding change within the homozygosity regions that fully segregated with the disease (p = 3.075x10-58). Second, the genetic defect results in the very early truncation abolishing the function of the protein. Third, the ANLN protein was found to be expressed predominantly in the lungs, which is the key affected organ in ARDS. ANLN was absent in the lungs of the affected dogs. Finally, ANLN is a relevant functional candidate gene, since it plays a role in cell division and in the assembly of intercellular junctions [12–14]. Histopathology of the affected dogs show a disorganized bronchiolar epithelial regeneration attempt and disturbed alveolar epithelial regeneration [10, 11], which could be due to improper ANLN contributions. Therefore, our study has important implications since it uncovers a novel candidate gene for ARDS and sheds new light on the understanding of the underlying pathophysiology.
The ANLN gene encodes a 124 kDa intracellular multi-domain protein (Fig 1C) that is expressed in several organs, including lungs, kidney and brain [14–16]. ANLN interacts with F-actin and CD2-associated protein (CD2AP) [12] and is implicated in cytoskeletal dynamics [14]. As ANLN interacts both with the organizing mitotic spindle during mitosis and the cytoskeletal actin in cellular migration [15], the absence of the protein and consequent disturbed migration and proliferation of PCII at the alveolar level could trigger the immense bronchiolar epithelial regeneration attempt seen on the histopathology of the affected Dalmatians. One of the key events in the repair of alveolar injury involves the proliferation and migration of ANLN-positive PCII cells [9]. In addition, the atypical, broad based and multinucleated PCII of the affected dogs correlate morphologically with the disturbed cell division, likely caused by the absence of functional ANLN.
An alternative or complementary hypothesis is that the loss of ANLN results primarily in the disorganized bronchiolar epithelia in the affected Dalmatians due to improper assembly of intercellular junctions. Hereby the hyper-and dysplastic epithelium acts as a mechanical hurdle at the bronchioalveolar junction during expiration and air is trapped at the alveolar level during passive exhalation, causing over-extension damage to the alveolar wall and ARDS. A comparable pathogenesis leads to ARDS in ventilator-induced lung injury in man and dog, where alveolar over-inflation, with consecutive alveolar edema and alveolar emphysema, progresses into interlobular emphysema and pneumomediastinum [9]. Seven of the affected dogs suffered from marked alveolar edema, five from marked alveolar emphysema [10, 11]. In addition, three dogs developed pneumomediastium [10], which is rare in dogs not suffering from perforating trauma of the esophagus, neck or trachea.
ANLN has been associated in diverse forms of neoplastic disease in man [15, 17–20] with a proposed role in regulating intercellular adhesion in the epithelia [21]. Interestingly, another scaffold protein Alix has recently been associated with the maintenance of epithelial cell polarity and assembly of intercellular junctions [16]. Abnormal structure of the choroid plexus epithelium and ependymal in the Alix knockout mouse results in enlargement of the lateral ventricles and hydrocephalus as the homeostasis of the blood-cerebrospinal fluid barrier requires intact tight junctions. Alix, just like anillin, interacts with F-actin and in addition, with tight junction protein ZO-1, being essential for the maintenance of epithelial cell polarity and barrier. Some of the ARDS-affected Dalmatians manifested also hydrocephalus and renal aplasia [10], which could be caused by abnormal assembly of intercellular junctions in the epithelium of the choroid plexus and the ureteric bud epithelium during early organogenesis.
Mutations in the ANLN gene have been linked to human FSGS and ANLN has been suggested to play a role in retaining the podocyte function in the glomerular filtration barrier [12]. We did not identify ANLN in the normal canine glomeruli by IHC staining. These findings are similar to the IHC findings in man, where ANLN expression was upregulated in the glomeruli affected by FSGS but not detected in normal glomeruli. Apart from unilateral renal aplasia in two affected dogs, serum biochemistry and renal histology of the affected Dalmatians did not reveal other renal disease [10, 11]. The suggested conclusion [12] that ANLN expression is induced in response to podocyte injury and repair, not in the end-differentiated mature podocyte, remains unverified in dogs as early onset and lethal outcome of the disease prevents further study of a potential renal phenotype. Unilateral renal aplasia may result from a developmental expression of ANLN.
Additional evidence for the physiological significance of the loss-of-function variants in genes can be explored utilizing available variant databases. We searched possible loss-of-function variants (frameshifts and nonsense variants) for canine and human ANLN in our canine variant database and in public Genome Aggregation Database (gnomAD) [22]. No loss-of-function variants were found in 136 unaffected dogs while the exploration of the GnomAD database revealed four heterozygous frameshifts and fourteen heterozygous nonsense variants with very low allele frequencies and most being singletons. It therefore appears that the loss-of-function variants in ANLN are extremely rare across species, which supports the vital role of the gene for survival and is in agreement with the observed lethal disease in the affected Dalmatians.
In summary, our study reveals a novel lethal pulmonary disease association with the ANLN gene and suggests that abnormal cytoskeletal dynamics and epithelial regeneration due to lack of functional ANLN result in the hyper- and metaplastic bronchiolar epithelium that predisposes the affected dogs to ARDS. A genetic test can be established to facilitate veterinary diagnostics and to eradicate the detrimental condition in the affected breed.
The experiments performed on dogs were approved by the “Animal Ethics Committee at the State Provincial Office of Southern Finland” (permits: ESAVI/6054/04.10.03/2012 and ESAVI/7482/04.10.07/2015, expire date 17.10.2020) and by the “University of Helsinki Viikki Campus Research Ethics Committee” (Statement 4/2014).
Altogether 188 Dalmatian dogs and 31 Pointers were included in the study. Samples were obtained from seven affected Dalmatian dogs from four litters presented to the Small Animal Hospital of Helsinki University as described previously [10]. The mean age at the onset of illness in the puppies included in this study was seven months (range 5–10 months). The mean duration of illness varied from one to six weeks, with a mean of three weeks. Four of the puppies were male, three female. DNA was extracted either from formalin fixed paraffin embedded (FFPE) tissue samples (four dogs), from bronchoalveolar lavage samples (two dogs) or from whole EDTA -blood of unaffected Dalmatian dogs in Finland. All dogs in this study were privately owned pets that were examined with the owners’ consent.
Genomic DNA from the FFPE and EDTA samples was extracted using the semi-automated Chemagen extraction robot (PerkinElmer Chemagen Technologie GmbH, Germany). DNA from the BAL samples were extracted using QIAamp DNA Micro Kit (Qiagen, Germany). DNA concentration was determined either with the NanoDrop ND-1000 UV/Vis Spectrophotometer (Thermo Fisher Scientific Inc., USA) or Qubit 3.0 Fluorometer (Thermo Fisher Scientific Inc., USA).
Genome-wide SNP genotyping of two affected, one healthy sibling, one obligate (parent) and one possible carrier (grandparent) and six healthy control Dalmatian dogs was performed at the GeneSeek facility (Neogen Corporation, USA) using Illumina’s CanineHD BedChips containing 173,662 validated SNPs. Genotypes were stored in BC/Gene database version 3.5 (BC/Platforms). The PLINK v 1.07 software was used to search for segments of extended homozygosity in the two affected dogs as described previously [23, 24]. Genotype data was filtered using a SNP call rate of > 95%, an array call rate > 95% and minor allele frequency of > 0.05. The genotype data is available for further use upon request.
We performed WGS of one affected Dalmatian dog and used 136 other dog genomes (48 whole exome sequences and 88 whole genome sequences) available as controls (S2 Table). A fragment library was prepared with a 290 bp insert size and collected to a single lane of Illumina HiSeq2000 paired-end reads (2 x 100 bp). The reads were processed using speedseq align module available in SpeedSeq suite to produce a duplicate-marked, sorted and indexed BAM file. The Genome Analysis Tool kit (version = 3.3.0-g37228af) was used to perform realignment around potential indel sites and base quality score recalibration using the known SNP variation available at the Broad Institute (https://www.broadinstitute.org/ftp/pub/vgb/dog/trackHub/canFam3/variation). Dual algorithms, Samtools mpileup (version samtools-1.2) and GATK haplotype caller module were used to detect variants and the variants from both algorithms were merged into variant call format (VCFv4.1). In summary, 98.37% of the reads from Dalmatian dog were mapped to the reference genome yielding a genome-wide mean coverage of 16X. We identified 1,475,318 indels and 4,804,627 SNPs of which 39.91% of the variants were known and 60.09% were novel compared to SNPs from Axelsson et al., Lindblad-Toh et al. and Vaysse et al. and dbSNP build 131 [25–27]. Annovar and SnpEff tools were used to annotate the variants to Ensembl, NCBI and Broad annotation databases to predict the functional effects of the variants. Canine genome build CanFam 3.1 was used as a reference sequence.
We used PCR and Sanger sequencing to perform targeted genotyping for selected variants in the candidate gene. PCR primers were designed with Primer 3 [28] to assess the prevalence of the mutation by Sanger sequencing in a cohort of Dalmatian dogs. We performed a standard PCR, including 0.5 U Biotools DNA Polymerase (Biotools, Madrid, Spain), 2.0 mM MgCl2 (Biotools, Madrid, Spain), 200 μM dNTPs (Finnzymes, Espoo, Finland), 1 x Biotools PCR Buffer (Biotools, Madrid, Spain), 0.5 μM forward primers and 0.5 μM reverse primers (S3 Table). All primers were custom ordered from Sigma Aldrich (St. Louis, MO, USA). Reaction mixtures were subjected to the thermal cycling program of 95°C for 10 min, 35 cycles of 95°C for 30 s, 30 s 57°C, 72°C for 40 s and final elongation state of 72°C for 10 min. Genomic PCR products were sequenced using a 3730xl DNA Analyzer (Applied Biosystems, Foster City California, USA) in the core facility, Institute for Molecular Medicine Finland (FIMM, Technology Centre, University of Helsinki, Helsinki, Finland). We analyzed the sequence data with Sequencher 5.3 software (Gene Codes Corp, Ann Arbor, MI, USA).
The UniProt database (http://www.uniprot.org) and SMART tool (http://smart.embl-heidelberg.de) were used to confirm the protein domain structure of ANLN [29–30]. The ANLN sequence alignment and cross-species conservation was analyzed with ClustalW2 algorithm (http://www.ebi.ac.uk/Tools/clustalw2/). All numbering with the ANLN gene correspond to the accessions ENSCAFG00000003243 (gene) and ENSCAFT00000005209 (protein).
Archived paraffin blocks of autopsy tissue samples from four affected Dalmatian dogs were available for immunohistochemical staining. Autopsy lung samples from a 5-month-old, male Dalmatian, euthanized due to epilepsy and without histopathological changes in the lungs and other internal organs was used as healthy controls. Organs available for assessment of the ANLN expression in the normal dog included lung, kidney, smooth and cross-striated muscle, heart, choroid plexus, testis, liver, spleen, pancreas and lymph node. Lung samples from a 5-year-old Chihuahua male, euthanized due to ARDS with histopathologically confirmed AIP and DAD, were used as comparison of PCII expression pattern in wild-type dogs. Paraffin blocks were sectioned at 4 μm thickness and deparaffinized, antigens were retrieved with 0.01M citrate buffer at pH 6 and heat for 20 minutes at 99°C. Overnight- incubation was used for the primary antibody (rabbit-polyclonal anti-Anillin Antibody aa1074-1124, LS-C288200 (LifeSpan BioSciences, Inc., USA). The sections were stained according to the UltraVision Detection System HRP/DAB kit (Thermo Fisher Scientific Inc., USA). Separate tissue sections from all of the dogs were also stained with hematoxylin and eosin (HE). The hematoxylin-eosin stained histological slides of kidneys from seven affected Dalmatians, including two adults, were histologically re-evaluated for a renal phenotype. Paraffin blocks from four puppies were available for further studies of the kidney and glomerular collagen was highlighted by Masson-trichrome (MTC) staining and basal membranes by periodic acid-Schiff (PAS) staining in order to reveal even subtle fibrosis.
|
10.1371/journal.pcbi.1004664 | A Computational Model for the AMPA Receptor Phosphorylation Master Switch Regulating Cerebellar Long-Term Depression | The expression of long-term depression (LTD) in cerebellar Purkinje cells results from the internalisation of α-amino-3-hydroxy-5-methylisoxazole-4-propionic acid receptors (AMPARs) from the postsynaptic membrane. This process is regulated by a complex signalling pathway involving sustained protein kinase C (PKC) activation, inhibition of serine/threonine phosphatase, and an active protein tyrosine phosphatase, PTPMEG. In addition, two AMPAR-interacting proteins–glutamate receptor-interacting protein (GRIP) and protein interacting with C kinase 1 (PICK1)–regulate the availability of AMPARs for trafficking between the postsynaptic membrane and the endosome. Here we present a new computational model of these overlapping signalling pathways. The model reveals how PTPMEG cooperates with PKC to drive LTD expression by facilitating the effect of PKC on the dissociation of AMPARs from GRIP and thus their availability for trafficking. Model simulations show that LTD expression is increased by serine/threonine phosphatase inhibition, and negatively regulated by Src-family tyrosine kinase activity, which restricts the dissociation of AMPARs from GRIP under basal conditions. We use the model to expose the dynamic balance between AMPAR internalisation and reinsertion, and the phosphorylation switch responsible for the perturbation of this balance and for the rapid plasticity initiation and regulation. Our model advances the understanding of PF-PC LTD regulation and induction, and provides a validated extensible platform for more detailed studies of this fundamental synaptic process.
| Changes in synaptic strength, which can include long-term potentiation and long-term depression, are important for learning and the encoding of memories across the brain. Long-term depression (LTD), in particular, is thought to be essential for motor learning in the cerebellum, and disruption of this process, by disease or injury, can result in severe motor dysfunction. Cerebellar LTD is achieved by reducing the population of AMPA receptors at the Purkinje cell postsynaptic membrane. This population is maintained by a dynamic trafficking loop, in which AMPA receptors are continuously removed from the postsynaptic membrane by endocytosis and reinserted by exocytosis. Specific phosphorylation sites on the AMPA receptors regulate their interaction with proteins that either stabilise the receptors at the membrane or promote their trafficking. We develop a detailed bidirectional computational model of this trafficking loop and its regulation. The model shows how perturbing the trafficking balance towards AMPA receptor mobilisation and endocytosis can account for rapid induction of cerebellar LTD, and suggests mechanistic explanations for numerous features observed experimentally. This deepens our understanding of cerebellar LTD and provides a foundation for further experimental studies of this synaptic process.
| The functional plasticity of neuronal synapses, including long-term potentiation (LTP) and long-term depression (LTD), is essential for learning and the encoding of memories [1]. The focus of this study is LTD at the parallel fibre-Purkinje cell (PF-PC) synapse in the cerebellum, which is believed to play an important role in motor learning [2–4]. This form of LTD requires [5, 6] the concurrent activation of a sufficiently large fraction of the around 175,000 excitatory en passant contacts made from cerebellar parallel fibres to the Purkinje cell dendritic tree [7] and of a climbing fibre comprising several thousand synaptic contacts [8]. PF-PC LTD is linked to the endocytic removal of α-amino-3-hydroxy-5-methylisoxazole-4-propionic acid receptors (AMPARs) from the Purkinje cell postsynaptic membrane [9–11]. The synaptic AMPAR population is dynamically controlled through lateral diffusion into and out of the synapse [8], and receptor endocytosis and exocytosis between the cell surface and the endosome [12]. Endosomes store the internalised AMPARs before they are directed to either reinsertion into the membrane during plasticity [2–4] or degradation [13]. The AMPAR degradation [11, 14] and de novo synthesis [15] provide additional regulation for the receptor population. PF-PC LTD is dependent on the increased internalisation of AMPARs relative to their reinsertion [16].
PF-PC LTD is induced by the activation of protein kinase C (PKC) [17], elevated intracellular calcium [18] and the concurrent inhibition of serine/threonine phosphatase activity [19, 20]. The mechanics of PF-PC LTD are partly controlled by two AMPAR-GluA2 subunit interacting proteins, glutamate receptor interacting protein (GRIP) and protein interacting with C kinase 1 (PICK1) [5, 6], both of which bind at the same site via their C-terminal PDZ domains [21]. The three GRIP isoforms are functionally indistinguishable [22], so we refer to them simply as GRIP. GRIP interacts with AMPARs, stabilising and clustering them both at the plasma membrane and at intracellular endosomal pools [23, 24]. This interaction prevents AMPAR trafficking [21, 23, 25], and AMPAR dissociation from GRIP is essential for the expression of PF-PC LTD [26]. AMPARs that lack the GRIP interaction are unable to stably incorporate into synapses [27].
PICK1 actively promotes AMPAR endocytosis in cerebellar Purkinje cells [6, 26, 28] and the PICK1-AMPAR interaction is indispensable for PF-PC LTD expression [6, 10, 28–30]. PICK1 also associates with the active form of PKCα [31], which phosphorylates the S880 C-terminus residue of the AMPAR-GluA2 subunit [32, 33] sustained by positive feedback mechanisms for at least 20 minutes during LTD induction [34, 35]. GluA2-S880 phosphorylation, which is elevated after the induction of LTD in hippocampal slices [36] and is required for PF-PC LTD [37], abolishes binding between GluA2 and GRIP. However, GluA2 binding to PICK1 is unaffected [38]. Dissociation of GRIP therefore allows PICK1 to bind at the same AMPAR-GluA2 site, promoting AMPAR internalisation. Disruption of the GluA2-GRIP interaction and AMPAR declustering are specifically associated with LTD induction [39]. PICK1 also interacts with GRIP and this enhances GluA2-S880 phosphorylation, possibly by directing PKCα to the GluA2 subunit [33]. The role of PICK1 in AMPAR reinsertion remains unclear, with several studies suggesting conflicting roles [21, 28, 40, 41].
Phosphorylation of the tyrosine GluA2-Y876 by Src family kinases (SFKs) negatively interferes with the GluA2-S880 phosphorylation, suggesting a regulatory role of GluA2-Y876 in LTD induction [42]. GluA2-Y876 phosphorylation levels are determined by the balance between endogenous SFK and protein tyrosine phosphatase activities. The GluA2-Y876 site is predominantly phosphorylated during basal conditions [42] and actively dephosphorylated during mGluR1-mediated LTD induction [43]. The δ2-glutamate receptor (GluD2)-associated tyrosine phosphatase, PTPMEG, actively dephosphorylates the GluA2-Y876 position in vitro [42], and PTPMEG-null mice display impaired motor learning and LTD [44]. By dephosphorylating the GluA2-Y876 site and hence facilitating GluA2-S880 phosphorylation, PTPMEG gates the induction of LTD in the cerebellum [42].
To gain insight into the regulation of AMPAR mobility in cerebellar LTD, we constructed a bidirectional kinetic computational model of PF-PC LTD that emphasises AMPAR trafficking as a dynamic recycling loop, and the role of GRIP, PICK1 and the relevant kinases and phosphatases in maintaining this loop. This is the first model to explicitly account for the dynamic regulation of AMPAR mobility by the interaction of the GluA2-Y876 and GluA2-S880 phosphorylation sites, now known to be a key regulatory switch for PF-PC LTD induction. Our conceptually simple model sheds light on LTD signalling beyond the well-established data showing that PF-PC LTD is dependent on PKC activation, Ca2+ elevation and serine/threonine phosphatase inhibition [19]. We predict that PTPMEG cooperates with PKC to drive LTD expression by gating the effect of PKC on the dissociation of AMPARs from GRIP and thus their availability for binding to PICK1 and internalisation from the postsynaptic membrane. We also show that serine/threonine phosphatase inhibition increases the degree of LTD expression, in line with experimental data [45, 46], and that SFK is not required for the induction of LTD, but negatively regulates LTD expression, as demonstrated experimentally [47]. These results advance our understanding of PF-PC LTD regulation and induction, suggest new hypotheses for experimental validation and provide a platform for further computational studies.
We model AMPARs as embedded at the cell membrane or the endosome, with all interactions with protein partners occurring in the sub-membrane and the ‘sub-endosome’ regions, respectively. These regions constitute the two main compartments of the model, and the bulk cytosol merely acts as a source/sink for smaller molecules. The sub-membrane contains three sub-compartments–the postsynaptic density (PSD), the extra-synaptic area and the endocytic zone, and AMPARs can diffuse laterally between these areas (Fig 1). The recruitment of AMPARs is a three-step process [48] comprising exocytosis at extra-synaptic areas, lateral diffusion to the PSD, and trapping by scaffold proteins (GRIP). Only AMPARs within the endocytic zone can be internalised [49, 50]. Trapping of AMPARs at the endocytic zone by dephosphorylated stargazin (TARP-γ2) is essential for LTD expression [51–53]. In line with this data, LTD is well expressed in our model only when the diffusion rate out of the endocytic zone is kept very low (<0.01s-1).
N-ethylmaleimide-sensitive factor (NSF) interacts with GluA2-containing AMPARs and has an essential role in the recruitment of AMPARs into the postsynaptic membrane, possibly by controlling SNARE-dependent exocytosis [54] or promoting lateral diffusion to the PSD [55, 56]. We model the potentially manifold roles of NSF by requiring that AMPARs are bound to NSF in order to undergo exocytosis [57]. As NSF disrupts the AMPAR-PICK1 interaction [58], and AMPARs bound to GRIP are not available for trafficking, only AMPARs bound to neither PICK1 nor GRIP can bind to NSF [54, 57].
We model AMPAR trafficking exclusively as a recycling loop, and LTD as a perturbation of this dynamic trafficking equilibrium. Therefore, we do not consider de novo synthesis and degradation of AMPARs, whose inclusion is likely to occlude the effect of the phosphorylation switch on AMPAR mobility and LTD expression. Furthermore, degradation of internalised AMPARs does not have functional consequences for the regulation of LTD [59], although the regulation of AMPAR recycling is essential for determining the degree of LTD expression [60].
Many published LTD models are unidirectional and measure LTD expression in terms of AMPAR internalization only, or even simply by the level of AMPAR phosphorylation [61]. This simplifies the modeling strategy but neglects the importance of the endocytosis-exocytosis balance in regulating the cell surface AMPAR population and the dynamic nature of AMPAR recycling. A sophisticated recent stochastic model of cerebellar LTD [35] does account for exocytosis of AMPARs, but disregards all other interactions within the intracellular compartment that are important in regulating AMPAR mobility and reinsertion. Our trafficking pathway is a bidirectional kinetic model (Fig 2) that emphasises AMPAR trafficking as a dynamic recycling loop. As with other models of LTD, we measure LTD expression purely in terms of the reduction of the postsynaptic membrane AMPAR population [35, 61], although additional mechanisms, such as AMPAR desensitisation, may also play a minor role in the biological system [62].
Unique to our model, the dissociation from GRIP, and the mobilisation and availability of AMPARs for trafficking between compartments are regulated by the mutually exclusive phosphorylation of the GluA2-S880 and GluA2-Y876 sites (Fig 3A) [42]. GluA2-S880 is phosphorylated by PKC and dephosphorylated by PP2A, while GluA2-Y876 is phosphorylated by SFKs [42] and dephosphorylated by PTPMEG [42]. Phosphorylation of the GluA2-S880 site abolishes the interaction between the AMPAR and GRIP, allowing PICK1 to bind. PICK1 can also associate with GRIP directly to form a tripartite complex (Fig 3B). The other interactions within the model are detailed in the Methods section.
To observe the effect of PKC and PTPMEG on the endocytic rate alone, we initially selectively blocked exocytosis. Under basal conditions, when PKC is inactive, approximately 125 AMPARs populate the PSD [63] and around 40% of these are estimated to be internalised within 20 minutes [57]. When PKC is activated in the absence of active PTPMEG, the average rate of endocytosis is only slightly elevated relative to basal conditions (44% of AMPARs internalised over 20 minutes with activated PKC, versus 38% when PKC is inactive) (Fig 4A). However, activation of PKC in the presence of active PTPMEG increases the internalisation rate 2-fold above that generated by activated PKC alone, with 89% of AMPARs being internalised over 20 minutes. This suggests that the role of PTPMEG is to gate the effect of active PKC in promoting AMPAR dissociation from GRIP and subsequent internalisation. The result is in agreement with experimental data, which shows that elevated PKC alone does not increase the AMPAR internalisation rate in cerebellar Purkinje cells [64].
According to our LTD model, under basal conditions, the AMPARs trafficked between the cell surface and endosome are predominantly the unphosphorylated and GluA2-Y876-phosphorylated forms. During LTD induction, we expected a shift towards internalisation of the GluA2-S880-phosphorylated form of the receptor as PKC is activated. We reinstated exocytosis and measured the flow of the three different forms of AMPAR (unphosphorylated, GluA2-Y876-phosphorylated and GluA2-S880-phosphorylated) between the plasma membrane and endosomal compartments and vice versa in 3000-second simulations of the system under basal conditions, and during PKC-induced LTD (Fig 4B). Under basal conditions, the cell surface AMPAR population remained stable and only the unphosphorylated form of AMPAR and the GluA2-Y876 phosphorylated form were internalised, each being trafficked at a rate of 0.03–0.04 receptors per second, equally in both directions. When PKC was activated in the presence of PTPMEG, the cell surface AMPAR population declined to 44% of its initial number over around 1000 seconds. This was followed by a steady state during which mainly the GluA2-S880-phosphorylated form of AMPAR was internalised, with 0.16 of these receptors being trafficked per second in both directions, in addition to a small number (0.04–0.06 per second for each) of the unphosphorylated and GluA2-Y876 phosphorylated forms of the receptor (Fig 4C).
When PKC is inactive, the cell surface AMPAR population remains stable, both in the presence and absence of PTPMEG. To analyse the effect of PKC activation, we ran 3000-second model simulations comprising 1000 seconds under basal conditions, followed by a step function activation of PKC that was maintained for the remaining 2000 seconds. This represents the approximate period for which PKC activation is maintained by a positive feedback mechanism during LTD induction, in line with experimental data [35]. As late phase effects maintain LTD after the PKC activation window, we do not consider deactivation of PKC or the maintenance of LTD after this time.
In the absence of PTPMEG, and in agreement with experimental results [42, 44], the activation of PKC does not result in a marked inward trafficking of plasma membrane AMPARs, with the cell surface population of AMPARs only falling to 92% of baseline when PKC is activated. Furthermore, there is no increase in the number of mobile AMPARs (i.e. not bound to GRIP), with fewer than 6% of the AMPARs being mobile during the PKC activation period, as during basal conditions (Fig 5A).
When PTPMEG is present, the activation of PKC leads to an immediate increase in the average percentage of cell surface AMPARs that are mobile from ~6% to ~18% (Fig 5B). This demonstrates cooperation between PKC and PTPMEG to mobilise the cell surface AMPARs for trafficking. Neither PKC activation nor PTPMEG alone is capable of eliciting LTD. Both enzymes are required concurrently, as suggested by experimental data demonstrating that LTD expression in cerebellar Purkinje cells requires PTPMEG activity [42]. The increase in mobile AMPARs during the PKC activation window triggers a decline in the cell surface AMPAR population towards a steady state as endocytosis dominates the trafficking dynamics (Fig 5B).
Experiments have shown that the population of GluA2-Y876-phosphorylated AMPARs declines during LTD induction [42], with the GluA2-S880-phosphorylated form increasing concurrently [36], as plasma membrane AMPARs are mobilised and internalised. Our simulations replicate and quantify this effect (Fig 5C). Under basal conditions in our model, approximately 20% of membrane AMPARs are GluA2-Y876 phosphorylated, with none of the receptors phosphorylated at the GluA2-S880 site. However, immediately upon PKC activation, the population of GluA2-S880-phosphorylated AMPARs increased to 18% of the total PSD AMPAR population, and this was maintained throughout the PKC activation window. Comparable with experimental observations [42], the population of GluA2-Y876 phosphorylated receptors declined from 20% to 9% upon PKC activation (Fig 5C).
It is well established that the inhibition of serine/threonine phosphatase activity accompanies LTD induction [19, 45, 65]. However, whether such inhibition is essential for LTD induction or merely augments is not understood. To study the effects of phosphatase inhibition on LTD induction, we performed simulations for PP2A concentrations ranging between 0–100% (Fig 6 and Table 1). Increasing phosphatase inhibition results in a corresponding increase in the degree of LTD achieved. Without PP2A inhibition, only a 39% reduction in cell surface AMPAR population is achieved after 20 minutes, rising to 77% reduction with 100% PP2A inhibition. This result is comparable to experimental results showing up to a 65% reduction in excitatory postsynaptic current amplitude in cerebellar Purkinje cells using PP2A inhibitors [45], and suggests that tuning of phosphatase inhibition could regulate the degree of depression achieved during LTD.
SFKs selectively phosphorylate the Y876 site of the AMPAR GluA2 subunit [42]. Under basal conditions, phosphorylation at this position limits GluA2-S880 phosphorylation. By allowing GRIP to bind, this stabilises the AMPARs at the cell surface or endosomal membrane. Active PTPMEG dephosphorylates GluA2-Y876, enabling GluA2-S880 phosphorylation and hence the dissociation of the AMPAR from GRIP and its mobilisation for trafficking. We performed simulations under standard LTD induction conditions, in the absence of SFKs, and with increasing SFK concentrations up to 5-fold greater than the basal concentration. Removing SFKs from the system slightly enhanced LTD expression, with 38% of cell surface (PSD) AMPARs remaining after 20 minutes, compared to 44% for the wild-type conditions. Increasing concentrations of SFK caused a proportional decrease in the magnitude of the LTD response, which was directly related to the degree of GluA2-Y876 phosphorylation (Fig 6B and Table 2).
This result is in agreement with experimental studies showing that SFK negatively regulates cerebellar LTD expression [47], although it appears to contradict earlier studies showing that SFKs are essential for LTD expression [66, 67], with SFK inhibitors abolishing LTD. However, the broad-spectrum tyrosine kinase inhibitors used in these studies (i.e. genistein and lavendustin A) are likely to affect kinases other than SFKs [42]. If a more specific SFK inhibitor is used to reduce tyrosine (GluA2-Y876) phosphorylation, LTD induction in cerebellar Purkinje cells is unaffected [42, 47], in agreement with our results. It should be noted that, in vivo, SFKs act on a broad range of substrates and, as such, their effect on AMPAR trafficking, both directly and indirectly, could be more complex than indicated by our model. However, the effect of SFK at the GluA2-Y876 phosphorylation site is sufficient to explain current experimental data.
Knockout of PTPMEG or the PTPMEG-interacting GluD2 abrogates LTD [42] by preventing AMPAR mobilisation. Expression of the mutant subunit, GluA2-Y876F, which cannot be tyrosine phosphorylated, rescues LTD in GluD2-null Purkinje cells [42]. We replicated this result by blocking GluA2-Y876 phosphorylation. Under these conditions, even when PTPMEG was knocked out, LTD was fully expressed (Fig 6C). This demonstrates the central role of GluA2-Y876 phosphorylation in the regulation of AMPAR mobility. The role of SFK activity thus appears to be in limiting AMPAR mobilisation under basal conditions, as well as being an active regulator of PF-PC LTD.
The AMPAR population at the Purkinje cell postsynaptic membrane is part of a continuous dynamic recycling loop. Even when the population is stable, under basal conditions, 90% of the internalised AMPARs are returned to the cell surface within 60 minutes [14]. It is this dynamism that ensures a rapid response to perturbation. Modelling both directions of AMPAR trafficking simultaneously is therefore essential for the accurate study of plasticity. Furthermore, a number of proteins and signalling pathways that regulate receptor internalisation may also affect reinsertion. Consequently, any LTD model that considers only the regulation of internalisation will necessarily be incomplete and may even produce misleading data. In a study of the effects of synaptic activity on AMPAR trafficking in cultured cortical neurons [14], manipulating the rate of AMPAR internalisation–using tetrodotoxin and picrotoxin–had no effect on the size of the cell surface AMPAR population, as the reinsertion rate was similarly affected. It is thus clear that the regulatory mechanisms controlling AMPAR internalisation overlap with those controlling reinsertion. As such, AMPAR trafficking is best described as a unified recycling loop rather than two separate processes.
The balance of kinase and phosphatase activity within cerebellar Purkinje cells is exquisitely poised to allow the AMPAR population to be stabilised at the cell surface and endosome, and yet rapidly mobilised for trafficking during LTD induction. Our simulations show that the GluA2-Y876 and GluA2-S880 phosphorylation sites together act as a ‘master switch’ both for the induction of PF-PC LTD and the regulation of its magnitude. Whilst PTPMEG acts as an overall facilitator of LTD induction, by gating the dissociation of AMPARs from their GRIP anchors, PP2A and SFK activity can tune the degree of depression achieved. This is an important insight that clarifies, and provides a straightforward molecular mechanism for, the role of kinase and phosphatase activity in LTD regulation. Experimental studies have established that PP2A inhibition enhances LTD expression [45], and that SFK activity negatively regulates it [47], in agreement with our simulations. Furthermore, Endo et al [68] produced mutant mice lacking the gene coding for G-substrate, a potent inhibitor of PP2A [69]. Surprisingly, the consequent elevated PP2A levels did not abolish LTD in cerebellar Purkinje cells. Our model explains this result, and demonstrates that PP2A inhibition regulates the magnitude of LTD achieved, but is not required for LTD induction (Table 3).
Although the orphan glutamate receptor δ2 (GluD2) is indispensable for PF-PC LTD expression [72], its specific role remains unclear. However, by binding to and potentially activating PTPMEG, it may concentrate this phosphatase at the plasma membrane and thus facilitate the selective mobilisation of cell surface AMPARs. Whilst GluD2 is only expressed in cerebellar Purkinje cells, several brain regions express GluD1 [73], which may function in a similar manner by binding and/or leading to PTPMEG activation, making this phosphatase a more global regulator of plasticity than currently known. Furthermore, PTPMEG has been shown to bind the NR2A subunit of NMDA receptors [74], which could also support this function.
The signalling pathways regulating synaptic plasticity are complex, both in terms of the number of signalling species involved and their spatiotemporal dynamics. This makes any bidirectional model of trafficking challenging to construct and implement, but essential for generating realistic data. Our model achieves this, is able to replicate a wide range of experimental observations of cerebellar parallel fibre-Purkinje cell LTD, sheds light on their underpinning mechanisms and provides a sound foundation for additional simulation experiments and for more detailed models of synaptic plasticity processes. Furthermore, our model is the first to explore the role of this type of mutually-exclusive phosphorylation switch, which is similar to switches found in other important systems, including receptors controlling insulin response [75], and NMDA receptor function [76].
The model was implemented in the well-established and validated open-source biochemical network simulator COPASI [77, 78], using kinetic parameters obtained from the literature (see supplementary information S1 Table for details). We used deterministic simulation to efficiently and accurately establish the average system behaviour for a wide range of scenarios and parameter ranges [79]. These simulations were performed using the COPASI built-in LSODA (Livermore Solver for Ordinary Differential Equations) solver, with particle number to concentration conversions performed by COPASI. Model can be found in S1 Model.
The model contains two compartments (Fig 1). The sub-membrane compartment comprises the volume of cytosol directly below the plasma membrane to a distance of 120nm [80], and consists of three sub-compartments: postsynaptic density (PSD), endocytic zone (EZ) and extra-synaptic area. The sub-endosome compartment is assumed to occupy the same volume as the sub-membrane. As AMPARs are entirely membrane-bound, they are concentrated in these regions and hence all of the key reactions occur here. The bulk cytosol, which is not explicitly modelled, merely acts as a source/sink for species that are distributed throughout the dendritic spine. Thus, when a species, such as GRIP or PICK1, binds to an AMPAR, it is immediately replaced, by diffusion, by a spare from the bulk cytosol. This approach is supported by experimental and modelling data suggesting that AMPAR scaffolds are never saturated [12]. However, we also produced an alternative model in which GRIP and PICK1 numbers were finite. This model produced results qualitatively the same as those produced with the model used in our paper. The alternative model, together with representative results, is included in the supplementary information S2 Model and S1 Fig.
The complete set of model reactions is summarised in Table 4 and is described below. Except where explicitly stated, these reactions occur in each compartment of the model, between species from the populations in that compartment.
AMPARs exist freely or associated with GRIP or with PICK1, forming an AMPAR-GRIP or AMPAR-PICK1 complex, respectively (Table 4, Reactions 1–6). PICK1 may associate with the GRIP of an AMPAR-GRIP complex and thus a tripartite complex, AMPAR-GRIP-PICK1, can form (Reactions 7–12). A dimeric GRIP-PICK1 complex is not considered, as preliminary experiments showed that it had no effect on the outcome of the simulations. The GRIP populations at the PSD and the endosome interact with AMPARs identically, anchoring the AMPAR to the PSD and the endosomal compartment, respectively [81]. AMPAR-GRIP interactions are not considered in the extra-synaptic area or the endocytic zone. PICK1 is a calcium sensor and the AMPAR-PICK1 binding rate increases 4-fold in the presence of a high calcium concentration [26], as during PF-PC LTD induction. Endosomal AMPARs can also associate with NSF, but only when not associated with either GRIP or PICK1 (Reactions 13 and 14). All binding interactions are assumed to occur with mass action kinetics.
AMPARs not bound to GRIP can diffuse laterally, in both directions, between the PSD and the extra-synaptic area (Reactions 15–20), and between the extra-synaptic area and the endocytic zone (Reactions 21–26). The rate constant for diffusion from one area to another is calculated as the ratio between the diffusion coefficient [12, 48] and the area of the sub-compartment [35].
To undergo endocytosis (Reactions 27 and 28), a GRIP-bound plasma membrane AMPAR must detach from GRIP and bind to PICK1. Furthermore, only AMPARs at the EZ can be internalised. AMPARs can only undergo exocytosis (Reaction 29) when NSF is bound to the receptor, with AMPARs being reinserted into the extra-synaptic area. As we do not consider AMPAR-NSF interactions within the plasma membrane, AMPARs are assumed to detach from NSF when exocytosis occurs.
We adopt a simple switch for activating and deactivating PKC (Reactions 30 and 31), in line with both experimental data [34] and computational simulations [35], which show that positive feedback mechanisms maintain PKC activity for the duration of early LTD induction (at least 20 minutes). PKC can exist freely in the cytoplasm or, when in its active form (PKC*), combined in a reversible complex with PICK1 (Reactions 32 and 33).
PKC* phosphorylates AMPAR at the GluA2-S880 site to generate AMPApS(880) (Reactions 34–37). The PICK1-PKC* complex can also phosphorylate the GluA2-S880 site. Once PICK-PKC* is bound to AMPAR, phosphorylation is assumed to occur at the turnover rate for PKC* (Reactions 38 and 39). The phosphorylation of GluA2-S880 reduces the affinity of the AMPAR for GRIP, as reflected by an increase in the AMPApS-GRIP unbinding rate (Reaction 2) [35]. The AMPAR GluA2-S880 site is dephosphorylated by PP2A, which we assume constitutively active and inhibited (60%) during LTD induction (Reactions 40–45). AMPAR is phosphorylated by SFKs at the GluA2-Y876 site to generate AMPApY(876) (Reactions 46–51). Dephosphorylation of GluA2-Y876 is performed by PTPMEG (Reactions 52–57).
All phosphorylation and dephosphorylation reactions are assumed to occur with Michaelis-Menten kinetics.
Experimentally, under basal conditions, the majority of AMPARs are unphosphorylated [42]. In line with experimental data [63], the system was initially populated with 125 submembrane AMPARs and 125 sub-endosome AMPARs, all unphosphorylated. The kinetics of PTPMEG were calibrated such that the proportion of AMPARs phosphorylated at GluA2-Y876 was consistently approximately 25%, in line with experimental data [42]. However, simulations using alternative initial AMPAR populations–increasing the proportion of GluA2-Y876-phosphorylated AMPARs, for example–did not affect the results obtained, either qualitatively or quantitatively.
Basal conditions were defined as corresponding to PKC inactive, PP2A uninhibited and AMPAR trafficking calibrated by setting the endocytosis rate such that approximately 40% of receptors were internalised over a 20-minute period when exocytosis was selectively blocked [57]. The exocytosis rate was set such that it balanced endocytosis under basal conditions.
When simulating LTD induction, PKC was activated and PP2A was inhibited by 60% throughout the simulation. This inhibition was modelled by removing 60% of the PP2A from the model. For time course simulations, a step function was used to activate PKC (Table 4, Reactions 31 and 32) after allowing the simulation to run for 1000 seconds. As PTPMEG has no effect on LTD induction or expression in the absence of active PKC, PTPMEG was present and active throughout the 3000-second simulation.
To simulate the knockout of specific species (e.g. PTPMEG, Figs 4 and 5), these species were removed from the model.
We carried out standard sensitivity analysis to measure the impact of variations in the model parameters (i.e., the reaction rates from Table 4) on the simulation results. To this end, we established the sensitivity of the steady-state plasma membrane AMPAR population n during LTD induction to changes in each reaction rate ri from Table 4. This involved calculating the scaled sensitivity coefficient of ri as the scaled partial derivative of the AMPAR population n by the reaction rate ri:
SSC(ri)=δnδrinri
The magnitude of the coefficient indicates the sensitivity of the AMPAR population n to changes in the reaction rate ri. The sign of the coefficient indicates whether n increases (SSC(ri) > 0) or decreases (SSC(ri) < 0) in response to an increase in the rate ri.
Table 5 shows these coefficients for the system operating with the rates shown in the supplementary material (S1 Table). Several model parameters have a small (<0.1) scaled sensitivity coefficient, indicating that the model is robust to significant changes in these parameters. The model is sensitive to the remaining parameters:
Experimental data from the literature was used to determine the values for these parameters that the model is sensitive to, as explained in the supplementary information S1 Table.
|
10.1371/journal.pgen.1007780 | Replicative and non-replicative mechanisms in the formation of clustered CNVs are indicated by whole genome characterization | Clustered copy number variants (CNVs) as detected by chromosomal microarray analysis (CMA) are often reported as germline chromothripsis. However, such cases might need further investigations by massive parallel whole genome sequencing (WGS) in order to accurately define the underlying complex rearrangement, predict the occurrence mechanisms and identify additional complexities. Here, we utilized WGS to delineate the rearrangement structure of 21 clustered CNV carriers first investigated by CMA and identified a total of 83 breakpoint junctions (BPJs). The rearrangements were further sub-classified depending on the patterns observed: I) Cases with only deletions (n = 8) often had additional structural rearrangements, such as insertions and inversions typical to chromothripsis; II) cases with only duplications (n = 7) or III) combinations of deletions and duplications (n = 6) demonstrated mostly interspersed duplications and BPJs enriched with microhomology. In two cases the rearrangement mutational signatures indicated both a breakage-fusion-bridge cycle process and haltered formation of a ring chromosome. Finally, we observed two cases with Alu- and LINE-mediated rearrangements as well as two unrelated individuals with seemingly identical clustered CNVs on 2p25.3, possibly a rare European founder rearrangement.
In conclusion, through detailed characterization of the derivative chromosomes we show that multiple mechanisms are likely involved in the formation of clustered CNVs and add further evidence for chromoanagenesis mechanisms in both “simple” and highly complex chromosomal rearrangements. Finally, WGS characterization adds positional information, important for a correct clinical interpretation and deciphering mechanisms involved in the formation of these rearrangements.
| Clustered copy number variants (CNVs) as detected by chromosomal microarray are often reported as germline chromoanagenesis. However, such cases might need further investigation by whole genome sequencing (WGS) to accurately resolve the complexity of the structural rearrangement and predict underlying mutational mechanisms. Here, we used WGS to characterize 83 breakpoint-junctions (BPJs) from 21 clustered CNVs, and outlined the rearrangement connectivity pictures. Cases with only deletions often had additional structural rearrangements, such as insertions and inversions, which could be a result of multiple double-strand DNA breaks followed by non-homologous repair, typical to chromothripsis. In contrast, cases with only duplications or combinations of deletions and duplications, demonstrated mostly interspersed duplications and BPJs enriched with microhomology, consistent with serial template switching during DNA replication (chromoanasynthesis). Only two rearrangements were repeat mediated. In aggregate, our results suggest that multiple CNVs clustered on a single chromosome may arise through either chromothripsis or chromoanasynthesis.
| Structural variants (SVs) contribute to genomic diversity in human [1] and include copy number variants (CNVs) (deletions, duplications), as well as copy number neutral (balanced) variants (inversions and translocations), and more complex rearrangements, resulting from chromothripsis and/or chromoanasynthesis [2,3]. Complex SVs (complex chromosomal rearrangements, CCRs) often result in congenital and developmental abnormalities, as well as in cancer development, although carriers with unaffected phenotypes have also been reported [4].
A rare phenomenon regularly observed in clinical genetic diagnostic laboratories is multiple CNVs co-localizing on the same chromosome. Even though a chromosomal microarray (CMA) may identify such rearrangements, further characterization with whole genome sequencing (WGS) may be useful. A previous WGS study of two closely located duplications revealed additional copy-neutral complex genomic rearrangements associated with paired-duplications, such as inverted fragments, duplications with a nested deletion and other complexities, which were cryptic to CMA [5].
Proposed mechanisms that could explain the formation of multiple CNVs on the same chromosome include chromothripsis and chromoanasynthesis [6,7] while the term chromoanagenesis, a form of chromosome rebirth, describe the two phenomena independent of the underlying mechanism [8].
Chromothripsis is a chromosome shattering phenomenon, where part of or an entire chromosome, or few chromosomes, are fragmented into multiple pieces and reassembled in a random order and orientation resulting in complex genomic rearrangements [9]. During this process, some of the generated fragments can be lost resulting in heterozygous deletions. One of the distinctive features of chromothripsis is that the rearrangement breakpoints (BPs) are localized to relatively small genomic regions, usually spanning a few Mb. The causes of such clustered fragmentations are still unclear, however some studies suggested that chromothripsis could be generated through the physical isolation of chromosomes within micronuclei, where the “trapped” lagging chromosome(s) undergo defective DNA replication and repair, resulting in chromosome pulverization [10,11]. Others hypothesized that the clustered DNA double-strand breaks (DSBs) during chromothripsis could be initiated by ionizing radiation [9,12], breakage-fusion-bridge cycle associated with telomere attrition [9,13], aborted apoptosis [14], as well as endogenous endonucleases [15]. The highly characteristic breakpoint-junction (BPJ) sequences in the derivative chromosomes point to non-homologous end-joining (NHEJ) [16] or microhomology-mediated end-joining (MMEJ) [17] as being likely underlying repair mechanisms for rejoining of the shattered DNA fragments [9,18,19]. Although non-allelic homologous recombination (NAHR) was excluded as a chromothripsis repair mechanism [20], our recent report showed that homologous Alu elements may also mediate germline chromothripsis [15]. Chromothripsis was deciphered by the help of whole genome next generation sequencing technologies (WGS) in microscopic complex chromosomal rearrangements involving three or more BPs [18,19,21,22], as well as in microscopically balanced reciprocal translocations [23,24].
Chromoanasynthesis [25], was described by high resolution chromosome microarray analysis (CMA) and refers to clustered copy number changes, including deletions, duplications, and triplications, that are flanked by regions of normal dosage state. Small templated insertions and microhomologies found at most BPJs pinpointed that chromoanasynthesis likely involves replication failures, such as fork stalling and template switching (FoSTeS) [26] and/or microhomology-mediated break-induced replication (MMBIR) [27]. Another rare but distinct underlying mechanism of formation is atypical chromoanasynthesis that seems to only involve single chromosomes and exclusively generate duplications [28], either clustering on one chromosome arm or scattered throughout the entire chromosome.
It has also been shown that clustered duplications confined to a single chromosome may not only be integrated into the chromosome-of-origin in tandem, but could be integrated at multiple positions in the derivative chromosome and have non-templated insertions at the BPJs, indicating a different mutational mechanism, such as alternative NHEJ mediated by the DNA polymerase Polθ [28]. Finally, evidence suggests that both chromothripsis and replicative errors are not only responsible for highly complex rearrangements involving several chromosomes or a large number of chromosomal segments. Even simpler rearrangements involving a small number of chromosomal segments on a single chromosome could have formed through shattering of a chromosome or replicative errors [21].
To delineate the chromosomes and analyze the plausible underlying mechanisms of formation of multiple CNVs on a single chromosome, we characterized 21 germline complex rearrangements initially detected by CMA. The rearrangements involved only duplications, only deletions or both deletions and duplications. Underlying mechanisms of rearrangement formation were inferred from the BPJ architecture as well as the overall connective picture.
We investigated the BPs of 21 individuals with clustered germline CNVs using WGS (mate-pair or paired-end sequencing) to elucidate potential underlying mechanisms of rearrangement formation and possibly clinically relevant genomic imbalances or gene disruptions. Cases were included if they harbored two or more CNVs on the same chromosome. The clinical symptoms were variable, including congenital malformations and neurodevelopmental disorders. Phenotypes and CMA results are presented in Table 1.
Segregation analysis had been performed in 20 cases and showed that the CNVs were inherited in 8 and de novo in 12. Parental DNA samples for further investigation of parental origin were available in seven of the de novo cases. It was found that the rearrangement was on the maternal chromosome in four cases and on the paternal chromosome in three cases (S1 Table). We also excluded presence of copy number neutral inversions in the parents. Among the eight inherited cases, the rearrangement segregated from a phenotypically unaffected mother (n = 6) or father (n = 2), indicating that the complex chromosomal rearrangement may be an incidental finding. We detected a complex overall picture with 83 BPs associated with deletions, duplications, inversions and insertions (Table 2; S1 Fig; S2 Table). Resolution was on single nucleotide level in 83 BPJs (75%) (Table 2).
In ten cases, two distinct patterns DEL-INV-DEL (n = 4) and DUP-DIP-DUP (n = 6) were observed (DEL, deletion; INV, inversion; DUP, duplication; DIP, diploid). In four of these (P2109_302, P2109_123, P2109_150, P2109_151), the initial CMA suggested a single deletion or duplication and the nature of the rearrangement was resolved with WGS (Table 3). The remaining 11 cases showed unique patterns (Table 3).
Based on the CNV type, all rearrangements were classified into deletions-only group (n = 8), duplications-only group (n = 7) and deletions-and-duplications group (n = 6) (S1 Fig). Examples from each group are presented in Fig 1. The average number of BPJs per case was 4 (range = 2–14). The rearrangements in the duplications-only group contained the fewest BPJs per case (average = 3, range = 2–5) and consisted mostly of DUP-DIP-DUP rearrangements (Table 1). The rearrangements in the deletions-only group contained slightly more junctions (average = 4, range = 2–7). The rearrangements belonging to the deletions-and-duplications group showed the highest degree of complexity with more BPJs per case (average = 6, range = 2–14).
In total, WGS revealed additional duplicated or deleted fragments not detected by CMA in 16 out of 21 cases (76%) (Table 3). In most of the cases, the obtained BPJs allowed us to resolve the exact nature of rearranged chromosomes. For one case (P5513_206) from the duplications-only group, there was no conclusive order for the duplicated fragments, hence three possibilities are shown in Fig 2. In one highly complex case (P1426_301) the full connective picture of rearranged chromosomes could not be established (Fig 3).
In four cases where CMA suggested two clustered duplications separated by a diploid fragment (P4855_511, P2109_150, P06 and P74), WGS revealed a nested deletion within the duplicated segment (S2 Fig). Notably, all these four rearrangements were maternally inherited indicating that the duplication and the deletion are located in cis. In addition, WGS allowed detection of copy-neutral segments (inversions and insertions); and in total, 37 inversions were detected within the clustered CNVs (Table 3). The deletions-only group contains a large number of inverted fragments similar to the deletions-and-duplications group, while the duplications-only group contains only four duplicated fragments with inverted orientation in three cases (P209_151, P4855_512 and P5513_206) (Table 3).
Several OMIM morbid genes were identified in clustered CNVs detected by CMA (S3 Table). A CNV was assessed as pathogenic or likely pathogenic in 11 cases, as benign in one case, and in the remaining cases as variants of unknown significance (Table 1). The pathogenicity classification was based on the American College of Medical Genetics and Genomics (ACMG) guidelines [29] and included the segregation analysis, amount of OMIM morbid genes or specific disease-related genes, size of the CNVs and/or if the CNVs had been reported previously in patients with similar phenotype. None of the CNVs disrupted an OMIM morbid gene but all CNVs that were classified as likely pathogenic or pathogenic was based on gene dosage sensitivity mechanisms. In four cases (P2046_133, P5513_206, P5513_116 and P1426_301) WGS enabled detection of further OMIM morbid genes, which could not be revealed by CMA (S3 Table).
Thirteen of the 21 rearrangements consisted of 36 duplicated fragments (Table 1): 17 of these fragments belong to the duplications-only group (7 individuals) and 19 fragments belong to the deletions-and-duplications group (6 individuals). In all cases, the WGS data analysis could detect whether the duplications were tandem (3 fragments) or interspersed (33 fragments).
Notably, the majority of the duplications were interspersed (92%). There was a single tandem duplication in the duplications-only group (P4855_512) and two tandem duplications in the deletions-and-duplications group (P5371_204 and P2109_176) (Fig 1B). All interspersed duplications were intrachromosomal and 46% of the duplicated fragments were inverted, indicating random orientation of the duplicates. The duplicates of the interspersed duplications clustered tightly: 79% of the duplicates were inserted next to another duplicate. P5513_206 represents such a rearrangement that consists of five interspersed duplications, all inserted in a clustered but seemingly random manner in the same region (Fig 2).
Of the 83 total BPJs, 63 (19 cases) were resolved to single nucleotide resolution (Table 2). SplitVision analyses suggested the following features for the BPJs: novel single nucleotide variants (SNVs) within 1 kb of the BPJ (absent in gnomAD and SweFreq), microhomology, short insertions and repeat elements. Most of the rearrangements contained at least one of these features (S2 Table, Table 2). In total, 30 BPJs (48%) contained microhomology stretches ranging from 2 to 32 nucleotides (median = 2) (S2 Table, S5 Fig, S6 Fig). Even though repeat elements were enriched in BPJs, fusions of similar repeats were only observed in 11 BPJs (13%). The longest stretch of microhomology was 32 nucleotides (P2109_123) and involved homologous Alu associated BPs (Fig 4A). Similarly, all the 11 BPs in P2109_176 contained LINE elements resulting in fusion LINEs at the BPJs (Fig 4B). The most complex case, P1426_301, contained deletions, duplications, and inversions and harbored 25 BPs (14 BPJs) where 16 (64%) were located within repeat regions (Fig 3, S6 Fig). In two cases (P4855_512 and P5371_204), two BPJs harbored novel SNVs within 1 kb of BPJs localized to non-coding regions. Lastly, 10 blunt BPJs were identified in 5 cases (P2046_133, P81, P00, P4855_511, P06) (Table 2, S2 Table, S6 Fig). P2046_133, P81 and P00 belong to the deletions-only group, and P4855_511 and P06 belong to the duplications-only group. No blunt BPJs were found in the deletions-and-duplications group (Table 2). Comprehensive analysis of the BPJ characteristics surrounding the BPJs in all cases and comparisons between the groups are presented in S5 Fig and S6 Fig.
Molecular signatures at the BPJs further enabled the reconstruction of underlying mutational mechanisms. For example, blunt joints, absent or short microhomology (1–4 bp) and small insertions or deletions at the BPJs are characteristic of DNA DSB repair through direct ligation by NHEJ. In the clustered CNVs studied here, we observed that most of the BPJs involved in the deletions-only group showed such signatures (Table 2, S2 Table) pinpointing involvement of NHEJ. Alternatively, DNA DSBs can also be repaired by alternative NHEJ (alt-NHEJ) mechanisms, such as MMEJ which is a more error prone repair pathway highly dependent on microhomology [17]. MMEJ may result in deletions of the DNA regions flanking the original BP, and longer stretches of both templated (sequences found within 100 nucleotides upstream or downstream of the junction) and non-templated (seemingly random nucleotides) insertions at the BPJs. One of the characterized BPJs in P2109_188 has very typical signatures of MMEJ: a 14bp non-templated insertion followed by a 26 bp templated insertion (chr21:45466217–45466242, (-) strand), followed by another 12 bp non-templated insertion, plus 3 bp and 4bp microhomologies at the 5’- and the 3’-sides of the BPJ (S3 Fig). Short stretches of microhomologies (2–3 bp) were also found at other BPJs in the deletions-only group (i.e. P00, P2046_133, P2109_190, P2109_302). It is important to note that these features are also overlapping with features consistent with alt-NHEJ mediated by PARP1, CTIP, MRE11, DNA ligase I/III and polymerase θ (Polθ) [28,30,31], which is associated with short single-strand overhangs after a DSB. This typically leads to inserts of 5–25 bp before ligation and hence leads to short stretches of microhomology seen in the BPJ [31], similar to what is seen in MMEJ. In addition, canonical NHEJ and alt-NHEJ can operate simultaneously in the same cell [32], and this possibility needs to be taken into consideration as well.
Overall, microhomologies were mostly prevalent at the BPJs of the complex rearrangements containing duplications (54% and 59% for duplications-only group and deletions-and-duplications group, respectively) (Table 2, S5 Fig). A model of replication-based mechanisms, for example multiple template switching, could better explain the formation of these complex rearrangements (Fig 3B, Fig 4). Such mechanisms are commonly associated with similar features as MMEJ, as well as de novo single nucleotide variants around the BPJs [33].
Seemingly identical rearrangements on 2p25.3 were identified in individuals P4855_511 (from Sweden) and P06 (from Denmark), belonging to the duplications-only group based on CMA results. However, these two cases were later redefined as having duplication with a “nested” deletion inside the duplicated fragment. An identical blunt BPJ without microhomology (the BPJ of the nested deletion) was detected in both P4855_511 and P06. The duplication junction was resolved at nucleotide level only in P4855_511 and a 3bp microhomology (TGC) was detected at the BPJ through split reads in the deep paired-end data. However, for case P06 no split-read was present for the BPJ showing the duplication in the shallow mate-pair WGS data. Several attempts were made to amplify the BPJ using breakpoint PCR and Sanger sequencing without success due to GC-rich sequences in the area. Hence, we could only compare the junction sequences of one junction, which were identical, including a SNV (rs4971462) in cis upstream of the junction (S4 Fig). This may suggest that the 2p25.3 could be a rare founder variant in Europe. However, using the WGS data from P4855_511 and the Affymetrix Cytoscan HD SNP array data from P06, we analyzed 100 common SNVs surrounding the rearrangement and found that the haplotypes for these variants varied in a way that would be expected for two unrelated individuals. Hence, it was not possible to assess whether the rearrangement in these two individuals have occurred through separate events or in a common ancestor. No evidence suggest that the region is a hotspot for CNV formation, no common repeat structure was present in the BPJs and we also assessed the junction sequence from the common BPJ (S4 Fig) in the Predict a Secondary Structure Web Server (https://rna.urmc.rochester.edu/RNAstructureWeb/Servers/Predict1/Predict1.html) and no significant structure was seen. Remaining rearrangements were all unique.
Finally, the junction architecture may indicate that the nested deletion occurred via non-replicative mechanisms (e.g. NHEJ), which require no microhomology. Although the tandem duplication might occur during replication process, we hypothesize that they occurred within a single cell cycle, as the duplication is co-segregated with deletion in both families.
We and others have previously shown that the sequence homology between Alu elements (average 71%) may facilitate unequal crossover between genomic segments and generate Alu-Alu mediated CNVs, inversions, translocations and chromothripsis [15,34,35]. In the current cohort, DEL-INV-DEL rearrangements on 17p13.3 are associated with fusion Alu–Alu elements at both junctions (P2109_123), suggesting an Alu-Alu mediated mechanism in this complex rearrangement. Sequence identity between the AluSx_AluSx1 and AluSq2_AluSq2 pairs are 73.3% and 78.6%, respectively. Notably, both AluSx_AluSx1 and AluSq2_AluSq2 pairs are in opposite orientation on the reference genome, which resulted in inversion of the fragment C (Fig 4A). As the sequence identity of involved Alu pairs is < 90%, it might not be sufficient for homologous recombination, while MMEJ or FoSTeS/MMBIR could potentially generate Alu-Alu mediated rearrangements here as previously suggested by other studies [34–36]. Indeed, 17p13.3 region is known to be Alu rich and consequently many Alu-Alu mediated CNVs and complex genomic rearrangements associated with multiple disorders have been reported [35]. Similarly, in P2109_176 involving a combination of deletions, duplications and other copy-neutral rearrangements on chromosome 2, we observed LINE elements at all 11 BPs, indicating underlying LINE-mediated mechanisms (Fig 4B). Here, we found 3–5 bp microhomologies at most of the BPJs, indicating replication based FoSTeS/MMBIR mechanisms likely being involved in this case.
Finally, 14 out of 25 BPs in the most complex case (P1426_301) containing deletions, duplications, and inversions are located within repeat regions of different classes likely providing microhomology for multiple template switching (Fig 3).
In the current study we present 21 individuals with two or more clustered non-recurrent CNVs confined to a single chromosome including both chromosomal arms (two cases) or to a single chromosomal arm (19 cases). WGS enabled us to decipher the true nature of the rearrangements including detection of copy neutral variants within or flanking the rearrangements. The individuals had a wide range of clinical symptoms, including congenital malformations and neurodevelopmental disorders. Dosage of the genes located within the deleted and/or duplicated fragments and/or the disruption of genes located in the BPJs could be responsible for the clinical manifestations. In the current cohort, the more exact resolution of WGS as compared to CMA resulted in a reduction of the number of morbid OMIM genes affected in three cases (14%) and in an increase in one individual (5%). However, this information did not influence the overall assessment of the clinical relevance.
WGS analysis revealed additional complexities such as inversions and interspersed duplicates in most cases, findings that are in line with previous findings in a cohort of autism spectrum disorder where 84.4% of large complex SVs involved inversions [3]. In addition, we detected that most of the interspersed duplications were inserted next to another in a seemingly random manner, similar to the few cases reported before [28].
For ultra-complex chromosomal rearrangements such as the ones seen in P1426_301 and P00, the large number of genomic pieces with breakpoints often located in repetitive regions complicates the mapping of the final structure of the derivative chromosome(s). Third-generation sequencing including Pacific Biosciences SMRT long-read sequencing platform or Nanopore MinION sequencing has showed promising results [37,38] for bridging repetitive sequences and hence overcoming one of the largest limitations with short-read sequencing. The current study is limited by the fact that we did not try any of these technologies, which would be the next step needed to completely solve the structure of the derivative chromosomes in this case (P1426_301). Long-read sequencing might also add information in case P5513_206 that is presented here with three possible rearrangements of the duplicated fragments.
By mapping all the BPs and resolving the links between the generated fragments, we observed several hallmarks of germline chromothripsis and chromoanasynthesis [4,25,39]. First, all the BPs associated with the complex rearrangements were clustered and confined to a single chromosome. Second, the rearranged fragments within the derivative chromosomes had random order and orientation. Third, the copy-number states detected in deletions-only group oscillated between one and two, typical to chromothripsis, while the rearrangements including duplications were mostly resembling chromoanasynthesis. Fourth, signatures of NHEJ and MMEJ pathways were mostly detected at the BPJs of the complex rearrangements included in the deletions-only group, which is compatible with the previous reports describing BPJs associated with chromothripsis [9,18,19,32]. Even though both chromothripsis and chromoanasynthesis are generally of paternal origin [6,40], the current de novo chromosomal rearrangements occurred on the maternal and paternal chromosomes to the same extent. Of the seven de novo cases where we had parental samples, three had characteristics of chromoanasynthesis and replicative errors and two of those arose on the maternal chromosome. This is in contrast to the expectation that replicative error-mediated chromosomal aberrations would be biased towards spermatogenic origin. In addition, among the four cases with characteristics of chromothripsis, two were of paternal origin and two of maternal origin. Finally, we confirmed that Alu- or LINE- mediated mechanisms may also underlie chromothripsis formation.
Most of the reported germline chromothripsis cases are nearly dosage-neutral, possibly due to embryonic selection against loss of dosage-sensitive genes. However, there are few reports of heavy imbalances detected by CMA, suggesting chromothripsis event [41–45]. Such cases need further investigations by paired-end or mate-pair sequencing in order to decipher the balanced rearrangements involved as well as to understand the underlying mechanisms. Our approach of applying high-resolution sequencing in such cases with clustered deletions, confirmed that additional copy-neutral SVs may coexist. Combined picture of such complex rearrangements resembled catastrophic phenomenon of chromosome “shattering”, where some of the fragments may be lost (deleted), while retained fragments would be resembled by repair machinery with random order and orientation. The fact that clustered duplications and combinations of deletions and duplications typical to chromoanasynthesis revealed both non-tandem and inverted nature of most duplicates, enriched with microhomologies at the BPJs, further supports the notion that replication based mechanisms, may explain the complex nature of these derivative chromosomes. In summary, we suggest that seven cases in the current study (P2109_190, P72, P2109_302, P2109_123, P2109_188, P81 and P00) represents chromothripsis, ten cases (P06, P4855_511, P2109_150, P2109_151, P74, P4855_512, P5513_206, P2109_162, P5513_116, P5371_204) are chromoanasynthesis events and four cases (P2109_185, P2109_176, P2046_133 and P1426_301) have ambiguous mutational signatures. All four ambiguous cases showed large non-templated insertions in the BPJ (typical to Polθ-driven atypical chromoanagenesis or retrotransposition-mediated chromothripsis), but three cases harbored both duplications and deletions (typical to chromoanasynthesis) and one case contained only deletions (typical to chromothripsis). Of the seven chromothripsis cases, one case was Alu-Alu mediated (P2109_123) and one was likely mediated by replicative errors and the DSBs were joined through alt-NHEJ (P2109_188), while remaining cases showed more consistent signatures of canonical NHEJ or MMBIR. Among the cases involving duplications or both duplications and deletions, most BPJs showed signatures of replicative errors with microhomology in the breakpoints, some possibly caused by repeat elements, except in three cases from the deletions and duplications-group (P2109_185, P2109_176, P1426_301) with non-templated insertions ranging in 8–52 bp in size and short microhomology (2–6 nt) in the BPJs. These features are not fully consistent with replicative joining mechanisms such as FoSTeS/MMBIR, but it is possible that these cases are mediated by replicative errors, and that Polθ is involved in the stitching of the chromosomes, hence two operating repair machineries in the same cell.
In two of the cases in our cohort (P5513_116 and P2109_185) the clustered CNVs were detected on both arms of the chromosomes involved (chromosome X and 5, respectively). Notably, these two cases show similar patterns, where a terminal duplication of one chromosomal arm is inserted in the place of terminal deletion of the other chromosomal arm with an inverted orientation. A breakage-fusion-bridge cycle process could explain parts of this kind of rearrangement. Briefly, the process starts when a chromosome loses its telomere and after replication the two sister chromatids will fuse into a dicentric chromosome [46]. Then, during anaphase the two centromeres will be pulled towards opposite nuclei, resulting in the breakage of the dicentric chromosome. Random breakage may cause large inverted duplications. After the breakage there will be new chromosome ends lacking telomeres resulting in a new cycle of breakage-fusion-bridge, the cycles will stop once the chromosome end acquires a telomere. This mechanism has previously been suggested to explain some cases of chromothripsis formation [9,13,47]. Here, with telomeric regions of both chromosome arms being involved, it is likely that the breakage-fusion-bridge cycle has been accompanied by a formation-attempt of a ring chromosome. However, chromosome analysis and FISH had previously shown that no ring chromosome was formed in either of these cases. In addition, as mentioned previously, case P2109_185 showed characteristics of Polθ involvement in the stitching with large non-templated insertions in the BPJs.
In conclusion, the BP characterization of the derivative chromosomes showed that multiple mechanisms are likely involved in the formation of clustered CNVs, including replication independent canonical NHEJ and alt-NHEJ, replication-dependent MMBIR/FoSTeS and breakage-fusion-bridge cycle, as well as Alu- and LINE-mediated pathways. WGS characterization adds positional information important for a correct interpretation of complex CNVs and for determining their clinical significance; and deciphers the mechanisms involved in formation of these rearrangements.
The local ethical board in Stockholm, Sweden approved the study (approval number KS 2012/222-31/3). This ethics permit allows us to use clinical samples for analysis of scientific importance as part of clinical development. Included subjects were part of clinical cohorts investigated at the respective centers and the current study reports de-identified results that cannot be traced to a specific individual. All subjects have given oral consent to be part of these clinical investigations.
The subjects included in this study (n = 21) were initially referred to the Department of Clinical Genetics at the Karolinska University Hospital (n = 13), Kennedy Center (n = 5), Sahlgrenska University Hospital (n = 2) or Linköping University Hospital (n = 1). All subjects were part of clinical cohorts investigated at respective centers with CMA due to congenital developmental disorders, intellectual disability or autism. Karyotypes and phenotypes are provided in Table 1.
Genomic DNA was prepared from whole blood using standard procedures. CMA was carried out using either SNP (single nucleotide polymorphism) or oligonucleotide microarrays. Fluorescent in situ hybridization (FISH) analysis or quantitative PCR (qPCR) with Power SYBR Green reagents (Applied Biosystems, Carlsbad, CA, USA) was employed to verify the structural variants. FISH-, qPCR-, or array comparative genomic hybridization (aCGH) analysis was used to investigate parental inheritance when possible.
In 13 cases (P2046_133, P2109_123, P2109_150, P2109_151, P2109_162, P2109_188, P2109_190, P2109_302, P4855_511, P4855_512, P2109_176, P1426_301, P2109_185), the CMA was performed with an 180K custom oligonucleotide microarray with whole genome coverage and a median resolution of approximately 18 kb (Oxford Gene Technology (OGT), Oxfordshire, UK). Experiments were performed at the Department of Clinical Genetics at Karolinska University Hospital, Stockholm, Sweden, according to the manufacturer’s protocol. Slides were scanned using an Agilent Microarray Scanner (Agilent Technologies, Santa Clara, CA, USA). Raw data were normalized using Feature Extraction Software (Agilent Technologies, Santa Clara, CA, USA), and log2 ratios were calculated by dividing the normalized intensity in the sample by the mean intensity across the reference sample. The log2 ratios were plotted and segmented by circular binary segmentation in the CytoSure Interpret software (OGT, Oxfordshire, UK). Oligonucleotide probe positions were annotated to the human genome assembly GRCh37 (Hg19). Aberrations were called using a cut-off of three probes and a log2 ratio of 0.65 and 0.35 for deletions and duplications, respectively.
For eight cases (P72, P81, P06, P74, P5513_206, P5513_116, P5371_204, P00) the CMA was performed using an Affymetrix CytoScan HD array and data were analyzed with ChAS software (Affymetrix, Santa Clara, CA, USA) using the following filtering criteria: deletions > 5 kb (a minimum of 5 markers) and duplications >10 kb (a minimum of 10 markers). Patients’ CNV data were reported to ClinVar (P2046_133, P2109_123, P2109_150, P2109_151, P2109_162, P2109_188, P2109_190, P2109_302, P4855_511, P4855_512, P2109_176, P1426_301, P2109_185, P5513_206, P5513_116, P5371_204) or to DECIPHER (P72, P81, P06, P74, P00).
Mate-pair libraries were prepared using Nextera mate-pair kit following the manufacturers’ instructions (Illumina, San Diego, CA, USA). The subjects were investigated with the gel-free protocol where 1 μg of genomic DNA was fragmented using an enzymatic method generating fragments in the range of 2–15 kb. The final library was subjected to 2x100 bases paired-end sequencing on an Illumina HiSeq2500 sequencing platform.
The PCR-free paired-end Illumina WGS data was produced at the National Genomics Infrastructure (NGI), Stockholm, Sweden. The WGS data was generated using the Illumina Hiseq Xten platform, which produced an average coverage of 30X per sample. The average insert size of the WGS libraries was 350 bp, and each read length was 2x150 bp.
The WGS data was aligned to GRCh37 (Hg19) using BWA-mem (version 0.7.15-r1140) [48], and duplicates were marked using Picard tools (http://broadinstitute.github.io/picard/). Structural variant calling was performed using FindSV (https://github.com/J35P312/FindSV), which combines CNVnator [49] and TIDDIT [50]. The variant call format (vcf) files of these two callers were merged and annotated using VEP [51] and filtered against an internal frequency database consisting of 350 individuals. The exact position of the BPs was pinpointed using split reads (S2 Table; cases P2046_133, P2109_123, P2109_150, P2109_151, P2109_162, P2109_188, P2109_190, P2109_302, P4855_511, P4855_512, P2109_176, P5513_116, P5371_204, P1426_301, P2109_185) or Sanger sequencing (cases P00, P06 and P81; Primers and PCR conditions will be provided upon request).
The WGS data and Sanger reads were analyzed for junction features such as microhomology, insertions, single nucleotide variants (SNVs), and repeat elements using blat (https://genome.ucsc.edu/cgi-bin/hgBlat?command=start) and an in-house developed analysis tool dubbed SplitVision (https://github.com/J35P312/SplitVision) (S1 Appendix). In short, SplitVision searches for split reads bridging each BPJ. A consensus sequence of these reads are generated through multiple sequence alignment using ClustalW [52,53] and assembly using a greedy algorithm; maximizing the length and support of each consensus sequence. The consensus sequences are then mapped to the reference genome using BWA. The exact BPs as well as any microhomology and/or insertions at the BPJs are found based on the orientation, position and cigar string of the primary and supplementary alignments of the consensus sequences. Additionally, SplitVision searches for repeat elements and SNVs close to the BPJs (<1 kb). Repeat elements are found using the USCS repeat masker [54] and SNVs are called using SAMtools [55]. Lastly, the SNVs were filtered based on the SweFreq (SweGen Variant Frequency Dataset) [56] and gnomAD (http://gnomad.broadinstitute.org). The allele frequency threshold was set to 0, removing any previously reported SNVs, and SNVs located in regions not covered by the SweGen dataset. The quality of the remaining SNVs was assessed using the Integrative Genomics Viewer (IGV) tool [57].
10X Genomics Chromium WGS was performed on sample P00 at NGI, Stockholm, Sweden. Libraries were prepared using the 10X Chromium controller and sequenced on an Illumina Hiseq Xten platform. Data was analyzed using two separate pipelines developed by 10X Genomics: the default Long Ranger pipeline (https://support.10xgenomics.com/genome-exome/software/downloads/latest) and a custom de novo assembly pipeline based on the Supernova de novo assembler (https://support.10xgenomics.com/de-novo-assembly/software/downloads/latest). The custom de novo assembler pipelines included mapping of raw Supernova contigs with the bwa mem intra-contig mode, as well as extraction of split contigs using a python script (https://github.com/J35P312/Assemblatron).
The bam files of all the sequenced samples indicating SVs are deposited in European Nucleotide Archive (ENA), (S4 Table). Patients’ CNV data are reported to ClinVar (P2046_133, P2109_123, P2109_150, P2109_151, P2109_162, P2109_188, P2109_190, P2109_302, P4855_511, P4855_512, P2109_176, P1426_301, P2109_185, P5513_206, P5513_116, P5371_204) or to DECIPHER (P72, P81, P06, P74, P00). The details of in-house developed analysis tool dubbed SplitVision is provided in S1 Appendix (https://github.com/J35P312/SplitVision).
|
10.1371/journal.pgen.1007040 | Shared genetic regulatory networks for cardiovascular disease and type 2 diabetes in multiple populations of diverse ethnicities in the United States | Cardiovascular diseases (CVD) and type 2 diabetes (T2D) are closely interrelated complex diseases likely sharing overlapping pathogenesis driven by aberrant activities in gene networks. However, the molecular circuitries underlying the pathogenic commonalities remain poorly understood. We sought to identify the shared gene networks and their key intervening drivers for both CVD and T2D by conducting a comprehensive integrative analysis driven by five multi-ethnic genome-wide association studies (GWAS) for CVD and T2D, expression quantitative trait loci (eQTLs), ENCODE, and tissue-specific gene network models (both co-expression and graphical models) from CVD and T2D relevant tissues. We identified pathways regulating the metabolism of lipids, glucose, and branched-chain amino acids, along with those governing oxidation, extracellular matrix, immune response, and neuronal system as shared pathogenic processes for both diseases. Further, we uncovered 15 key drivers including HMGCR, CAV1, IGF1 and PCOLCE, whose network neighbors collectively account for approximately 35% of known GWAS hits for CVD and 22% for T2D. Finally, we cross-validated the regulatory role of the top key drivers using in vitro siRNA knockdown, in vivo gene knockout, and two Hybrid Mouse Diversity Panels each comprised of >100 strains. Findings from this in-depth assessment of genetic and functional data from multiple human cohorts provide strong support that common sets of tissue-specific molecular networks drive the pathogenesis of both CVD and T2D across ethnicities and help prioritize new therapeutic avenues for both CVD and T2D.
| Cardiovascular disease (CVD) and type 2 diabetes (T2D) are two tightly interrelated diseases that are leading epidemics and causes of deaths around the world. Elucidating the mechanistic connections between the two diseases will offer critical insights for the development of novel therapeutic avenues to target both simultaneously. Because of the challenging complexity of CVD and T2D, involving numerous risk factors, multiple tissues, and multidimensional molecular alterations, few have attempted such an investigation. We herein report a comprehensive and in-depth data-driven assessment of the shared mechanisms between CVD and T2D by integrating genomics data from diverse human populations including African Americans, Caucasian Americans, and Hispanic Americans with tissue-specific functional genomics information. We identified shared pathways and gene networks informed by CVD and T2D genetic risks across populations, confirming the importance of well-established processes, as well as unraveling previously under-appreciated processes such as extracellular matrix, branched-chain amino acid metabolism, and neuronal system for both diseases. Further incorporation of tissue-specific regulatory networks pinpointed potential key regulators that orchestrate the biological processes shared between the two diseases, which were cross-validated using cell culture and mouse models. This study suggests potential new therapeutic targets that warrant further investigation for both CVD and T2D.
| Cardiovascular disease (CVD) and type 2 diabetes (T2D) are two leading causes of death in the United States [1]. Patients with T2D are at two to six times higher risk of developing CVD compared to those without T2D [2], indicating the importance of targeting common pathogenic pathways to improve the prevention, diagnosis, and treatment for these two diseases. While decades of work has revealed dyslipidemia, dysglycemia, inflammation, and hemodynamic disturbances as common pathophysiological intermediates for both CVD and T2D [3–5], very few studies have directly investigated the genomic architectures shared by the two diseases. While genetic factors are known to play a fundamental role in the pathogenesis of both CVD and T2D [6], a direct comparison of the top risk variants between these diseases has revealed few overlapping loci in genome-wide association studies (GWAS) from multiple large consortia. Aside from the speculation that the strongest genetic risks may be disease-specific, the agnostic approach requiring the application of strict statistical adjustment for multiple comparisons also increases false negative rate because of the lack of “genome-wide significance”.
To meet these challenges, we and others have previously shown that hidden disease mechanisms can be unraveled through the assessment of the combined activities of genetic loci with weak to moderate effects on disease susceptibility by integrating GWAS with functional genomics and regulatory gene networks [7–11]. Importantly, such high-level integration approaches are able to overcome substantial heterogeneity between independent datasets and extract robust biological signals across molecular layers, tissue types, and even species [8, 12–14]. This advantage is mainly conferred by the aggregation of genetic signals from individual studies onto a comparable ground–molecular pathways and gene networks, before conducting meta-analysis across studies [14, 15]. In other words, even if the genetic variants and linkage architecture can be different between studies, the biological processes implicated are more reproducible and comparable across studies [16]. In the current investigation, we employed a systematic data-driven approach that leveraged multi-dimensional omics datasets including GWAS, tissue-specific expression quantitative trait loci (eQTLs), ENCODE, and tissue-specific gene networks (Fig 1). GWAS datasets were from three well-characterized and high-quality prospective cohorts of African Americans (AA), European Americans (EA), and Hispanic Americans (HA)—the national Women’s Health Initiative (WHI) [8], the Framingham Heart Study (FHS) [17], and the Jackson Heart Study (JHS) [18]. To maximize the reproducibility of our findings across different populations, we also incorporated meta-analyses of CVD and T2D genetics from CARDIoGRAMplusC4D [19] and DIAGRAM [20]. Further, we comprehensively curated functional genomics and gene networks derived from 25 tissue or cell types relevant to CVD and T2D. A streamlined integration of these rich data sources using our Mergeomics pipeline [14, 15] enabled the identification of shared pathways, gene subnetworks, and key regulators for both CVD and T2D across cohorts and ethnicities. Finally, we validated the subnetworks using adipocyte and knockout mouse models, and confirmed their associations with cardiometabolic traits in the Hybrid Mouse Diversity Panel (HMDP) comprised of >100 mouse strains [21–23].
We first investigated whether genetic risk variants of CVD and T2D from GWAS of each cohort/ethnicity were aggregated in a functionally coherent manner by integrating GWAS with tissue-specific eQTLs or ENCODE information and gene co-expression networks that define functional units of genes (Fig 1A). Briefly, co-expression networks were constructed from an array of transcriptomic datasets of various tissues relevant to CVD and T2D (details in Methods). These modules were mainly used to define sets of functionally related genes in a data-driven manner. Genes within the co-expression modules (a module captures functionally related genes) were mapped to single nucleotide polymorphisms (SNPs) that most likely regulate gene functions via tissue-specific eQTLs or ENCODE information. SNPs were filtered by linkage disequilibrium (LD) and then a chi-square like statistic was used to assess whether a co-expression module shows enrichment of potential functional disease SNPs compared to random chance using the marker set enrichment analysis (MSEA) implemented in our Mergeomics pipeline (details in Methods) [14]. Subsequently, meta-analyses across individual MSEA results at the co-expression module level were conducted using the Meta-MSEA function in Mergeomics to retrieve robust signals across studies. Among the 2,672 co-expression modules tested, 131 were found to be significant as defined by false discovery rate (FDR) < 5% in Meta-MSEA across studies (Table 1, S1 Table). Moreover, the majority of the disease relevant tissues or cell types included in our analysis yielded informative signal, supporting the systemic pathogenic perturbations known for CVD and T2D (S1 Fig). Of the significant modules identified, 79 were associated with CVD and 54 with T2D. Two modules were associated with both diseases, with one enriched for “carbohydrate metabolism” genes and the other over-represented with “other glycan degradation; known T2D genes” (Fig 2A, S1 Table). Examination of these two shared modules showed that the genetic signals driving the module significance were largely different between CVD and T2D, with 14.8% lead SNPs overlapping for the carbohydrate metabolism module and 5.8% lead SNPs overlapping for the glycan degradation module between diseases. These results indicate that the GWAS signals for the two diseases in each module do not necessarily overlap, but the CVD and T2D genes are likely functionally connected since they are co-expressed in the same modules and annotated with coherent functions. Additionally, the majority of the CVD modules and T2D modules were identified in more than one ethnic group based on MSEA analysis of individual studies, supporting consistency across ethnicities (Fig 2B).
Apart from the two directly overlapping modules, between the CVD- and T2D-associated modules there were many overlapping genes, indicating additional shared functions that contribute to both diseases (S2 Fig). Upon annotating the disease-associated modules using functional categories curated in Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome while correcting for the overlaps between pathways (method details in S1 Text; S3 Fig; S2 Table), we found significant functional overlaps between the CVD and T2D modules (overlap p = 3.1e-15 by Fisher’s exact test, Fig 2C). We further ranked all the enriched functional categories by the number of CVD/T2D modules that were annotated with each functional term (Fig 3), which showed a wide spectrum of biological processes shared by both CVD and T2D across ethnicities and cohorts. Of the top ranked processes for the significant co-expression modules identified, we observed well-established pathogenic processes such as lipid and fatty acid metabolism [24], glucose metabolism [25], oxidation [26], and cytokine signaling [27]. Pathways previously implicated mainly for T2D such as beta-cell function were also found to be shared for both CVD and T2D. Interestingly, our completely data-driven approach also identified extracellular matrix (ECM) and branched chain amino acids (BCAA) metabolism as top functional categories whose roles in the development of cardiometabolic disorders have only been implicated in recent experimental work [28–30]. Furthermore, our analysis also revealed under-appreciated processes involving the neuronal system and transport of small molecules.
The coexpression networks used above mainly served to capture coexpression patterns between genes and to define data-driven gene sets that contain functionally related genes, but they do not carry detailed topology information on gene-gene regulatory relationships. To dissect the gene-gene interactions within and between the 131 disease-associated modules, and to identify key perturbation points shared for both CVD and T2D modules, we used the GIANT networks [31] and Bayesian networks (BNs) from 25 CVD and T2D relevant tissue and cell types, which provide detailed topological information on gene-gene regulatory relationships necessary for the downstream network analysis. The BNs used in our study were generated using similar sets of mouse and human gene expression datasets as used for the co-expression networks, but additionally incorporated genetic data to model causal gene regulatory networks, whereas the GIANT networks were derived based on independent gene expression datasets and protein interaction information. We included both types of gene regulatory networks to increase the coverage of functional connections between genes and only considered KDs identified in both to enhance the robustness of KD prediction.
Specifically, all genes in each of the co-expression modules genetically associated with CVD or T2D as identified in our study were mapped onto the GIANT and BN graphical networks to identify KDs using the weighted key driver analysis (wKDA) implemented in Mergeomics [14], where KDs were defined as genes whose local network neighborhoods demonstrate significant enrichment of genes from disease-associated modules (details in Methods; concept depicted in S4 Fig). Of note, wKDA gives higher weight to network edges that are consistent across network models constructed from independent studies, therefore alleviating potential bias caused by dataset heterogeneity. We identified 226 KDs that were consistently captured in Bayesian and GIANT network at Bonferroni-corrected p-value < 0.05 (Fig 1B), among which 162 were KDs for both CVD and T2D associated modules. Bonferroni-correction was used here to focus on the strongest KDs for prioritization purposes. To further prioritize these 162 shared KDs, tissue-specific subnetworks of these KDs were evaluated using Meta-MSEA to rank the magnitude of their genetic association with CVD and T2D across cohorts, yielding 15 top-ranked KDs at FDR<10% in Meta-MSEA for CVD and T2D separately (combined FDR<1% for both diseases simultaneously) (Fig 1B, Table 2). The top KD subnetworks were related to similar pathogenic processes highlighted in the previous section, including cholesterol biosynthesis, respiratory electron transport, immune system and ECM. We further inferred the directionality of the effects of each specific KD on both diseases using GWAS signals mapped to each KD based on eQTLs or chromosomal distance (details in Methods; results in S5 Fig). This analysis differentiated the KDs into those showing consistent direction of association for both CVD and T2D (ACLY, CAV1, SPARC, COL6A2, IGF1), inverse directions with CVD and T2D (HMGCR, IDI1), and uncertain directions (Table 2). Therefore, the shared KDs do not necessarily affect the risks for the two diseases in the same direction.
The KDs and subnetworks were identified based on the full spectrum of genetic evidence (from strong to moderate and subtle) from the various GWAS datasets examined in the current study. To assess whether the top KD subnetworks were enriched for previously known disease genes that mostly represent the strong and replicated genes as a means of cross-validation, we manually curated previously reported genes associated with CVD, T2D, and intermediate metabolic traits related to CVD, T2D (glucose, insulin, lipids, obesity) from DisGeNET [32] and the NHGRI GWAS Catalog [6] (Fig 1C, genes listed in S3 Table). The connection between the top 15 KDs and known genes for CVD, T2D and relevant cardiometabolic traits was confirmed by the significant over-representation of the known disease genes in KD subnetworks, with fold enrichment as large as 8, confirming the strong biological importance of these KDs (Fig 4A). Further, the top 15 KDs showed direct connections to 28 GWAS hits reaching genome-wide significance (p < 5e-8) for CVD and 16 for T2D, which account for 35% (fold = 3.35, p = 7.18e-10) and 22% (fold = 2.16, p = 8.08e-4) of all reported significant GWAS signals for CVD and T2D in GWAS catalog, respectively. Two of the 15 top KDs, namely HMGCR and IGF1, were previously identified as signals of genome-wide significance for obesity, lipids and T2D, all risk factors of CVD. Additionally, network visualization revealed tissue-specific KDs and interactions of CVD and T2D genes in many disease-relevant tissues including adipose, adrenal gland, artery, blood, digestive tract (small intestine, colon), hypothalamus, islet, liver, lymphocyte, skeletal muscle, thyroid, and vascular endothelium (Fig 4B). PCOLCE represents an intriguing hypothalamus KD that interacts with important energy homeostasis genes like leptin receptor LEPR, suggesting a role of neurohormonal control in CVD and T2D pathogenesis. In contrast, CAV1 appeared to interact extensively with other KDs in peripheral tissues, especially in the adipose tissue.
CAV1 is a robust KD for CVD- and T2D-associated modules across multiple tissues, with the adipose tissue subnetwork of CAV1 containing the largest number of neighboring genes (Fig 4B). In addition, adipose tissue is the only tissue where CAV1 is a KD in both the Bayesian networks and GIANT networks. These lines of evidence implicate the potential importance of CAV1 adipose subnetwork in the shared pathogenesis for both diseases. Indeed, Cav1-/- mice have been shown to alter the lipid profile, susceptibility to atherosclerosis, and insulin resistance [33, 34]. To assess whether perturbation of this potential KD induces changes in the subnetwork genes as predicted by our network modeling, we performed validation by conducting siRNA-mediated knock down of Cav1 in differentiating mouse 3T3-L1 adipocytes and by evaluating the whole transcriptome alteration in mouse gonadal adipose tissue between wild type and Cav1-/- mice [33] (Fig 1C; details in Methods). Of the 12 adipose network neighbors of Cav1 that were tested in vitro, 6 exhibited significant changes in expression level on day 2 after ~60% Cav1 knockdown using two siRNAs against Cav1. In contrast, none of the 5 negative controls, which were randomly selected among adipocyte genes that are not connected to Cav1 or its first level neighbors in the adipose network, were affected after Cav1 perturbation (Fig 5A). Cav1 knockdown also led to decreased expression of Pparg, a major adipocyte differentiation regulator (S6 Fig), supporting a role of Cav1 in adipocyte differentiation as previously observed [35].
In 3-month-old Cav1-/- mice which showed perturbed lipid and insulin sensitivity profiles, we observed 1,474 differentially expressed genes (DEGs) at FDR<1%. We found that the first and second level neighbors of CAV1 in our predicted subnetwork showed significant enrichment for DEGs in adipose tissue induced by Cav1 knockout, with the degree of fold enrichment increasing as the statistical cutoff used to define DEGs became more stringent (Fig 5B; subnetwork view with DEGs in S7 Fig). On the contrary, the third and fourth level neighbors of CAV1 in our predicted subnetwork did not exhibit such enrichment of DEGs (Fig 5B). These experimental findings support that CAV1 is a key regulator of the subnetwork and the network structure predicted by our network modeling is reliable, although it is difficult to discern whether the network changes are related to alterations in adipocyte differentiation status. We also observed strong enrichment for the focal adhesion pathway in both the predicted Cav1 adipose subnetwork (p = 9.6e-14 by Fisher’s exact test, fold enrichment = 6.0) and the differential adipose genes in Cav1-/- mice (p = 6.9e-9, fold enrichment = 3.5).
We further assessed the transcriptomic profiling in adipose (relevant to T2D and CVD) and aorta tissue (main site of CVD) in relation to 7 cardiometabolic phenotypes including adiposity, lipid levels (triglyceride, LDL, HDL), fasting glucose, fasting insulin and HOMA-IR, across >100 mouse strains in two HMDP panels [21–23]. HMDP is a systems genetics resource that comprises more than 100 commercially available mouse strains differing in genetic composition, and has emerged as a power tool to study complex human diseases [22, 36]. The biological relevance of HMDP to human pathophysiology has been reproducibly demonstrated [37–39]. Moreover, HMDP data was completely independent of the human-focused genetic datasets and the network datasets used in our primary integrative analysis (Fig 1C). Here we selected two specific HMDP panels, high-fat (HF) and atherogenic (ATH), in which mice were either fed with a high-fat high-sucrose diet or underwent transgenic expression of human APOE-Leiden and CETP gene as a pro-atherogenic background, respectively. These two panels were chosen for their representativeness of human T2D (the HF panel) and CVD (the ATH panel) pathology. First, we investigated the correlation between the expression of 14 top KDs (no probe for KD MSMO1 in HMDP) and cardiometabolic traits in the adipose and aorta tissues assessed in HMDP. All 14 KDs displayed significant trait association in HMDP, with the association for 11 KDs replicated in both the HF and ATH HMDP panels (Fig 6A). Next, we retrieved the adipose and aorta gene-trait correlation statistics for the top KD subnetwork genes, and used MSEA to test whether genes in the KD subnetworks displayed an overall overrepresentation of strong trait association in HMDP. Again, the 14 KD subnetworks showed significant trait association after Bonferroni correction (Fig 6B). These findings support that the close involvement of the KDs in cardiometabolic trait perturbation we predicted based on human datasets can be cross-validated in mouse models.
Cav1 knockout in mice led to dysreuglation of the predicted subnetwork (Fig 5B) and significant alterations in cardiometabolic phenotypes [33, 34], supporting the causal role of CAV1 in both CVD and T2D. To further investigate the potential causal role of the top KDs and their subnetworks in CVD and T2D, we conducted integrative analysis of the KD subnetworks to assess their disease association using GWAS results for the 7 cardiometabolic traits from HMDP and tissue-specific cis-eQTLs (Fig 1C). By mapping GWAS signals to genes using adipose or aorta eQTLs and testing for enrichment of genetic association with cardiometabolic traits within the KD subnetwork genes using MSEA, we found consistent and significant association between cardiometabolic traits and the subnetworks of KDs ACAT2, CAV1, COL6A2, IGF1, PCOLCE, and SPARC across adipose and aorta (Fig 6C). These results informed by mouse GWAS support a potential causal role of these top KDs in perturbing gene networks in multiple tissues to trigger CVD and T2D.
CVD and T2D are highly correlated complex diseases and share many common risk factors. Multiple genetic variants may individually exert subtle to strong effects on disease pathogenesis, and in aggregate perturb diverse pathogenic pathways [8, 9, 13, 19, 20, 40]. In this systems-level, data-driven analysis of GWAS from several large and high-quality cohorts of diverse ethnicities, integrated with functional data (from ENCODE, eQTLs, tissue-specific co-expression and regulatory networks constructed from human and mouse experiments), we identified both known and novel pathways and gene subnetworks that were genetically linked to both CVD and T2D across cohorts and ethnicities. Further, KDs in tissue-specific subnetworks appear to regulate many known disease genes for increased risk of CVD and T2D. Lastly, we experimentally validated the network topology using in vitro adipocyte and data from in vivo gene knockout models, and confirmed the role of the top KDs and subnetworks in both CVD and T2D traits in independent sets of mouse studies.
The data-driven nature of the current study offers several strengths. First, we incorporated the full-scale of genetic variant-disease association from multiple cohorts, ethnicities and disease endpoints, allowing for the detection of subtle to moderate signals, as well as comparison and replication of results across diseases and populations. More importantly, by focusing on results that demonstrate consistent significance at pathway and network level, we overcome the difficulties in harmonizing independent datasets that are complicated by substantial heterogeneity due to platform differences and population substructure. This is because disease signals across populations are more conserved at pathway level than at individual variant and gene levels [12, 14, 16]. Second, the comprehensive incorporation of tissue-specific eQTLs, coupled with the use of tissue-specific networks, enhances our ability to achieve better functional mapping between genetic variants and genes, and uncover systems-level regulatory circuits for CVD and T2D in a tissue-specific fashion. Third, data-driven modules and networks used in this study increase the potential for novel discovery as gene-gene interactions are defined by data rather than prior knowledge. As the network models were from many independent studies reflecting diverse physiological conditions, leveraging these datasets and network models offers more comprehensive coverage of biological interactions than any given dataset can provide and has proven a valuable approach to unveil novel biological insights [9, 13, 41]. While some of our findings confirmed those from previous canonical pathway-based analysis on disease processes including ECM-receptor interaction and cell-adhesion, and KDs such as SPARC [8], our data-driven approach in the current study uncovered numerous novel genes, pathways, and gene subnetworks. A likely reason for the enhanced discovery potential of co-expression modules is that several interacting pathways could be co-regulated in a single module, or a pathway could interact with other poorly annotated processes in a module to together confer disease risk. The use of modules capturing such interactions improves the statistical power, in contrast to testing the pathways individually. Lastly, we conducted cross-validation studies in support of the functional roles of specific KDs and subnetworks in CVD and T2D using independent experimental models.
We acknowledge the following limitations in our study. First, our analyses were constrained by the coverage of functional datasets that are currently available, which causes uneven tissue coverage between data types and statistical bias towards more commonly profiled tissues such as adipose and liver, making it difficult to achieve precise inference for all relevant tissues. Although we believe this does not necessarily undermine the validity of the main findings from our study, we acknowledge that we likely have missed relevant biology from tissues with fewer studies and smaller sample sizes. Further investigation is needed when additional relevant datasets become available. Secondly, our FDR estimates in MSEA do not take into consideration the gene overlap structure among co-expression modules, due to the challenge in accurately adjusting for the various degrees of overlaps between module pairs. To alleviate this limitation, we focus on modules and pathways demonstrating consistency across datasets and merge overlapping modules subsequently. Thirdly, although we conducted validation experiments on the CAV1 subnetwork in both in vitro and in vivo models and cross-validated the importance of the predicted top key drivers and subnetworks in two independent large-scale mouse population studies, further experiments are warranted to thoroughly test the causality of the predicted KDs and elucidate the detailed tissue-specific mechanisms of the KDs on CVD and T2D. This is particularly important considering the limited overlaps in the modules and KDs identified from our study and the ones identified in two recent multi-tissue network analysis of cardiometabolic diseases [10, 11]. Only 7 KDs overlapped including APOA1, CD2, CEBPD, CENPF, CSF1R, CTSS, UBE2S. Methodological differences in network inference and key driver analysis and differences in the pathophysiological conditions of the study populations could explain the discrepancies. Lastly, ethnic-specific and sex-specific mechanisms await future exploration.
There are several direct implications that can be drawn from the results of our integrative analyses of both observational and experimental data. First, it appears that pathogenic pathways for CVD and T2D are indeed common in ethnically diverse populations. These shared pathways capture most of the critical processes that have been previously implicated in the development of either T2D or CVD, including metabolism of lipids and lipoproteins, glucose, fatty acids, bile acids metabolism, biological oxidation, coagulation, immune response, cytokine signaling, and PDGF signaling. Second, BCAA metabolism and ECM are among the top and common pathways identified. Our finding on BCAA is consistent with recent work relating serum levels of BCAA to risk of CVD and T2D in large prospective cohorts [42, 43], although whether BCAA is a “pathophenotype” or strong pathogenic factor has been debated [28, 44]. Our findings support a causal role of BCAA because 1) both CVD and T2D risk variants were enriched in the co-expression modules related to BCAA degradation, and 2) 15 genes in the BCAA pathway were part of the top KD subnetworks, representing a significant enrichment of BCAA genes (fold enrichment = 3.02, Fisher’s exact test p = 1.4e-5). Of note, BCAA genes themselves carry few genetic risk variants for CVD and T2D, albeit their network neighboring genes are highly enriched for disease variants, which may result from negative evolutionary pressure due to the critical role of BCAA. More recently, Jang and colleagues have shown BCAA catabolism can cause insulin resistance, providing further support for the causal role of BCAA for both CVD and T2D [45]. Our finding on the role of ECM in both CVD and T2D is also in line with recent reports [8, 13, 29, 30, 46]. In the top enriched subnetworks, ECM genes appear to exert strong effect (Fig 4B) coordinating other processes such as cholesterol metabolism, energy homeostasis, and immune response across a wide range of peripheral tissues and endocrine axis. This substantiates the importance of ECM modeling as a mechanistic driver for CVD and T2D.
Secondly, our comprehensive network modeling identified critical disease modulators and key targets whose functional roles were subsequently supported by multiple cross-validation efforts. This supports the use of network modeling to unravel and prioritize promising top targets that may have high pathogenic potential for both CVD and T2D. The KDs we identified can be considered as “highly confident” for the following reasons: 1) they are KDs for both CVD and T2D associated modules, 2) the tissue-specific subnetworks of these KDs show significant and replicable association with both diseases, 3) their subnetworks are highly enriched with known CVD and T2D genes, 4) in vitro siRNA knockdown and in vivo knockout mouse experiments confirm the role of a central KD CAV1 in regulating the downstream genes as predicted in our network model, and 5) both the expression levels of KDs and the genetic variants mapped to the KD subnetworks are significantly associated with CVD and T2D relevant traits in independent mouse populations with naturally occurring genetic variations.
Thirdly, most KDs are not GWAS signals reaching genome-wide significance, nor are they rare-variant carrying genes, indicating that standard genetic studies miss important genes that orchestrate known CVD and T2D genes. The phenomenon may reflect a negative evolutionary pressure experienced by the KDs due to their crucial functions. In support of this hypothesis, we found a significant enrichment of human essential genes lacking functional variations among the 162 KDs identified in our study [47] (Fold = 1.41, p = 9.02e-3). This is consistent with previous findings [8, 9, 13] reaffirming the power and reliability of our approach in uncovering hidden biological insights particularly in a systematic integrative manner.
The connections between KDs and other disease genes revealed by our study warrant future investigation into the potential gene-gene interactions. Indeed, a closer examination of the biological functions from the top shared KDs further corroborates their disease relevance. For instance, our network modeling identified HMGCR as a top KD, consistent with its primary role as the target for cholesterol-lowering HMG-CoA inhibitors, namely statins. Our directionality inference analysis indicates that HMGCR is associated with CVD and T2D in opposite directions. This is consistent with the recent findings that genetic variations in HMGCR that decrease CVD risk cause slightly increased T2D risk, and statin drugs targeting HMGCR reduces CVD risk but increases T2D risk [48–50]. CAV1 and IGF1 represent two tightly connected multi-functional KDs. CAV1 null mice were found to have abnormal lipid levels, hyperglycemia, insulin resistance and atherosclerosis [33, 34]. Consistent with these observations, we found strong association of CAV1 expression levels as well as CAV1 network with diverse cardiometabolic traits in both human studies and mouse HMDP panels. Our data-driven approach also revealed the central role of CAV1 in adipose tissue by elucidating its connection to a large number of CVD and T2D GWAS genes and to genes involved in focal adhesion and inflammation (Fig 4), which could drive adipocyte insulin resistance [51, 52]. The regulatory effect of CAV1 on neighboring genes was subsequently validated using in vitro adipocyte and in vivo mouse models. Moreover, our network modeling also captured the central role of CAV1 in muscle and artery tissues, suggesting multi-tissue functions of CAV1 in the pathogenic crossroads for CVD and T2D. The other multi-functional KD, IGF1, is itself a GWAS hit for fasting insulin and HOMA-IR. Despite being primarily secreted in liver, in our study IGF1 demonstrated an adrenal gland and muscle specific regulatory circuit with CVD and T2D genes, suggesting that it may confer risk to these diseases through the adrenal endocrine function and muscle insulin sensitivity. The three ECM KDs we identified, SPARC, PCOLCE and COL6A2, were especially interesting due to their consistent and strong impact on diverse cardiometabolic traits shown in our cross-validation analyses in HMDP (Fig 4, Fig 6). SPARC encodes osteonectin, which is primarily circulated by adipocytes. It inhibits adipogenesis and promotes adipose tissue fibrosis 50. SPARC is also associated with insulin resistance and coronary artery lesions 51, 52. PCOLCE (procollagen C-endopeptidase enhancer) represents a novel regulator for hypothalamus ECM that could potentially disrupt the neuroendocrine system. COL6A2, on the other hand, provides new insights into how collagen may affect cardiometabolic disorders: in adrenal tissue COL6A2 is connected to IGF1R, the direct downstream effector for KD IGF1. Importantly, our directionality analysis suggests that while some KDs such as CAV1 may have similar directional effects on CVD and T2D, cases like HMGCR that show opposite effects on these diseases are also present. Therefore, it is important to test the directional predictions to prioritize targets that have the potential to ameliorate both diseases and deprioritize targets with opposite effects on the two diseases.
In summary, through integration and modeling of a multitude of genetics and genomics datasets, we identified key molecular drivers, pathways, and gene subnetworks that are shared in the pathogenesis of CVD and T2D. Our findings offer a systems-level understanding of these highly clustered diseases and provide guidance on further basic mechanistic work and intervention studies. The shared key drivers and networks identified may serve as more effective therapeutic targets to help achieve systems-wide alleviation of pathogenic stress for cardiometabolic diseases, due to their central and systemic role in regulating scores of disease genes. Such network-based approach represents a new avenue for therapeutic intervention targeting common complex diseases.
Detailed GWAS information including sample size, ethnicity and genotyping platform was described in S4 Table and S1 Text. Briefly, p-values of qualified single nucleotide polymorphisms (SNPs) at minor allele frequency > 0.05 and imputation quality > 0.3 for CVD and T2D were collected for all available GWAS datasets (WHI-SHARe, WHI-GARNET, JHS, FHS, CARDIoGRAMplusC4D [19], and DIAGRAM [20]). SNPs meeting the following criteria were further filtered out: 1) ranked in the bottom 50% (weaker association) based on disease association p-values and 2) in strong linkage disequilibrium (LD) (r2 > 0.5) according to ethnicity-specific LD data from Hapmap V3 [53] and 1000 Genomes[54]. For each GWAS dataset, LD filtering was conducted by first ranking SNPs based on the association p values and then checking if the next highest ranked SNP was in LD with the top SNP. If the r2 was above 0.5, the SNP with lower rank was removed. The step was repeated by always checking if the next SNP was in LD with any of the already accepted ones.
Using the Weighted Gene Co-expression Network Analysis (WGCNA)[55], we constructed gene co-expression modules capturing significant co-regulation patterns and functional relatedness among groups of genes in multiple CVD- or T2D-related tissues (including aortic endothelial cells, adipose, blood, liver, heart, islet, kidney, muscle and brain) obtained from nine human and mouse studies (S5 Table). Modules with size smaller than 10 genes were excluded to avoid statistical artifacts, yielding 2,672 co-expression modules. These coexpression modules were used as a collection of data-driven sets of functionally connected genes for downstream analysis. The potential biological functions of each module were annotated using pathway databases Reactome and KEGG, and statistical significance was determined by Fisher’s exact test with Bonferroni-corrected p< 0.05.
eQTLs establish biologically meaningful connections between genetic variants and gene expression, and could serve as functional evidence in support of the potential causal role of candidate genes in pathogenic processes[56, 57]. We therefore conducted comprehensive curation for significant eQTLs in a total of 19 tissues that have been identified by various consortia (including the Genotype-Tissue Expression (GTEx) [58], Muther [59] and Cardiogenics [60], and additional independent studies; S6 Table). Additional functional genomics resources from ENCODE were also curated to complement the eQTLs for SNP-gene mapping (S1 Text).
MSEA was used to identify co-expression modules with over-representation of CVD- or T2D-associated genetic signals for each disease GWAS in each cohort/ethnicity in a study specific manner. MSEA takes into three input: 1) Summary-level results of individual GWAS (WHI, FHS, JHS, CARDIoGRAM+C4D, DIAGRAM) for the LD-filtered SNPs; 2) SNP-gene mapping information, which could be determined by tissue-specific cis-eQTLs, ENCODE based functional annotation and chromosome distance based annotation. Cis-eQTLs is defined as eQTLs within 1MB of the transcription starting sites of genes. For ENCODE, we accessed the Regulome database and used the reported functional interactions to map SNPs to genes by chromosomal distance. Only SNPs within 50kb of the gene region and have functional evidence in Regulome database were kept; 3) Data-driven co-expression modules from multiple human and mouse studies as described above. Tissue-specificity was determined by the tissue origins of eQTLs and ethnic specificity was determined by the ethnicity of each GWAS cohort, since the human disease genetic signals and human eQTL mapping were the main driving factors to determine the significance of the modules. MSEA employs a chi-square like statistic with multiple quantile thresholds to assess whether a co-expression module shows enrichment of functional disease SNPs compared to random chance [14]. The varying quantile thresholds allows the statistic to be adoptable to studies of varying sample size and statistical power. For the list of SNPs mapped to each gene-set, MSEA tested whether the SNP list exhibited significant enrichment of SNPs with stronger association with disease using a chi-square like statistic: χ=∑i=1nOi−EiEi+κ, where n denotes the number of quantile points (we used ten quantile points ranging from the top 50% to the top 99.9% based on the rank of GWAS p values), O and E denote the observed and expected counts of positive findings (i.e. signals above the quantile point), and κ = 1 is a stability parameter to reduce artefacts from low expected counts for small SNP sets. The null background was estimated by permuting gene labels to generate random gene sets matching the gene number of each co-expression module, while preserving the assignment of SNPs to genes, accounting for confounding factors such as gene size, LD block size and SNPs per loci. For each co-expression module, 10000 permuted gene sets were generated and enrichment P-values were determined from a Gaussian distribution approximated using the enrichment statistics from the 10000 permutations and the statistics of the co-expression module. Finally, Benjami-Hochberg FDR was estimated across all modules tested for each GWAS.
To evaluate a module across multiple GWAS studies, we employed the Meta-MSEA analysis in Mergeomics, which conducts module-level meta-analysis to retrieve robust signals across studies. Meta-MSEA takes advantage of the parametric estimation of p-values in MSEA by applying Stouffer’s Z score method to determine the meta-Z score, then converts it back to a meta P-value. The meta-FDR was calculated using Benjamini-Hochberg method. Co-expression modules with meta-FDR < 5% were considered significant and included in subsequent analyses.
We used graphical gene-gene interaction networks including the GIANT networks [31] and Bayesian networks (BN) from 25 CVD and T2D relevant tissue and cell types (S7 Table, S1 Text) to identify KDs. If more than one dataset was available for a given tissue, a network was constructed for each dataset and all networks for the same tissue were combined as a union to represent the network of that tissue, with the consistency of each network edge across datasets coded as edge weight. The co-expression modules genetically associated with CVD or T2D identified by Meta-MSEA were mapped onto these graphical networks to identify KDs using the weighted key driver analysis (wKDA) implemented in Mergeomics [14]. wKDA uniquely consider the edge weight information, either in the form of edge consistency score in the case of BNs or edge confidence score in the case of GIANT networks. Specifically, a network was first screened for suitable hub genes whose degree (number of genes connected to the hub) is in the top 25% of all network nodes. Once the hubs have been defined, their local one-edge neighborhoods, or “subnetworks” were extracted. All genes in each of the CVD and T2D-associated gene sets that were discovered by meta-MSEA were overlaid onto the hub subnetworks to see if a particular subnetwork was enriched for the genes in CVD/T2D associated gene sets. The edges that connect a hub to its neighbors are simplified into node strengths (strength = sum of adjacent edge weights) within the neighborhood, except for the hub itself. The test statistic for the wKDA is analogous to the one used for MSEA: χ=O−EE−κ, except that the values O and E represent the observed and expected ratios of genes from CVD/T2D gene sets in a hub subnetwork. In particular, E=NkNpN is estimated based on the hub degree Nk, disease gene set size Np and the order of the full network N, with the implicit assumption that the weight distribution is isotropic across the network. Statistical significance of the disease-enriched hubs, henceforth KDs, is estimated by permuting the gene labels in the network for 10000 times and estimating the P-value based on the null distribution. To control for multiple testing, stringent Bonferroni adjustment was used to focus on the top robust KDs. KDs shared by CVD and T2D modules are prioritized based on the following criteria: i) Bonferroni-corrected p< 0.05 in wKDA, ii) replicated by both GIANT networks and Bayesian networks, and iii) the genetic association strength between the KD subnetworks (immediate network neighbors of the KDs) and CVD/T2D in Meta-MSEA. Finally, Cytoscape 3.3.0 [61] was used for disease subnetwork visualization.
We used the genetic effect direction of KDs as a proxy for probable effect direction of the KD subnetworks. For each KD, we retrieved their tissue-specific eQTLs as well as variants within 50kb of the gene region, whose genetic association information was available in both CARDIoGRAMplusC4D and DIAGRAM, the two large meta-consortia of GWAS for CVD and T2D. CVD/T2D association beta-values of mapped variants of KDs were then extracted, and the signs of beta-values were examined to ensure they were based on the same reference alleles between GWAS. Lastly, for all mapped variants on each KD, the signs of the beta-value for CVD and T2D were compared and statistical significance of the proportion of variants with similar or opposite effect direction between diseases was determined by z-test.
We chose to validate the predicted adipose subnetwork of a top ranked KD of both CVD and T2D, Cav1, in 3T3-L1 adipocytes. Cells were cultured to confluence and adipocyte differentiation was induced using MDI differentiation medium (S1 Text). Two days after the initiation of differentiation, cells were transfected with 50 nM Cav1 siRNAs (3 distinct siRNAs were tested and two of the strongest ones were chosen) or a scrambled control siRNA. For each siRNA, two separate sets of transfection experiments were conducted, with three biological replicates in each experiment. Two days after transfection, cells were collected for total RNA extraction, reverse transcription and quantitative PCR measurement of 12 predicted Cav1 subnetwork genes and 5 random genes not within the subnetwork as negative controls (S1 Text). β-actin was used to normalize the expression level of target genes.
We accessed the gonadal white tissue gene expression data of 3-month-old wild type and Cav1-/- male mice (N = 3/group) from Gene Expression Omnibus (GEO accession: GSE35431). Detailed description of the data collection procedures has been described previously [33]. Gene expression was profiled using Illumina MouseWG-6 v2.0 expression beadchip and normalized using robust spline. Differentially expressed genes (DEGs) between genotype groups were identified using linear model implemented in the R package Limma and false discovery rate was estimated using the Benjamini-Hochberg procedure [62]. DEGs at different statistical cutoffs were compared to CAV1 subnetwork genes at different levels (i.e., 1, 2, 3, or 4 edges away from CAV1) to assess overlap and significance of overlap was evaluated using Fisher’s exact test.
To further validate the role of KD subnetworks in CVD and T2D, we incorporated genetic, genomic and transcriptomic data from HMDP (comprised of >100 mouse strains differing by genetic composition) [21–23]. HMDP data was from two panels, one with mice fed with a high-fat diet (HF-HMDP)[22], and the other with hyperlipidemic mice made by transgenic expression of human APOE-Leiden and CETP gene (ATH-HMDP)[23]. For HF-HMDP, we retrieved gene-trait correlation data for adipose tissue (due to its relevance to both CVD and T2D) and 7 core cardiometabolic traits including adiposity, fasting glucose level, fasting insulin level, LDL, HDL, triglycerides and homeostatic model assessment-insulin resistance (HOMA-IR). For ATH-HMDP, we retrieved aorta gene-trait correlation (aorta tissue is the main site for CVD in mice) for all 7 traits. In addition to assessing the trait association strengths of individual KDs, we also used MSEA to evaluate the aggregate association strength of the top CVD/T2D KD subnetworks with the traits at both transcription and genetic levels through transcriptome-wide association (TWAS) and GWAS in HF-HMDP and ATH-HMDP (S1 Text).
|
10.1371/journal.ppat.1003788 | Genome-Wide Detection of Fitness Genes in Uropathogenic Escherichia coli during Systemic Infection | Uropathogenic Escherichia coli (UPEC) is a leading etiological agent of bacteremia in humans. Virulence mechanisms of UPEC in the context of urinary tract infections have been subjected to extensive research. However, understanding of the fitness mechanisms used by UPEC during bacteremia and systemic infection is limited. A forward genetic screen was utilized to detect transposon insertion mutants with fitness defects during colonization of mouse spleens. An inoculum comprised of 360,000 transposon mutants in the UPEC strain CFT073, cultured from the blood of a patient with pyelonephritis, was used to inoculate mice intravenously. Transposon insertion sites in the inoculum (input) and bacteria colonizing the spleen (output) were identified using high-throughput sequencing of transposon-chromosome junctions. Using frequencies of representation of each insertion mutant in the input and output samples, 242 candidate fitness genes were identified. Co-infection experiments with each of 11 defined mutants and the wild-type strain demonstrated that 82% (9 of 11) of the tested candidate fitness genes were required for optimal fitness in a mouse model of systemic infection. Genes involved in biosynthesis of poly-N-acetyl glucosamine (pgaABCD), major and minor pilin of a type IV pilus (c2394 and c2395), oligopeptide uptake periplasmic-binding protein (oppA), sensitive to antimicrobial peptides (sapABCDF), putative outer membrane receptor (yddB), zinc metallopeptidase (pqqL), a shikimate pathway gene (c1220) and autotransporter serine proteases (pic and vat) were further characterized. Here, we report the first genome-wide identification of genes that contribute to fitness in UPEC during systemic infection in a mammalian host. These fitness factors may represent targets for developing novel therapeutics against UPEC.
| Uropathogenic E. coli is a major cause of bacterial bloodstream infections in humans. Dissemination of E. coli into the bloodstream during urinary tract infections may lead to potentially fatal complications. This pathogen is becoming increasingly resistant to currently used antibiotics. To develop additional tools to treat such infections, a thorough understanding of the mechanism of pathogenesis is required. Here, we report major progress towards that goal by identifying bacterial genes that are critical for the ability of this pathogen to cause bloodstream infections using a mouse model of infection. This study sheds light on the conditions encountered by E. coli during systemic infection. Further research on the genes identified in this study may reveal bacterial targets that can be used to develop novel therapeutics against bloodstream infections caused by E. coli.
| Uropathogenic Escherichia coli (UPEC), one of the most common bacterial pathogens infecting humans, is the primary etiological agent of urinary tract infections (UTI) in otherwise healthy individuals [1]. UPEC is a subset of extraintestinal pathogenic E. coli (ExPEC), which causes a broad spectrum of conditions including colibacillosis in poultry, and UTIs, bacteremia, and neonatal meningitis in humans [2]. A subset of patients with UTI develops pyelonephritis and is at risk for developing bacteremia that may result in life threatening sepsis. UTI is the source of E. coli in >70% of both young and elderly patients with bloodstream infections [3], [4]. E. coli strains isolated from the bloodstream are becoming increasingly resistant to trimethoprim/sulfamethoxazole and ciprofloxacin, two first line antibiotics used to treat bacterial UTIs [5]. Despite the prevalence of these infections and potential difficulties in treatment, little is known about the fitness and virulence mechanisms employed by E. coli to establish a systemic infection.
The marriage between transposon mutagenesis and high-throughput (HT) sequencing has resulted in the emergence of powerful techniques that can be harnessed for global functional genomic studies [6]. Here, we utilize an adaptation of transposon directed insertion-site sequencing (TraDIS) [7] to identify genes required for optimal fitness of UPEC during colonization and survival in a murine model of bacteremia. Recently, such approaches were used to determine virulence and fitness factors in Yersinia pseudotuberculosis [8] and Salmonella enterica serovar Typhimurium [9] utilizing animal models of infection and colonization.
Genes that encode microbial proteins and organelles that specifically aid in pathogenesis are known as virulence genes. Bacterial pathogens are adept at co-opting genes that are otherwise used in non-pathogenesis related roles for gaining fitness advantage during infection. In this context, fitness refers to enhanced survival and growth within a given niche. Genes that promote colonization and survival of UPEC within murine hosts are referred to as fitness factors in this manuscript. A subset of the fitness factors reported here, represent virulence factors that meet the criteria defined by molecular Koch's postulates [10]. This report, to our knowledge, represents the first global functional genomic screen aimed at identification of in vivo fitness factors in a pathogenic E. coli strain involving a targeted-sequencing approach.
In this study, a murine model of invasive UPEC infection, previously developed in our laboratory [11], was used in conjunction with transposon mutagenesis to identify bacterial fitness mechanisms involved in establishing systemic infection. Mice were inoculated intravenously with an inoculum derived from a saturating transposon mutant library of a clinical bacteremia isolate, E. coli CFT073. Transposon mutants that colonized and survived in mouse spleens (output) were isolated. Transposon insertion sites in the input and output samples were mapped to the genome of the UPEC strain CFT073. 242 candidate fitness genes that are required for optimal survival in the spleen were identified in the primary screen. Genetically defined mutants were constructed and tested for in vivo and in vitro fitness phenotypes using assays relevant to the infection biology of UPEC. A subset of these fitness factors are also involved in the development of UTI in a mouse model and suggests the existence of shared fitness mechanisms used at these disparate body sites. In summary, we present a comprehensive study of fitness factors that augment the survival of UPEC during systemic disseminated infection in a mammalian host.
E. coli CFT073, isolated from the urine and blood of a patient hospitalized with pyelonephritis and bacteremia [12], was used to construct a high-density transposon mutant library. An estimated 48,174 transposon mutants are required to obtain a 99.99% saturation of the CFT073 genome [13], which is 5.2 Mbp in length [14]. A genome-supersaturating Tn5 transposon mutant library, containing 360,000 kanamycin-resistant transformants, was generated for this study. The library was passaged three times in lysogeny broth (LB) to enrich for mutants that did not exhibit a fitness defect in vitro. This enriched mutant pool was used as the inoculum for infection experiments.
A murine model of systemic disseminated UPEC infection [11] was used to determine the highest dose of wild-type CFT073 that consistently resulted in non-lethal infection. Three doses (106, 107 and 108 CFU/mouse) were compared in the CBA/J mouse model of systemic disseminated infection. Mice were inoculated via tail vein, and livers and spleens were collected 24 h post inoculation (hpi). An inoculum of 107 CFU resulted in consistent colonization without causing distress in the inoculated animals (Fig. 1A). Inoculation with 106 CFU led to poor colonization, whereas a dose of 108 CFU resulted in 20% mortality (Fig. 1A). Twenty mice were inoculated with 107 CFU of transposon-insertion mutants (input) and euthanized 24 hpi (Fig. 1B). As a major reticuloendothelial organ, the spleen is a critical site of active bacterial killing during systemic infection [15]. Therefore, a bacterium that successfully survives in the spleen should contain the full complement of fitness factors that are critical for survival in that niche, including the ability to overcome host defenses activated during systemic bacterial infection. Bacteria that grew from splenic homogenates were harvested (output) and used to isolate genomic DNA for Illumina sequencing.
Transposon insertion sites in the inoculum (input) and bacteria colonizing the spleens (output) were determined using transposon directed insertion-site sequencing (TraDIS), a HT sequencing-based approach [7]. Genomic DNA from the input and output samples were used to generate TruSeq sequencing libraries (Illumina). Libraries were amplified using a transposon-specific forward primer and a custom adapter-specific reverse primer (Table S1). Resulting amplicons were used for cluster generation and each library was sequenced with a Tn-specific primer (Table S1) in an IIlumina HiSeq 2000 sequencer. Fifty nucleotide single-end reads in FASTQ format were aligned to the E. coli CFT073 genome [14] using the short read aligner, BOWTIE [16]. The number of reads processed was 75,935,499 and 87,030,926 for the input and output samples, respectively. 76.7% (58,209,557) of the reads from input and 77.9% (67,810,765) of the reads from output were aligned unambiguously to the CFT073 genome. TFAST [17] was used to determine the exact genomic location of the Tn-insertion and the frequency of reads that map to a given insertion site.
Reads for each transposon insertion site were normalized to the total number of reads obtained from that sample and a fitness factor was calculated for each Tn-insertion mutant as the ratio of normalized frequency of reads in the input to that of the output (Fig. 2A, Table 1 and Table S2). Therefore, a fitness factor ≥1 indicates that a given mutant is underrepresented in the output pool. For example, the sensitive to antimicrobial peptide (sap) gene cluster in E. coli CFT073 is depicted along with the frequency of representation of transposon-insertion mutants in the input and the output pools (Fig. 2A). Transformants containing an insertion in the sap genes are less well represented in the output pool compared to the input pool. A total of 6732 unique Tn5 insertion sites were mapped in the CFT073 genome (Table S2) with a mean fitness factor of 3.27±1.57. The insertion sites were distributed throughout the length of the CFT073 genome. At least a single transposon insertion site was observed in 3020 genes and an additional 843 intergenic regions, which could exert polar effects of downstream genes. A relatively short region in the genome (30 Kbp in length) reveals the presence of transposon insertion mutants with a broad range of fitness factors (Fig. 2B). The median distance between independent insertion sites was 561 bp. A total of 372 transposon insertion mutants, resulting in inactivation of 242 genes, exhibited fitness factors >6.41 (mean+2 standard deviations) and were considered as candidate fitness genes. 50 transposon mutants with the highest in vivo fitness defect phenotype are listed in Table 1. Seven (1.9%) of 372 candidate transposon mutants were previously designated as essential genes in a laboratory strain of E. coli K-12 [18]. These genes encode NrdB, an aerobic ribonucleotide reductase; MviN, a peptidoglycan lipid II flippase; LolD, lipoprotein releasing system ATP-binding protein; MinD, septum site-determining protein; GapA, glyceraldehyde-3-phosphate dehydrogenase; FabG, 3-ketoacyl-acyl carrier protein reductase; and MsbA, lipid flippase.
Genes involved in nucleotide metabolism were previously described to play a critical role in growth of a non-pathogenic strain of E. coli in human blood [19]. The following genes were identified in our primary screen: guaB, inosine monophosphate dehydrogenase involved in guanosine monophosphate biosynthesis; ntpA, dATP pyrophosphohydrolase involved in degradation of dATP; pyrC, dihydroorotase catalyzes the conversion of carbamoylaspartate to dihydrooratate; and yeiA, dihydropyrimidine dehydrogenase catalyzes first step in degradation of uracil and thymine (Table S2).
Several genes encoding surface structures that could potentially be involved in direct interaction with host cells were identified in our screen (Table S2). Periplasmic murein-peptide binding protein precursor gene (mppA) and a periplasmic protease that processes penicillin-binding protein 3 (prc, c2239) [20] are peptidoglycan biosynthetic genes that were identified in our primary screen. arnT and yfbH genes involved in resistance to polymyxin B, a peptide antibiotic that mimics the activity of host-derived cationic antimicrobial peptides, were identified in our primary screen. ArnT reduces the negative charge on the lipopolysaccharide due to its 4-amino-4-deoxy-L-arabinose transferase activity [21]. YfbH is a homolog of PmrJ, a deacetylase involved in biosynthesis of amino-arabinose-modified lipid A [22]. Two outer membrane porins, ompC and ompG were also identified in our screen. Colanic acid is a surface polysaccharide that is associated with biofilm formation in E. coli [23]. Two genes involved in colonic acid biosynthesis, wcaM and wcaL, [24] were identified as candidate fitness genes in the mouse bacteremia model. pgaABD, c2394-95 and yddB are other genes associated with surface structures identified in our screen and were subjected to further investigation.
Mammalian hosts actively limit the bioavailability of iron to hamper the growth of invading pathogens. Multiple genes involved in distinct iron acquisition systems were identified in our screen (Table S2). The sitC gene harbored on a bacteriophage, part of the SitABCD system involved in manganese and iron transport [25], was among the candidate fitness genes identified in this study (Table S2). A chuA hma double mutant, lacking two heme receptors, was previously found to exhibit fitness defect during bacteremia [11]. In the current study, mutation in hma alone reveals a fitness defect (Table S2) suggesting that heme is a major source for iron during systemic infection in a mammalian host. Multiple insertion sites were found within the genes involved in enterobactin biosynthesis, export, uptake and utilization with a mean fitness factor of 3.2. Salmochelin is a glycosylated derivative of enterobactin that evades chelation by host protein lipocalin-2 [26]. Inactivation of genes involved in salmochelin biosynthesis (iroB) and uptake (iroN) also resulted in attenuated fitness (Table S2). Insertion mutants in yersiniabactin biosynthesis and uptake genes also revealed a minor fitness defect (Table S2). Yersiniabactin, however, is not produced by E. coli CFT073 due to a previous insertion event at this locus. Coprogen and hydroxamate siderophore receptor gene, fhuE, was found among the candidate fitness genes (Table S2). Our results are consistent with the previously established role of iron acquisition genes in fitness of UPEC during systemic infection.
Tn-insertions leading to fitness defect in multiple genes within an operon/cluster and genes that were previously not known to affect fitness of UPEC during systemic infection were selected for further validation. Additionally, we tested the role of pic and vat in fitness primarily to establish that we have utilized a conservative threshold to delineate fitness genes. Co-infection experiments were performed by inoculating mice intravenously with equal numbers of both wild-type and mutant bacteria lacking select genes identified in the primary screen. Since most cases of bacteremia caused by UPEC are a result of ascending UTIs, we also tested the ability of a subset of these mutants to colonize murine urinary tract.
Growth kinetics of all the mutants used in the experiments described in the following sections is indistinguishable from that of the wild-type strain (Fig. S1). Homogenates of organs (spleen and liver from bacteremia model; urinary bladder and kidneys from ascending UTI model) were plated on plain and selective media. Differential plate counts were used to determine bacterial loads of wild-type and mutant strain in each tissue. Competitive indices (CI) were calculated using colony counts as: [mutant CFU/wild-type CFU (output)]/[mutant CFU/wild-type CFU (input)]. CI values less than 0 (log10 scale) indicate a comparative fitness defect for the mutant with respect to wild-type strain (Fig. 3A and 3B, 4A and 5A). Nine of the 11 mutants (82%) were out-competed by the wild-type strain during co-infection indicating that a high proportion of the candidate fitness genes identified in the primary screen indeed function as fitness factors during systemic infection (Fig. 3A and 3B, 4A and 5A). After successful validation of the primary screen, we probed the function of select candidate fitness genes in UPEC.
Transposon insertions within pgaA, pgaB, and pgaD resulted in reduced fitness, corresponding to fitness factors of 7.43, 9.41, and 4.32, respectively (Tables 1 and S2). The pgaABCD operon is involved in the biosynthesis and export of an extracellular polysaccharide, poly-N-acetyl glucosamine (PNAG), in E. coli [27]. Loss of PNAG biosynthetic operon resulted in a fitness defect in a mouse model of bacteremia (spleen, P = 0.002; Fig. 4A). The pgaABCD mutant was out-competed by the wild-type strain, ∼10-fold, both within the urinary bladder and the kidneys demonstrating that PNAG acts as a fitness factor in vivo within the murine urinary tract (Fig. 4A). The pgaA gene was upregulated 32-fold and pgaC transcript was detected by RT-PCR in urine collected from mice infected with E. coli CFT073, indicating that these genes are highly expressed during UTI. Furthermore, transcriptome analysis of UPEC CFT073 revealed that the pgaABCD genes are upregulated (∼2-fold) during culture in human urine compared to LB (unpublished results).
Since PNAG is involved in biofilm formation in a non-pathogenic strain of E. coli [27], we tested the contribution of PNAG to biofilm formation in the UPEC strain CFT073. Biofilm-forming ability of wild-type and pgaABCD mutant was tested using a crystal violet binding assay. Loss of PNAG did not affect biofilm formation on polystyrene (Fig. 4B) or glass surface (data not shown). Since UPEC is decorated with several surface structures, including multiple fimbriae and autotransporter adhesins, which might compensate for the loss of PNAG-dependent adhesion, the effect of overexpression of the pgaABCD operon on biofilm formation was also tested. Full-length pgaABCD operon including the native promoter was cloned into a multi-copy vector (pSS1); PNAG could be readily detected, by immunoblot analysis, in the overexpression strain but not in the vector control (Fig. 4C inset). Upon overexpression, PNAG promotes robust biofilm formation in UPEC strain CFT073 (Fig. 4B). E. coli K-12 strain MG1655 also displayed a profound, PNAG-dependent increase in biofilm formation, indicating that PNAG promotes biofilm formation in E. coli using a non strain-specific mechanism (Fig. 4B).
Factors involved in adherence are known to affect motility in UPEC [28]. In strain CFT073, loss of PNAG production results in a significant increase in motility (Fig. 4C) that is accompanied by a 4-fold increase in the expression of fliC (data not shown). A higher level of fliC expression (encoding flagellin, the major structural subunit of flagella) explains the increased motility observed in the pgaABCD mutant. Conversely, overexpression of PNAG diminishes motility (Fig. 4C) suggesting that PNAG production and motility could be controlled in a reciprocal manner. Overexpression of PNAG also resulted in decreased motility in E. coli K-12 (Fig. 4C). Taken together, motility is adversely affected during PNAG overexpression in a non strain-specific manner. Additionally, known repressors of flagellar motility, PapX and FocX, are not involved in this crosstalk between PNAG levels and swimming motility (data not shown).
Intact epithelial surface precludes the access of pathogens to ECM proteins; however, inflammation-associated mucosal denudation results in contact with ECM proteins. A plate-based adherence assay was used to determine whether PNAG is involved in adherence to common ECM proteins collagen I, collagen IV, fibronectin and laminin. PNAG overexpression resulted in significantly higher adherence of UPEC strain CFT073 to collagen I, collagen IV and laminin (Fig. 4D) compared to vector control. PNAG does not affect binding to fibronectin under the assay conditions tested. To determine if PNAG protects UPEC from killing by macrophages, survival of wild-type, pgaABCD, wild-type (pSS1), and wild-type (pTopo) within the murine macrophage cell line RAW264.7 was assessed. Under our experimental conditions, PNAG did not contribute to adherence or intracellular survival (data not shown).
c2394 encoding PilV was identified in the primary screen as a putative fitness gene (fitness factor = 6.7, Table S2). pilV (c2394) and pilS (c2395), encoding pilin subunits of type IV pilus two, are highly associated with UPEC strains compared to fecal E. coli isolates [29]. Additionally, these genes are more prevalent in E. coli isolated from humans than from animals [29]. Co-infection experiments revealed that the mutant strain lacking pilV and pilS genes was significantly out-competed (P<0.05) by wild-type E. coli CFT073 in the bladder, kidneys, spleen and liver (Fig. 5A). Our data demonstrate that these putative type IV pilin subunit genes are involved in colonization during both systemic infection and UTIs.
The ability of the isogenic mutant, c2394-95 (pGEN), and the complemented strain, c2394-95 (pGEN- c2394-95) to adhere to the immortalized epithelial cell lines UMUC-3 (human bladder), HEK293 (human embryonic kidney), VERO (green monkey kidney), and MM55K (mouse kidney) was compared to that of wild-type (pGEN) strain. Compared to wild-type, c2394-95 mutant was less adherent to UM-UC-3 (P = 0.032), HEK293 (P = 0.031), and VERO (P = 0.012) (Fig. 5B) cells. However, no significant difference was observed on MM55K (P = 0.675) cells. Complementation restored adherence to wild-type levels on all cell lines (Fig. 5B). This suggests that c2394-95 encode proteins involved in adherence to uroepithelial cells and the receptor for type IV pilin is likely expressed by both bladder (human) and kidney (human and monkey) epithelial cells, but not by the mouse kidney cell line. Electron microscopy was used to determine if the type IV pilus is indeed found on the cell surface. Wild-type and complemented mutant cells are densely piliated compared to the mutant strain that is sparsely piliated (Fig. 5C).
A c2394-95 mutant had a swimming diameter of 44.9±7.7 mm, significantly lower than that of wild-type E. coli CFT073 (P = 0.005), which swam 59.8±4.3 mm. Motility was not restored to wild-type levels by complementation (38.4±6.0 mm) with c2394-95 in trans; instead, the motility defect was increased upon expression of the pilin genes, which suggests that there may be a decrease in motility due to the level of expression from a multi-copy plasmid. Deletion of c2394-95 does not appear to affect cell aggregation, biofilm formation or invasion of kidney epithelial cells in E. coli CFT073 (data not shown).
Multiple peptide uptake genes (oppABD, sapACF and tppB) were identified as candidate fitness genes in the primary screen (Tables 1 and S2). Mutants lacking the sap gene cluster or oppA, but not the tppB were found to exhibit fitness defects in spleen and liver during co-challenge experiments, compared to the wild-type strain (Fig. 3). The opp gene cluster harbors the genes involved in oligopeptide uptake and multiple transposon insertion sites were observed within these genes (Table S2). This observation suggests that the ability to utilize oligopeptides as a source of carbon and nitrogen is critical for UPEC survival in murine spleens. Transposon insertions and corresponding fitness factors for the sap genes are depicted in Fig. 2A.
Cationic antimicrobial peptides represent a major antimicrobial defense system that aids in clearing invading pathogens. Polymyxin B (PB) is a peptide antibiotic that emulates the activity of host-derived cationic antimicrobial peptides [30]. The role of peptide uptake systems in resistance of UPEC to PB was tested. A mutant defective in dipeptide uptake (dppA) [31], not identified in our primary screen, was also used to determine if multiple peptide uptake systems are involved in PB resistance. Bacterial cultures in exponential phase of growth were exposed to PB and percent survival was calculated using colony counts from PB-treated and control cultures. Fold-change in resistance was calculated as the ratio of survival percentage of a given mutant to that of wild-type strain (Fig. 6A). Compared to the wild-type strain, the sapR mutant, that lacks the sapABCDF genes, exhibited increased sensitivity to PB (Fig. 6A, P<0.0001). The oppA mutant exhibited decreased sensitivity to PB compared to wild-type strain (Fig. 6A, P = 0.03) and the dppA mutant also showed a trend towards decreased sensitivity to PB (Fig. 6A).
Gentamicin protection assay was used to determine if the peptide uptake systems contributed to intracellular survival of UPEC in murine macrophage cells (RAW 264.7). Plate counts were used to determine the number of bacteria that entered and survived within RAW264.7 cells for 2 h. Ratio of killing percentages were determined and values >1 indicate that a given mutant was defective in intracellular survival within RAW 264.7 cells (Fig. 6B). A modest, but statistically significant reduction (P = 0.03) in intracellular fitness of the sapR mutant was observed (Fig. 6B), whereas the oppA and dppA mutants were not defective in intracellular survival compared to wild-type strain.
Although tppB was identified as a fitness gene in the primary screen, a co-challenge experiment revealed no role for this gene in fitness (Fig. 3). This discrepancy could be due to the differences in the nature of competition during infection with the transposon mutant library in the primary screen versus one-to-one competition between wild-type and mutant strains in our secondary validation experiments.
In the E. coli CFT073 genome, yddA, yddB and pqqL encode an ABC transporter ATPase, an outer membrane β-barrel protein and an inner membrane-associated zinc metallopeptidase, respectively. yddB and pqqL were identified as fitness genes in our primary screen and median fitness factors for multiple insertion mutants in these genes are depicted in Fig. 7A. A BLAST search revealed that this gene cluster is found only among E. coli and Shigella strains. yddB and pqqL genes are involved in fitness during systemic infection in a mammalian host (Fig. 3). RT-PCR experiments revealed that these genes are indeed co-transcribed as a single mRNA (Fig. 7B).
YddB exhibits a high degree of sequence similarity to ligand-gated outer membrane β-barrel proteins such as ferrienterobactin receptor, FepA in E. coli. Since outer membrane β-barrel proteins are usually involved in iron uptake and a putative Fur box (GGGAATGGTTATCATTAG) is found overlapping the start codon of yddA, we tested whether these genes are differentially expressed during culture in human urine, an iron limited milieu. RNA was extracted from CFT073 bacterial cells cultured to mid-exponential phase in either LB or filter-sterilized human urine and gene expression was quantified using RNAseq (unpublished results). The yddA, yddB, and pqqL genes are highly upregulated (>30-fold) during growth in human urine compared to LB (Fig. 7C). We also tested whether iron levels directly regulate the expression of yddA-yddB-pqqL genes. Transcript levels were determined in wild-type strain cultured in LB, LB supplemented with an iron chelator (Dipyridyl) or additional iron. Iron levels alone do not affect the expression of these genes in UPEC (Fig. S2A). However, yddA, yddB, and pqqL genes are upregulated in the Δfur mutant that lacks ferric uptake regulator (Fur), compared to the wild-type strain (Fig. S2B). Taken together, our data indicate that these genes are upregulated during growth in human urine but not subject to regulation by iron levels alone.
Proteins in the SPATE (serine protease autotransporter proteins of Enterobacteriaceae) family have previously been implicated in the pathobiology of UPEC [32]. Genes encoding members of SPATE family, protease involved in colonization (Pic) and vacuolating autotransporter toxin (Vat, previously known as Tsh) were identified in our primary screen (Table S2). The pic and vat transposon insertion mutants exhibited a fitness factor of 5.2 and 4.4, respectively. Co-infection experiments were performed with these genes to test whether a conservative threshold was used to delineate fitness genes. Competitive indices reveal that both pic and vat, which exhibit fitness factors lower than the cutoff used to delineate fitness genes, play a role in the fitness of UPEC during systemic infection in mice (Fig. 3).
3-deoxy-D-arabino-heptulosonic acid-7-phosphate synthase (DAHPS), encoded by c1220, catalyzes the formation of 3-deoxy-D-arabino-heptulosonate 7-phosphate (DAHP) from phosphoenolpyruvate and erythrose 4-phosphate, an early step in shikimate biosynthesis [33]. In CFT073, c1220 is located on the serX pathogenicity island [34]. DAHPS encoded by c1220 is the fourth isozyme, in addition to aroF, aroG and aroH that catalyzes the production of DAHP in E. coli CFT073. Two transposon insertion sites mapped to this gene resulted in reduced fitness during survival in spleen. Co-infection experiments with a c1220 mutant and wild-type strain confirmed that the mutant has a fitness defect in spleen and liver during systemic infection in mice (Fig. 3).
A gene encoding an EAL domain protein, ycgF, was identified in our primary screen. EAL domain proteins are usually associated with phosphodiesterase activity that reduces the intracellular levels of an important intracellular messenger, cyclic-di-GMP [35]. YcgF has been designated as an inactive phosphodiesterase that nevertheless positively regulates swimming motility by increasing flagellin levels in CFT073 [35]. Although ycgF was identified as a candidate fitness gene, co-infection experiments failed to reveal a role for ycgF in fitness in a mouse model of systemic infection (Fig. 3).
Uropathogenic Escherichia coli (UPEC) is a major cause of bacteremia in humans, yet, there is limited understanding of the fitness mechanisms used by this important pathogen during bacteremia and systemic infection. Here, we describe screening transposon mutants of E. coli CFT073 in a murine model of systemic disseminated infection and identifying 242 candidate fitness genes. Specific mutations were introduced in 11 candidate fitness genes and the contribution of the following nine gene or gene clusters in fitness was confirmed: pgaABCD (biosynthesis and export of poly-N-acetyl glucosamine), c2394-95 (major and minor pilin of type IV pilus two), oppA (oligopeptide uptake periplasmic-binding protein), sapABCDF (sensitive to antimicrobial peptide), yddB (putative outer membrane receptor), pqqL (zinc metallopeptidase), c1220 (a shikimate pathway gene), and pic and vat (autotransporter serine proteases). 82% of the specific mutants in representative candidate fitness genes were significantly outcompeted by the wild-type strain, validating the TraDIS approach in our murine model of systemic infection.
Transposon mutagenesis has been an indispensable tool in unraveling gene function. The complex nature of experiments involving transposon mutant pools, including bottlenecks when screening signature-tagged mutants in animal models of infection [36], has resulted in screens with fewer mutants than required to achieve genome saturation. Recently, HT sequencing and bioinformatic analyses have been used in tandem to identify transposon insertion sites in genome-saturating transposon mutant pools [6]. Several variants of this approach include HITS, high-throughput insertion tracking by deep sequencing [37]; INSeq, insertion sequencing [38]; Tn-Seq, transposon sequencing [39]; and TraDIS, transposon directed insertion-site sequencing [7]. These techniques utilize chromosomal regions flanking the transposons as unique tags in lieu of the synthetic tags used in signature-tagged mutagenesis. Availability of a large number of bacterial genome sequences combined with cost-effective HT sequencing is poised to make these approaches a staple of functional genomic studies in the near future.
Understanding the fitness strategies employed by UPEC during infection of a mammalian host has the potential to identify targets for novel intervention strategies. Here, we describe the first comprehensive identification of fitness factors involved in systemic infection by an ExPEC strain, CFT073, in a mammalian host. The original work describing TraDIS catalogued the essential genes in Salmonella enterica subsp. enterica serovar Typhi [7]. Since the primary objective of our study was to identify in vivo fitness factors, the mutant pool was passaged in LB to deplete mutants with in vitro fitness defects from the inoculum. This might have led to a reduction in the diversity of the mutant pool used for infection, compared to the original pool comprising of 360,000 transformants and may explain the fact that we identified only 6732 independent transposon insertion sites. Notwithstanding the reduced complexity of the input pool, we have identified novel fitness factors from this study.
TFAST was previously developed in our laboratory to determine the transcription factor binding sites [17] and facilitated successful identification of PapX binding site in the flhDC promoter [40]. Here, TFAST was applied to determine the chromosomal location and the frequency of detection of a given transposon mutant. Potential fitness genes identified in this study could be an underestimate because genes pic and vat did not meet the threshold (mean+2 standard deviations) but were confirmed as fitness genes in the co-infection experiments. The EZ-Tn5 transposon used for random mutagenesis was not modified to incorporate promoter regions at either end; therefore, the transposon insertion mutations could exert polar effects on co-transcribed genes. Additionally, random events could result in the loss of a transposon mutant during infection and could result in misinterpretation as a fitness gene. In co-infection experiments, nine of the 11 (82%) tested mutants revealed a fitness defect confirming the validity of our primary screen.
Seven of 372 predicted transposon mutants with a fitness defect (1.9%) were found in another study as essential genes in a laboratory strain of E. coli [18]. Studies on gene essentiality in E. coli have been conducted primarily on non-pathogenic, laboratory-adapted strains. Genomes of UPEC are usually larger than these laboratory strains. For instance, genome of UPEC CFT073 is ∼590 Kbp longer than the widely studied E. coli K-12 strain MG1655 [14]. Depending on the transposon insertion site and growth conditions, it is possible that transposon insertions could be tolerated in some essential genes. For instance, degS is designated as an essential gene in E. coli [18]. However, a degS mutant has been successfully constructed in E. coli CFT073 and used to demonstrate that DegS, likely by modulating members of Sigma E regulon, affects the fitness of UPEC during peritonitis as well as during UTI in a mouse model of infection [41]. Another possible explanation is the emergence of suppressor mutations that negate the effects of original mutation. It is also possible that these are artifacts due to sequence similarity to parts of other non-essential genes or gene duplication events. Potentially, some of the essential genes involved in non-structural components could be complemented by other transformants within the mutant pool. These genes constitute only a small fraction of all the fitness genes unraveled in this study.
A previous study in ExPEC during systemic infection in a mammalian host, led to the identification of type 1 pilus; P fimbria; Hma and ChuA, heme receptors; TonB, iron uptake energy transducer; Ksl, K2 capsule biogenesis; and NanA, N-acetylneuraminate aldolase as fitness determinants [11]. This study has greatly expanded the potential bacteremia fitness determinants in UPEC and offers evidence for the role of nine of these novel fitness determinants in a murine model of systemic infection. Furthermore, 81 (33.5%) of the candidate fitness genes are predicted to encode hypothetical proteins and constitute a unique resource that can be exploited to identify previously unknown fitness determinants.
Biosynthetic mutants defective in either salmochelin or both salmochelin and enterobactin production revealed reduced fitness in the primary screen. Since these mutants retain the ability to utilize catecholate siderophores synthesized by other transformants, it is likely that the observed fitness defect emerges from iron uptake-independent roles. Recently, catecholate siderophore biosynthesis, but not uptake-alone, was demonstrated to mitigate the effects of oxidative stress in both Salmonella Typhimurium and E. coli [42]. It is, therefore, plausible that UPEC utilizes catecholate siderophore biosynthesis not only for canonical iron acquisition functions but also for protection against oxidative stress encountered during systemic infection.
PNAG, an extracellular polysaccharide, has been associated with the virulence in a broad spectrum of bacterial pathogens, including Aggregatibacter actinomycetemcomitans [43], Bordetella pertussis [44], Staphylococcus aureus [45], and Yersinia pestis [46]. Antibodies raised against S. aureus-derived PNAG confer passive protection against systemic infection with a clinical UPEC strain [47]. Here, we provide evidence that biosynthesis of PNAG is required for optimal fitness of UPEC during both UTI and systemic infection (Fig. 4).
Type IV pili are filamentous organelles found at the bacterial surface that affect adherence and motility in several bacterial species, including enteropathogenic E. coli [48]. We found that the genes encoding predicted major and minor type IV pilins (c2394-95) are critical for fitness during both bacteremia and UTIs. Although the mutant did not exhibit reduced adherence to MM55K cells, an immortalized cell line derived from the kidneys of AKR strain mice, the mutant revealed colonization defect in murine kidneys in a mouse model of ascending UTI. Differences in the expression of surface receptors on MM55K cells compared to those found within the nephrons of live CBA/J mice used for infection experiments could account for the discrepancy between adherence phenotypes observed for the type IV pilus mutant during in vitro and in vivo assays. Electron micrographs revealed reduced number of pili in the mutant compared to wild-type and complemented mutant strain. However, UPEC strain CFT073 produces multiple fimbria [29]; therefore, this observation must be verified with immunostaining to enable specificity. UPEC also harbors a locus similar to that encoding type IV pilus in E. coli K-12 and has been demonstrated to affect the fitness in mouse urinary tract [49]. Mutants in both type IV pilus loci exhibit fitness defects independent of each other and here we demonstrate that type IV pilus two is a novel fitness factor in UPEC.
Oligopeptide uptake system gene oppA was previously shown to be critical for fitness of UPEC in the urinary tract [31]. Gene clusters involved in peptide uptake, opp and sap, were found to contribute to the fitness of UPEC during systemic invasion in the current study. The sapABCDF gene cluster contains homologs of genes involved in sensitivity to antimicrobial peptides in Salmonella enterica subspecies Typhimurium [50] and non-typeable Haemophilus influenzae [51]. Our data support a model in which the sap gene cluster, but neither oppA nor dppA, is required for optimal protection against polymyxin B and intracellular survival in murine macrophages (Fig. 6).
Targets of Fur in E. coli MC4100 were detected using a macroarray and yddABpqqL was determined as a Fur-regulated gene cluster in E. coli [52]. A transposon mutant screen in E. coli strain CC118, revealed that yddA and yddB are required for optimal growth in rich medium at 37°C [53]. However, UPEC CFT073 mutants lacking yddB or pqqL genes exhibited no difference in growth rate compared to wild-type strain in vitro (Fig. S1). yddA acts as a colonization factor in enterohemorrhagic E. coli O26:H− in a calf model of intestinal colonization [54]. Enhanced expression in urine (Fig. 7C) and the high degree of identity of YddB protein to ligand-gated outer membrane siderophore receptors allowed us to speculate that these genes could be involved in iron uptake. Although these genes are regulated by Fur (Fig. S2B), they do not appear to be regulated by iron levels alone in CFT073 (Fig. S2A). Cues, other than reduced iron levels, sensed by UPEC in human urine likely govern the regulation of yddABpqqL genes. pqqL from E. coli has been previously shown to complement pyrolloquinoline quinone (PQQ) biosynthetic genes pqqE and pqqF in Methylobacterium organophilum [55]. PQQ is a cofactor for quinoproteins, including glucose dehydrogenase in E. coli. It must be noted that E. coli is not capable of PQQ biosynthesis [55]. Studies are in progress to address whether this gene cluster is involved in uptake and processing of PQQ.
We have identified Pic and Vat, autotransporter serine proteases, to be involved in fitness during bacteremia (Fig. 3). In E. coli CFT073, pic was upregulated during UTI in a murine host and Pic exhibited serine protease activity in vitro [32]. On the contrary, Vat (referred to as Tsh in [32]) did not exhibit a detectable serine protease activity and both of these genes did not appear to affect fitness of UPEC during UTI. Key human leukocyte adhesion molecules such as CD43, CD44, CD45 and CD93, are targeted by Pic resulting in deregulation of leukocyte migration and inflammation [56]. Therefore, it is possible that reduced fitness of Pic mutants during systemic infection could emerge from its role in modulating inflammatory response to systemic infection with UPEC.
Shikimate is a critical intermediary molecule in chorismate biosynthetic pathway and chorismate is a precursor for the biosynthesis of aromatic amino acids, catecholate siderophores, folate, menaquinone and ubiquinone in bacteria [33]. Biosynthesis of aromatic amino acids has been associated with virulence in several bacterial species in various animal models of infection, including Neisseria meningitidis [57], Proteus mirabilis [58], Salmonella enterica subspecies enterica serovar Typhimurium [36], and Staphylococcus aureus [59]. Taken together, these findings indicate that aromatic amino acids and other compounds derived from the chorismate pathway are critical for optimal fitness of multiple bacterial pathogens during infection.
In summary, a combination of transposon mutagenesis and HT sequencing was used to determine genes in UPEC that contribute to fitness in a mouse model of systemic infection. The role of multiple candidate fitness genes was confirmed by independent experiments using a mouse model of infection and in vitro assays. Further characterization of the fitness genes unraveled in this study has the potential to identify targets for developing novel intervention strategies against bacteremia caused by UPEC.
All animal experiments were performed in accordance to the protocol (08999-3) approved by the University Committee on Use and Care of Animals at the University of Michigan. This protocol is in complete compliance with the guidelines for humane use and care of laboratory animals mandated by the National Institutes of Health.
E. coli CFT073, a prototypical uropathogenic strain that caused bacteremia of urinary tract origin [12] was used to generate a saturating Tn5 insertion mutant library. Strains and plasmids used in this study are listed in Table 2. Bacterial strains were cultured in LB containing 0.05% NaCl, unless otherwise noted. Tn5 transformants were cultured in LB containing kanamycin (12.5 µg/ml). Lambda red recombineering was used to introduce specific mutations in strain CFT073 [60]. Genetically defined mutants used in this study were cultured in LB containing either kanamycin (25 µg/ml) or chloramphenicol (20 µg/ml). Oligonucleotide primers used in this study are listed in Table S1. Growth kinetics of the wild-type and mutant strains were determined using a Bioscreen C system (Growth Curves USA). Overnight cultures were diluted 1∶100 in LB and incubated at 37°C. Optical density values were recorded at 600 nm, every 15 min, for 22 h and included three biological replicates, comprised of two technical replicates.
Tn5 insertion mutants were generated in E. coli CFT073 using the EZ-Tn5 transposome kit (Epicentre). Briefly, transposome complexes were electroporated into E. coli CFT073 and bacteria were allowed to recover in SOC broth for 50 min prior to plating on LB agar containing kanamycin using an automated plater (Spiral Biotech). Plates were incubated overnight at 37°C and CFUs were enumerated using a Qcount colony counter (Spiral Biotech). A total of 360,000 transformants were generated for this study and archived in pools of 1800 CFUs. The entire Tn5 mutant collection was passaged three times in LB for 16 h at 37°C and the resulting pool was used as the inoculum for experimental infections.
CBA/J mice (6–7 week old, Harlan Laboratories) were inoculated with 106 (n = 5), 107 (n = 10) or 108 (n = 5) CFU of CFT073 bacteria via tail vein. Mice were euthanized after 24 h and livers and spleens were harvested. Homogenates of these organs were plated on LB plates containing kanamycin and bacterial burden was determined. Mice were inoculated with 107 (n = 20) CFU of CFT073::Tn5 mutants. After 24 h, mice were euthanized to collect spleens. Homogenates of spleens were plated in their entirety, as described above and the bacterial burden was calculated. Colonies from splenic cultures were harvested and pooled from all 20 mice before archiving bacterial pellets at −80°C.
For co-infection experiments, wild-type and specific mutants, cultured overnight, were resuspended in PBS to yield 2×108 CFU/ml, containing equal number of wild-type and mutant CFUs. Inoculum (100 µl) was administered via the tail vein and mice were euthanized 24 h pi. For the ascending UTI model, female mice were inoculated intravesically via a transurethral catheter with 50 µl of the inoculum containing 109 CFU/ml (equal number of wild-type and mutant CFUs) and animals were euthanized after 48 hpi. Homogenates of spleen and liver (intravenous infection) or urinary bladder and kidneys (intravesical infection) were plated on plain and antibiotic-containing plates. Both wild-type and mutant strains grow on LB plates, whereas only a mutant strain can grow on antibiotic containing LB plates. Plate counts were used to calculate the number of wild-type and mutant bacteria surviving in vivo. Competitive indices (CI) were calculated as the ratio of mutant over wild-type in tissues to the ratio of mutant over wild-type in the inoculum. Urine was collected from a group of 5 mice infected with CFT073, periodically over 48 hours and immediately stabilized with RNAprotect (Qiagen) prior to RNA extraction.
Genomic DNA was isolated from the inoculum used for infections (input) and from cultures derived from infected spleens (output) using DNeasy blood and tissue DNA extraction kit (Qiagen). Genomic DNA (5 µg) was sheared to yield ∼300 bp fragments (Covaris). Illumina Truseq adapters were ligated to DNA fragments and used for Tn-specific amplification. A Tn-specific primer composed of the flowcell binding region of the Truseq adapter and Tn-specific region was used in conjunction with the Truseq adapter-specific primer to amplify transposon-chromosome insertion junctions (Table S1). Briefly, 25 ng of the TruSeq library was used as template for 30 cycles of amplification. Amplicons were further processed for Illumina sequencing (cluster generation) according to manufacturer's recommendations and sequenced using a Tn-specific primer (Table S1). Libraries from input and output samples were sequenced in two separate lanes of a single sequencing run in an Illumina HiSeq2000 sequencer. Library preparation and sequencing were performed at the University of Michigan DNA core facility.
Reads from the input and output libraries, in FASTQ format, were aligned to the genome of E. coli CFT073 (NCBI accession no. NC_004431) [14] using the short read aligner BOWTIE [16]. The alignment files, in SAM format, were then used in the TFAST [17] program to determine the number of reads that map to a given chromosomal location in the input and output libraries.
To assess biofilm production, strains were cultured in tryptic soy broth containing 1% glucose in 96-well tissue culture-treated polystyrene plates for 24 h, at 37°C. Supernatants were aspirated and plates were washed three times with water and stained with 0.3% crystal violet solution for 10 min. Unbound crystal violet was removed by three additional washes with water. Biofilm-bound crystal violet was dissolved in 200 µl of 33% acetic acid and absorbance was measured at 590 nm (μQuant, Bio Tek instruments, Inc.). This experiment was repeated at least three times, independently.
The protocol described by Cerca et al. [47] was adapted. Cultures, incubated overnight in tryptic soy broth containing 1% glucose, were adjusted to an OD600 of 1.5. Cell pellets from 1 ml samples were resuspended in 300 µL of 0.5M EDTA (pH 8.0) and boiled for 5 min. Samples were centrifuged at 13,000 RPM for 6 min. Supernatants were treated with Proteinase K (2 µg/µL), heat inactivated and diluted 3-fold in Tris-buffered saline (TBS; 20 mM Tris-HCl, 150 mM NaCl, pH 7.4). Extracts (200 µl/sample) were immobilized on nitrocellulose membranes and blocked with 5% skim milk in TBST (TBS containing 0.1% Tween20) for 2 h. Blots were incubated for 2 h with an affinity-purified anti S. aureus PNAG antibody (1∶2000) raised in rabbits [61]. Horseradish peroxidase-conjugated secondary anti-rabbit IgG antibody (1∶20,000) was used in conjunction with ECL Plus enhanced chemiluminescence detection system (GE Healthcare) to determine the presence of PNAG. These experiments were repeated at least three times, independently.
Agar (0.25%) plates containing NaCl (0.5%) and tryptone (1%) were used to measure swimming motility. Ampicillin (100 µg/µl) was added for plasmid maintenance, when required. Cultures were stab-inoculated and incubated at 30°C for 16 h. Diameters (mm) of swimming zone were determined. Four independent experiments were performed with at least two technical replicates.
E. coli CFT073 (pTopo) and CFT073 (pSS1) were cultured overnight in TSB with 1% glucose and ampicillin (100 µg/ml). Bacteria, washed and resuspended in PBS to an OD600 of 1, were incubated in ECM protein coated plates (Biocoat plates, Becton Dickinson) for 2 h at 37°C. The number of bacteria in the inoculum and the number of bacteria that remain in the supernatant (non-adherent) were determined by plate counts and used to calculate percentage of adherent bacteria. Fold-change in adherence was calculated as the ratio of adherence percentages of CFT073 (pSS1) over CFT073 (pTopo).
The following immortalized cell lines were used in adherence assays: human bladder epithelium, UM-UC-3 (ATCC CRL-1749); murine kidney, MM55.K (ATCC CRL-6436); green monkey kidney, VERO (ATCC CCL-81); and human embryonic kidney, HEK293 (ATCC CRL-1573). Cells were cultured to confluence in 24-well cell culture plates (Corning) in Dulbecco's Modified Eagle Medium supplemented with 10% fetal bovine serum, penicillin (100 U/ml), streptomycin (100 µg/ml) and L-glutamine (0.3 mg/ml), referred to as DMEM-PSG, at 37°C in a humidified atmosphere with 5% CO2. Epithelial cell cultures were washed once with PBS, and inoculated with a 250 µl suspension containing 1×108 CFU of E. coli CFT073 (wild-type), c2394-95 mutant, or the complemented mutant in DMEM without antibiotics. Infected epithelial cells were incubated at 37°C with 5% CO2 for 30 min, and then washed three times with PBS. Epithelial cells, along with any adherent bacteria, were lifted by incubation in 1 ml sterile distilled water containing 5 mM EDTA. Colony counts were used to enumerate CFUs in the inocula and cell-associated bacteria. Adherence was expressed as cell-associated CFU/initial CFU. Adherence of each mutant was normalized to the wild-type control; assays were performed in triplicate, each with three technical replicates.
E. coli CFT073 (wild-type), c2394-95 mutant, and the complemented mutant were cultured for 3 h at 37°C. Samples were swirled gently and 10 µl of the culture was dropped onto formvar carbon support film on TEM specimen grids (Electron Microscopy Sciences). Grids were incubated at room temperature for 5 min, and excess medium was wicked off with filter paper. Grids were washed once with 10 µl of deionized water, and then stained for 2 min with 10 µl of 1% phosphotungstic acid (pH 6.8). Excess stain was removed; grids were washed immediately with deionized water and dried. Grids were visualized using a Philips CM-100 transmission electron microscope.
Overnight cultures, diluted 1∶100 in fresh medium, were incubated at 37°C for 2 h. Cultures were exposed to polymyxin B (4 µg/mL) for 30 min. Colony counts were determined by plating and percent survival upon exposure to Polymyxin B was calculated as the ratio of CFU in the treated samples to untreated controls. The experiment was repeated at least three times, independently.
RAW 264.7 cells were cultured in RPMI1640-PSG supplemented with 10% fetal bovine serum and seeded in 24-well tissue culture plates. CFT073 and mutant strains, cultured overnight in LB, were washed in PBS and resuspended in RPMI1640 to an OD600 of 0.004. Cells were washed with PBS, overlaid with the inoculum at an MOI of 10 and incubated for 30 min. Two identical plates, for 0 h (T0) and 2 h (T2), were set up during each experiment. Supernatants were aspirated and cells were washed three times with PBS. Fresh RPMI1640 supplemented with gentamicin (200 µg/ml) was added and incubation was continued. The T0 plate was removed at 15 min post gentamicin addition and cells were lysed with saponin (10%, w/v in water). Lysates were plated to determine the number of intracellular bacteria. T2 plates were processed as described here at 2 h post gentamicin addition. Percent killing was calculated as the percent of intracellular bacteria that were killed within RAW 264.7 cells. Killing percentages of mutants were compared to that of wild-type bacteria to determine the comparative fitness of a given mutant during survival within RAW 264.7 cells. The experiment was repeated three times with three technical replicates per strain.
RNA was extracted from E. coli CFT073 cultured in LB or in filter sterilized human urine to mid-exponential phase or from cells harvested from urine of mice infected with wild-type strain using RNeasy mini kit (Qiagen). Contaminating DNA was removed by incubation with DNase (Turbo DNA-free, Ambion) and reverse transcribed using SS RT III (Invitrogen). To determine co-transcription, cDNA, genomic DNA and RNA samples were used as templates in standard PCR reaction (primers listed in Table S1). The entire experiment was repeated twice, independently.
Overnight cultures of E. coli CFT073 were diluted 1∶100 in LB or LB with 300 µM dipyridyl (Sigma) or LB with 10 µM ferric chloride and incubated for 2 h at 37°C. RNA extraction and cDNA synthesis were performed as described above. Expression of yddA, yddB and pqqL transcripts was determined by qPCR using SYBR green chemistry (Agilent Technologies) in a Stratagene Mx3000P thermal cycler (Stratagene). Transcripts were normalized to gapA mRNA (Table S1) and relative quantification was performed using CFT073 cultured in LB as the calibrator.
Overnight cultures of E. coli CFT073 and Δfur mutant were diluted 1∶100 in LB and incubated for 2 h. RNA extraction, cDNA synthesis and qPCR were performed as described above. Relative quantification was performed using CFT073 as the calibrator. Both qPCR experiments were repeated three times with two technical replicates.
DNase-treated RNA from mouse UTI urine was used to determine levels of pgaA transcript by qPCR, essentially as described above. Mid-exponential phase cells from LB were used as calibrator and all samples were normalized to gapA levels.
Statistical tests were performed using Prism 5 (www.graphpad.com). Data were analyzed using the following tests: co-infection experiments, Wilcoxon signed-rank test against a theoretical median of 0; biofilm assay, swimming motility assay and adherence to epithelial cells and ECM proteins, two-way ANOVA with Bonferroni's multiple comparison test; polymyxin B resistance assay and intracellular survival assay, student's t test. P<0.05 was considered as a statistically significant difference. Error bars in the figures represent standard error of the mean.
The raw reads can be accessed under the accession number, SRP027190 in NCBI SRA.
|
10.1371/journal.pcbi.1006593 | A regularity index for dendrites - local statistics of a neuron's input space | Neurons collect their inputs from other neurons by sending out arborized dendritic structures. However, the relationship between the shape of dendrites and the precise organization of synaptic inputs in the neural tissue remains unclear. Inputs could be distributed in tight clusters, entirely randomly or else in a regular grid-like manner. Here, we analyze dendritic branching structures using a regularity index R, based on average nearest neighbor distances between branch and termination points, characterizing their spatial distribution. We find that the distributions of these points depend strongly on cell types, indicating possible fundamental differences in synaptic input organization. Moreover, R is independent of cell size and we find that it is only weakly correlated with other branching statistics, suggesting that it might reflect features of dendritic morphology that are not captured by commonly studied branching statistics. We then use morphological models based on optimal wiring principles to study the relation between input distributions and dendritic branching structures. Using our models, we find that branch point distributions correlate more closely with the input distributions while termination points in dendrites are generally spread out more randomly with a close to uniform distribution. We validate these model predictions with connectome data. Finally, we find that in spatial input distributions with increasing regularity, characteristic scaling relationships between branching features are altered significantly. In summary, we conclude that local statistics of input distributions and dendrite morphology depend on each other leading to potentially cell type specific branching features.
| Dendritic tree structures of nerve cells are built to optimally collect inputs from other cells in the circuit. By looking at how regularly the branch and termination points of dendrites are distributed, we find characteristic differences between cell types that correlate little with other traditional branching statistics and affect their scaling properties. Using computational models based on optimal wiring principles, we then show that termination points of dendrites generally spread more randomly than the inputs that they receive while branch points follow more closely the underlying input organization. Existing connectome data validate these predictions indicating the importance of our findings for large scale neural circuit analysis.
| The primary function of dendritic trees is to collect inputs from other neurons in the nervous tissue [1,2]. Different cell types play distinct roles in wiring up the brain and are typically visually identifiable by the particular shape of their dendrites [3]. However, so far no branching statistic exists that reliably associates individual morphologies to their specific cell class [4,5], indicating that we have not yet identified the morphological features that are characteristic for the differences in how neurons connect to one another. Theoretical considerations have provided systematic qualitative insight into the question of how dendrite shape relates to specific connectivity. Dendrites are thought to collect their inputs using the shortest amount of cable and minimizing conduction times in the circuit [3,6–9] and they have been proposed to maximize the possible connection repertoire [10]. Of the possible connections that a neuron could make by anatomical proximity only a small, relatively invariable number become functional synapses [11]. But it has generally been proposed that the connection probability between a dendrite and an axon can be determined by the amount of anatomical overlap between the two [12–14]. Furthermore, dendrite shape has been linked to the number of synapses based on the optimal wiring assumptions described above, linking total dendrite length and the number of synapses that determine the morphology [15]. This leads to the question whether specific axonal arrangements or synapse distribution patterns may coincide with specific typical dendritic morphological characteristics [16].
A useful concept to relate dendritic trees with their underlying connectivity comes from extended minimum spanning trees (MSTs) that connect a set of target points while minimizing total cable length and path lengths in the tree toward the root where signals get integrated [6,7]. Such MSTs were shown to produce accurate dendritic morphologies when the corresponding target points were selected adequately [7,15,17,18]. This approach has previously linked the distribution of target points to actual synapse locations, as well as to the number of branch points (BPs) and termination points (TPs) [15]. Here, we study the relationship between local statistics of spatial input distributions and the respective dendritic morphology using MSTs generated on different target point distributions. In order to do this we use a regularity index R that measures the degree of clustering of points in a given volume based on the average nearest neighbor (NN) distance. Specifically, the statistic R is defined as the ratio between the observed mean NN distance of a set of points in a given volume and the expected mean NN distance of the same number of points distributed uniformly in the same volume, i.e. in a setting of complete spatial randomness. Randomly distributed points from a uniform distribution (i.e. samples from a Poisson process) thereby yield a value of R = 1 in the limit. When R > 1, NN distances are on average larger than in a uniform random distribution, meaning that the points are distributed more regularly. When R < 1, NN distances are on average shorter, indicating that the points are more clustered than expected by chance. The measure R has been used in a wide variety of scientific disciplines, such as physics, biology, geography and astronomy [19–22]. In neuroscience, a variant of R was used to describe the regular spacing of ganglion cells in the retina [23–25]. In particular, Cook performed a comparison of measures based on NN distances to study retinal mosaics [26], including R, which they refer to as the dispersion index. The statistic R has also been applied to graph theoretical problems such as minimum spanning trees [27], but to the best of our knowledge it has not yet been considered to characterize neuronal morphology. In the following, we first estimate R for BPs and TPs in real dendrites and then use MST-based morphological models to compare distributions of BPs and TPs with the underlying distribution of target points as a proxy for their corresponding synaptic input distributions.
The value of the statistic R, defined as the quotient of the average nearest neighbor (NN) distance to the one expected from a matching uniform distribution can be obtained for any given set of points in Euclidean space. We first calculated R for the set of BPs and TPs in real dendrites to estimate how regularly the dendrites spread in the circuit (Fig 1). One issue when computing R on finite point clouds that has been given little attention in the literature so far is that a naïve calculation yields a biased result due to boundary effects. This bias is due to the fact that the enclosing volume of the point cloud is usually not known. In the special case of planar convex volumes a bias correction can be performed analytically [28], but these methods are not available for higher-dimensional and non-convex carrier volumes such as the ones occurring in our case. We addressed this issue by developing a Monte Carlo (MC) based approach to estimate for a given volume the reference value for a Poisson process numerically and obtain the value of R without edge effects. This method also provided us with specific confidence intervals for our estimates (see Methods and S1–S4 Figs). We then estimated R for BPs and TPs using our MC based approach, resulting in separate statistics RBP and RTP for the sets of branch and termination points, respectively (Fig 2).
The mean estimated values of R in four three-dimensional (3D) and four two-dimensional (2D) cell types varied widely. For almost all cases we observed a tendency of RTP being slightly larger than RBP. To study whether there were significant differences between the RBP and RTP values of different cell types, and because not all of the necessary assumptions for ANOVA were satisfied, we used the Kruskal-Wallis test and then applied the Mann-Whitney test with the Bonferroni method to adjust the p-values for pair-wise comparisons. Using the Kruskal-Wallis test we found significant differences (p-value < 2.2 × 10−16) in the four tests performed, namely between the RBP values in 3D cells, RTP in 3D cells and the two analogous cases in 2D. In 3D dendrites, the spatial distribution of BPs and TPs was most clustered in motoneurons, followed by hippocampal pyramidal cells, neocortical pyramidal cells, and finally dentate granule cells (Fig 2A). Pair-wise comparisons revealed that there was no significant difference between the RBP values of neocortical pyramidal cells and hippocampal pyramidal cells or between RTP values of neocortical pyramidal cells and granule cells. In the case of the four planar cell types (Fig 2B), dendritic arborization (da) neurons in the fly larva were well characterized by the clustering of their BPs. We found no significant difference between the RTP values of Lobula Plate tangential cells (TCs) in the fly, cerebellar Purkinje cells and Retinal ganglion cells, a large inhomogeneous group of cell types. Pair-wise comparison also showed no significant difference between the RBP values of Purkinje cells and Retinal ganglion cells. In view of the above results, the only cell types where no differences were detected either in the R values for their branch or termination points were Purkinje cells and Retinal ganglion cells. However, it should be noted that our data set only contained 15 reconstructions of Purkinje cells. The results for Purkinje cells could therefore be dependent on the small number of data in that group. It is important to note that all eight populations in Fig 2 were composed of subgroups with strong differences in their functional role in the nervous system. Moreover, morphologies within the separate subgroups were partly obtained in different species, preparations and developmental ages. To illustrate the effect this can have on the analysis, we dissected fly da neurons and TCs into their respective characteristic subgroups (Fig 3). Da neurons are known to subdivide into morphologically distinct classes (I–IV) and using our statistic R these can be separated into clusters (Fig 3A), corresponding to their specific R values. In particular, class III da neurons with their large number of small terminal segments (STSs) exhibited small R values consistent with the clustering of BPs and TPs due to these STSs. On the other hand, sub-classes of TCs (two types of horizontal system cells—HSN and HSE, and three types of vertical system cells—VS2, VS3 and VS4) did not separate into different clusters according to their R values (Fig 3B). This was not surprising since TCs were previously characterized in detail using morphological models and shown to have similar inner branch rules even though their spanning areas are easy to distinguish [17]. In all TC classes, TPs were more regularly distributed than BPs and all RTP values were close to 1, indicating configurations close to complete spatial randomness.
We also tested if the average distance to the NN of BPs and TPs of each individual cell corresponds to that of a uniform random pattern. The p-values for the tests were computed using the simulations of Poisson point cloud instances with the observed number of points, generated with our MC based approach. Considering 2D cells, da neurons showed a strong tendency to clustering in both their BPs and TPs: in 97.06% and 79.41% of the cases we rejected that the distribution of BPs and TPs, respectively, was uniform random in favor of a clustered distribution. All 3D cell types, except for dentate granule cells, showed rejection of a random distribution in their BPs with high confidence, in favor of a clustered distribution. In general, only few 2D or 3D cell types hinted to regular BP or TP distributions (see S1 Table for detailed results of all analyzed cell types). The statistic R for BPs and TPs is therefore a useful measure to distinguish between cell classes and characterize the relationship between dendritic tree structure and input architecture. However, it remains to be shown that the use of this local statistic in dendritic morphology is not simply an altered version of another traditional branching statistic. In order to test this and to check whether the input architecture as measured by R is reflected in other branching statistics, we computed the pairwise correlations between R and other commonly used statistics in 3D (Fig 4A) and 2D (Fig 4B) dendrites. We found that RBP and RTP do not have strong correlations with other typical branching statistics of dendritic trees in both cases. Since RBP and RTP were different in distinct cell types and were weakly correlated with other branching statistics, we postulate that these measures are a useful addition to the collection of branching statistics used to classify dendritic morphology.
In order to estimate how clustered or regular distributions of synapse locations co-depend with the clustering of BPs and TPs, we generated morphological models targeting different sets of input points with specified values RInput of their statistic R. This process required a target point cloud generator for specified R values. To obtain a wide range of sets of target points with specific R values and number of target points, we started with a set of points obtained from a uniform random distribution; we then iteratively moved the input points until the point cloud reached a set target R value (see Methods and Fig 5). Dendrites were considered as tree structures connecting these target points [7,15].
We computed synthetic branching structures connecting the target points with minimal resources using the extended minimum spanning tree (MST) algorithm [7]. The MST connecting the target points minimizes at the same time both total cable length and the path length from any point along the tree to the root, using a parameter which weighs the two costs via a balancing factor (bf; see Methods). Fig 6A shows planar and 3D sample trees obtained from connecting 100 target points in a 200 μm x 200 μm square and 200 μm x 200 μm x 200 μm cube with different values of RInput, using the extended minimum spanning tree for different values of bf. Tree structures on sets of target points connected with the MST algorithm were previously studied for periglomerular neurons in the olfactory bulb and they were shown to approximate numbers and features of actual synapses [15]. Studying the relationship between target points and MSTs will therefore be informative about the relationship between synapses and dendritic trees.
We generated morphological models on different sets of 2D and 3D target points with specific values of RInput and found that with higher RInput the trees became denser and the branches were more regularly distributed. This can be clearly observed in the 3D sample trees of Fig 6 as RInput increases. As might be expected, more regularly distributed inputs generally resulted in more regular branching structures (Fig 6B for 3D cases in Fig 6A). Compared with the input point distributions, BPs and TPs describing the dendritic geometry were more regularly distributed in all cases where RInput < 1 and were slightly less regular than the input point distribution in cases where RInput > 1. Furthermore, the spatial organization of BPs and TPs was clearly different: in line with reconstructions of real dendritic trees, RTP values were consistently closer to 1 while RBP followed RInput more faithfully. We obtained similar relations between RInput, RBP, and RTP for different numbers of points. RBP is therefore most likely a better estimate for the regularity of a neuron’s underlying synaptic input organization. To show that this was also the case in more detailed morphologies we used a morphological model from layer 5 cortical pyramidal cells [7] and obtained similar results in a realistic range of RInput = 0.7 ± 0.1 (Fig 7).
We compared this result with one biological dataset where synapse locations and dendritic reconstructions were available (Fig 8A), specifically for the aforementioned adult-born periglomerular neurons of the olfactory bulb [15,29,30]. Interestingly, the results for both actual reconstructions and MSTs constructed on the actual synapse locations were in line with our simulations from this study. These results indicate that RInput could be estimated using RBP whereas termination points in dendrites and MSTs are more randomly distributed, characterized by a value of RTP close to 1. We then compared RInput, RBP, and RTP in two connectome datasets from the fly (Fig 8B and 8C) [31,32]. Using serial electron microscopy it has increasingly become possible to obtain data of morphological reconstructions and the exact position of synapses between the same cells. Interestingly, we found the predictions from our model to be validated by these two datasets with RBP and RInput values being more similar and RTP being closer to 1. Overall, we believe that constructing synthetic versions of real dendritic trees with more detailed morphological models would be useful for inferring the underlying spatial organization of the synaptic inputs. This will provide invaluable predictions for connectome and large-scale neural circuit analyses.
Apart from the relations between RInput, RBP, and RTP, it is interesting to study the relation of RInput with branching statistics typically used to characterize dendritic trees (Fig 9). As was the case for RBP and RTP in real dendritic tree reconstructions, RInput was weakly correlated with other branching statistics, suggesting that input architecture is not well captured by traditional branching statistics whereas RBP and RTP would be useful measures for this and to classify dendritic morphology accordingly. However, both total length and number of branch points increased reliably with RInput, requiring the minimum spanning tree to use more cable to connect the points that are more widely spread and more branches to reach out to all distributed inputs in space. This correlation clearly affected the scaling behavior that was previously observed between number of inputs and total length as well as between number of inputs and number of branch points [15]. Here, the previously reported 2/3 power between these measures was not affected by RInput, but a clear increase in total length was observed as an offset in the relationship (Fig 10). MST-based dendrites connecting target points with an increased RInput required much more cable length.
We have presented here a new branching statistic for dendrites, the regularity index R, which is based on the average nearest neighbor (NN) distance between branch, termination or input points of a given dendritic tree, capturing the regularity of their respective distributions. Specifically, R is defined as the ratio of the observed average NN distance to the one expected in a matching random point cloud. This makes R independent of the absolute scale of the dendritic arbor, but rather captures the clustering characteristic of the branch and termination points and allows for comparison of cells of different sizes. We found that the measure allowed to distinguish dendritic trees from different cell classes for which the local statistics of the spatial input distribution differed. The values of R computed for the sets of branch points (RBP) and termination points (RTP) of reconstructions of real dendritic trees correlated little with most other commonly considered branching statistics, indicating that these measures provide new descriptive power for dendritic trees that was not captured by existing measures.
Using morphological models, we then found that, in the range of observed RBP and RTP values in real cells, overall RBP values were good predictors of the input distribution (RInput) while RTP values were generally closer to 1, implying more randomly distributed termination points in dendritic trees. We also showed that more regular input distributions with higher values of RInput showed increased dendritic total length in the model. An analogous increase in total length was observed in real dendrites with increasing RBP. Input targets can loosely be compared to synaptic inputs for certain cell types such as the periglomerular neurons of the olfactory bulb in Fig 8A [15]. While this indicates a specific link between the input organization and dendritic morphology it by no means implies a causal relation between input locations and resulting dendrite growth. Rather, dendrites and axons organize to implement possible connectivity patterns between neurons and specific dendritic morphology could constrain the space of potential synapses. However, the precise biological processes are not yet known.
On the one hand, the correlation between the synapse distribution and RBP can be a direct consequence of optimal wiring since minimum spanning trees shows the same correlation as real dendrites. But it is not known how the biological growth process leads to such optimal wiring. On the other hand, the precise distribution of branches within the range of optimally wired dendrites can vary substantially and the details of the growth process can lead to these differences in morphology. Increased self-avoidance of branches during dendrite development through molecular mechanisms such as Dscam would for example most likely lead to larger values of R [33]. Increased localized branching due to some accumulation of molecular cues could in the other extreme lead to clustered branches in mature dendrites leading to R values lower than 1. Overall, we expect the proposed measure R to be able to better predict these type of local features of the input organization of given dendritic tree types compared to other existing branching statistics. As more realistic morphological models based on minimum spanning trees become available, as is the case for example for TCs [17] and dentate gyrus granule cells [18], this information can be further refined.
The expected NN distance of a uniform random distribution used for calculating R has traditionally been obtained analytically, assuming infinitely many points in an unbounded volume. Yet, in practice all of our volumes V containing a given point cloud C were finite and bounded resulting in strong biases and edge effects if R is estimated in a naïve way (see Methods). Our approach to remedy these adverse effects was to use Monte Carlo (MC) simulations to predict the expected NN distances of uniform distributions numerically, using instances of Poisson point clouds. This procedure furthermore allowed for the specification of confidence intervals for the estimated values of R (S2–S4 Figs). We found that confidence intervals were mostly dependent on the number of points in the point cloud (decreasing as the number of points increases) and only to a lesser extent on the shape of its supporting volume. For small numbers of points (N < 20), the confidence intervals were large, making the estimated values of R less reliable (S3 Fig). Our MC based approach will be useful in further studies beyond the scope of dendritic morphology since point processes in any type of finite volumes will necessarily exhibit similar important boundary effects.
There are several ways in which the measure R could be generalized. First of all, for simplicity we assumed point clouds with uniform homogeneous densities when computing R. This can be extended to non-homogenous cases by studying local estimates of point densities. This would lead to a localized version of measure R. Secondly, we only considered one nearest neighbor per point. This allows for point clouds with different global clustering behavior to get assigned the same value of R (e.g. a point cloud where all points have the same coordinates and a point cloud where pairs of points have the same coordinates both get assigned a value of R = 0). But the statistic R can easily be extended to consider neighborhoods of higher order, containing the k nearest neighbors for each point, k ≥ 2. Furthermore, other quantities commonly used for the analysis of spatial point patterns such as the cumulative distribution function of the nearest-neighbor distances, known as G function, Ripley’s K function [34] and related quantities can be used. These extensions are subject of future studies. If neighbor distances are used to define simplicial complexes on a given point cloud, and the resulting complexes are examined using methods from algebraic topology, this leads to techniques used in topological data analysis [35]. Related techniques most recently also were proposed as methods to examine dendritic structures [36].
Overall, we presented a new statistic, the regularity index R, for dendrites that allows to relate the morphology of a neuron with the specific connectivity that it implements. It has low correlation with most commonly used statistics of dendritic branching and is extendable in several ways, providing a useful new statistic for the classification of dendritic trees.
The analysis was performed with the TREES toolbox [7] (www.treestoolbox.org), an open-source software package for MATLAB (Mathworks, Natick, MA). Functions r_mc_tree to calculate R values in existing trees, as well as PP_generator_tree to generate point distributions with given R values were newly implemented and are now available in the TREES toolbox package.
The average nearest neighbor (NN) ratio R=r¯0/r¯E compares the observed average NN distance r¯0 between a set of N points with the expected average distance r¯E between nearest neighbors under the assumption of a uniform random distribution (with the same number of points covering the same total area or volume). This approach was first described in [20,37].
R provides a measure of the clustering of the points in a point cloud C. Concretely, the closer the points are to a random (Poisson) distribution, the closer to 1 the value of R becomes (as the values of r¯0 and r¯E are more similar). Values of R less than 1 correspond to clustering (r¯0<r¯E). When all points overlap (R = 0) the most clustered condition is reached. For values of R greater than 1, nearest neighbors are further apart than it would be expected for a random distribution (r¯0>r¯E). In 2D arrangements, the most dispersed situation is the one in which the points are spaced on a triangular lattice, yielding a value of R = 2.1491 [20]. The measure R has the advantage of being easily interpretable. For example, R = 0.5 indicates that nearest neighbors are, on average, half as distant as expected under random conditions (c.f. Fig 5D). Note though that R only considers pairwise NN distances and e.g. cannot distinguish the case in which all pairwise NN distances are 0 (pairs of points have the same coordinates) from a fully clustered situation (all points of the point cloud have the same coordinates).
Formally, for a finite point cloud C, i.e., a set of N points, the average NN distance is
r¯0=1N∑i=1i≠jNmin{di,j},
where di,j denotes the Euclidean distance between the i-th and the j-th point in C. This is the numerator in the definition of R=r¯0/r¯E. The denominator in R is the expected NN distance r¯E for a Poisson process that can be analytically computed as r¯E=1/2ρ in the 2D case and as r¯E=Γ(4/3)/4πρ/33 in the 3D case, where Γ(∙) is the gamma function and ρ is the point density, i.e., the mean number of points per unit area or volume V. For a uniform random distribution, an unbiased estimator of ρ is ρ^=N/V. Thus, to obtain the point density, an accurate estimate of the supporting volume V of the point cloud C is required.
In order to estimate R, a volume V supporting a given point cloud C needs to be estimated. The most common way to do this is to use the convex hull of C. Yet, with this choice the supporting volume is overestimated if it is non-convex, which results in incorrect values of R (see for examples S1 Fig). Better estimates of R were obtained using α-shapes. α-shapes were devised to characterize the shapes of point clouds and can be seen as an extension to the notion of a convex hull [38,39]. Formally, to any given finite point cloud C in 2D or 3D Euclidean space a one parameter family of curves or surfaces Sα called α-shapes can be constructed, with α ϵ [0,∞]. By construction, S∞ corresponds to the convex hull and S0 to the point cloud itself. For any finite C, Sα is a finite set and a smallest value α0 exists (called critical value of α) such that Sα0 is connected and contains all points of C. Furthermore a smallest value αk < ∞ exists for which Sαk=S∞. The α-spectrum of C is defined as the monotonically increasing, finite sequence of values (αi)0≤i≤k, 0 ≤ αi < ∞, αi < αi+1 for which each Sαi is a distinct α-shape and the shapes do not change between two consecutive values αi,αi+1. To compute what we call a “tight hull” around a point cloud C we selected the center point αk/2 of the α-spectrum, for which we rounded the index k/2 to the next integer value. Especially for point clouds with non-convex supporting volumes, this yielded much better estimates of the true volume and thus less biased estimates of R. See S1B and S1C Fig for an example of the convex hull compared to the tight hull of point clouds with non-convex support, and the resulting values of R.
After estimating a supporting volume V using α-shapes as described in the previous section, naïve calculations of R still yielded biased results due to boundary effects. This is due to points in C for which the distance to the boundary of V was smaller than the average NN distance r¯0 in C. As the calculation of r¯E assumes that balls of radius r¯E surrounding all points are always completely contained in V, R was overestimated. Two proposed techniques correcting for such boundary induced biases are the so-called toroidal edge correction and the border area edge correction. The first one removes boundaries by transforming a (bounded) rectangular study region into a torus by identifying opposing edges. The second one specifies a buffer zone around the boundary of the study region and uses the part remaining in the middle as the new study region. Points in the buffer zone are used only to take measurements of points (NN distances in our case) that are within the new study region and are further discarded. Yet, both techniques have their drawbacks: the toroidal edge correction cannot be used for non-rectangular regions, as is the case of dendrites, and the border area edge correction discards a large number of available points, which makes it inappropriate for many dendrites for which the number of points was not large enough. Moreover, analytical bias corrections were also derived [28], but those require convex planar surface areas as supporting volumes.
Since most dendritic arbors form non-convex areas and volumes (S1A Fig) and many of them do not have a high number of BPs and TPs, instead of computing r¯E analytically from an estimate of the point density and using an edge correction technique, we used a Monte Carlo (MC) simulation approach to estimate r¯E. For a point cloud C consisting of N points contained in a volume V, we first computed r¯0 as the observed mean NN distance in C. We then sampled M = 100 uniform random point clouds within V, each containing N points. Importantly, we scaled each sampled point cloud so that its supporting volume (again computed using α-shapes) matched V. When this process, which we call volume correction, was not performed the estimates of R values were positively biased, especially for small point clouds (S2 Fig). For each of those point clouds we then computed the average NN distance, r¯Ei,i=1,…,M, and obtained an estimate of r¯E as the mean of the r¯Ei,i=1,…,M, values of the simulations. No edge corrections were necessary since all the average NN distances were biased by the same edge effects. To check the correctness and convergence properties of this approach, we generated point clouds with known R values and compared them to the R values estimated from our MC based method (S4 Fig). In each MC iteration we additionally estimated confidence intervals [ci−,ci+] for r¯Ei,i=1,…,M with confidence level 1 − α using 1,000 bootstrap samples. We then obtained the corresponding confidence intervals [c−,c+] for r¯E by computing the sample mean of the set of ci− and ci+,i=1,…,M, to obtain c− and c+, respectively. The confidence interval [c−,c+] for r¯E in turn leads to the statistic R as [r¯0c+,r¯0c−]. Throughout this study we used α = 0.05. We observed that the confidence intervals were mainly influenced by the number N of points in the point cloud and to a much lesser extent by the shape of the supporting volume V. We assessed this by computing confidence intervals both for planar sample configurations with known R value (S2 Fig) in a square area and for the 3D cell classes considered in this work (S3 Fig).
To evaluate the measure R on real cells, we obtained a number of reconstructions of dendritic trees from NeuroMorpho.org [40], Version 7.0 (released on 09/01/2016) using the TREES toolbox [7]. Specifically, we chose reconstructions belonging to eight well-known cell classes for our investigations, namely cortical pyramidal cells, hippocampal pyramidal cells, dentate granule cells, motoneurons, retinal ganglion cells, cerebellar Purkinje cells, fly larva dendritic arborization (da) neurons and fly Lobula Plate tangential cells (TCs). The first four classes were 3D cells and the last four classes were 2D. For selecting the reconstructions, we obtained all reconstructions from NeuroMorpho.org that were classified as either having "moderate" or "complete" reconstructions of their dendritic trees and belonged to the control group (to exclude mutant cells). We then grouped all reconstructions by archive and sorted out archives that contained poor reconstructions by manual visual inspection as well as archives containing one cell only. This left us with a number of reconstructions of each cell type, denoted in parentheses in the following list: cortical pyramidal cells (3784), hippocampal pyramidal cells (399), dentate granule cells (154), motoneurons (82), retinal ganglion cells (322), cerebellar Purkinje cells (15), fly da neurons (68), fly TCs (55). After downloading the reconstructions in the SWC format, these were read into and pre-processed using the TREES toolbox. For each reconstruction, this process involved deleting the soma and the axon if present (function delete_tree for the SWC regions 1, 2, 5 and 10) and then re-joining the parts of the tree if the deletion operation yielded several roots (function catx_tree), followed by a final removal of higher order multifurcations (function repair_tree). This process was not necessary for fly TCs that were available with the TREES toolbox [17]. Da neurons were furthermore subdivided into da class I-IV cells and TCs into horizontal system northern (HSN) cells, horizontal system equatorial (HSE) cells, and vertical system (VS2, VS3, and VS4) cells (Fig 3). For each tree we then computed a number of statistics: total dendrite length, number of BPs, mean branch order of BPs and TPs, mean branch angle, mean asymmetry [41], mean path length, R for the set of BPs (RBP), R for the set of TPs (RTP), volume of the convex hull, volume of the tight hull and cable density as total length per volume in the tight hull. The tight hulls as well as their volumes needed for estimating R were computed using α-shapes as described previously (function boundary).
In order to study a wide range of different spatial input organizations in the morphological model we implemented a procedure for obtaining point clouds with specified R values. First, we generated a number N of random points within a square or cube. We then iteratively estimated the R value using our MC method and moved each point in the direction of or away from its NN, depending on whether the target R was smaller or greater than the current R, respectively, (Fig 5B) until the target R value was reached. The shift was proportional to the difference between the current R value and the target R, i.e., the closer the values of both, the smaller the movements. Fig 5C shows the number of iterations required for our algorithm to reach different values of R, from highly clustered (R = 0.2) to highly regular (R = 1.8) given 1,000 initial points. We obtained very similar results for different numbers of points.
Using the point pattern generator described above we generated a large number of point clouds in 2D or 3D spaces. Planar arrangements were fixed to 200 μm x 200 μm and 3D arrangements were set to 200 μm x 200 μm x 200 μm. A large variety of number of points (50–400 points) and R values (0.2–1.8) were computed. We subsequently computed morphological models based on optimal wiring principles that connected these point clouds. Optimal wiring of the underlying point cloud was implemented using a minimum spanning tree as described in [7] using the algorithms available in the TREES toolbox [7,42]. Briefly, optimal wiring minimizes both total cable length and the path length from any point along the tree to the root, using a balancing factor bf to weigh the second cost (that is: total cost = cable length cost + bf ∙ path length cost). For bf = 0 the algorithm only seeks to minimize the total cable length while for large bf it seeks also to minimize the length of the connections from the root to any point. Values of bf greater than 0 and less than 1 represent a mixture of the two objectives that are realistic for real dendrites. As a further constraint, we did not allow multifurcations (more than two daughter branches at each BP) in the computed synthetic trees.
The minimization was achieved via a greedy minimum spanning tree algorithm [43]. We computed synthetic dendritic trees from all the point clouds, connecting the points to a root in the center and using bf values from 0.2 to 0.8. We obtained 100 trees for each individual condition (point density, R and bf value). For each synthetic dendritic tree, RInput of its target points was known (since the inputs were obtained using the point pattern generator), and we estimated R values of its BPs (RBP) and TPs (RTP) in order to study the relation between these measures. In addition, we analyzed the relation between RInput and other branching statistics commonly used to describe dendritic morphology. Specifically, we studied the total length, the number of branching points, the mean path length from the root of BPs and TPs, the mean branching angle, the mean branching order and the mean asymmetry at the branching points of each synthetic dendritic tree. The asymmetry for each branching point was defined as the ratio of v1/(v1 + v2) for v1 < v2, where v1 and v2 are the counts of TPs in each of the two daughter branches. All statistics were computed using the TREES toolbox.
We also considered a possible exclusion zone around the target points due to the physical dimension of synapses that these might represent. This is an optional parameter ε in the PP_generator_tree function, a minimal distance between all points in the resulting point pattern. All morphological models from Figs 6, 9 and 10 were recalculated with ε = 0.5 μm with very similar results (S5 Fig). The simpler results without considering volume exclusion are shown in the main manuscript. It is worth noting that real neurons grow in a packed tissue and volume exclusion also most likely plays a role in the growth process of the dendritic trees. However, simulating dendrite growth in the packed tissue was outside of the scope of this study.
We used a model previously described in Cuntz et al., 2010 [7], to check whether the conclusions for simple abstract morphological models hold for more realistic morphologies. The construction pipeline has been well described previously but was adapted to rearrange the targets for apical, oblique and basal dendrites separately to match a given target RInput using our new PP_generator_tree function. Only then were the targets connected to MSTs as described earlier. Synthetic morphologies (N = 332) were generated for Fig 7 with RInput values from a normal distribution around 0.7 ± 0.1. The numbers were chosen to approximately simulate a distribution of dendritic parameters for cortical pyramidal cells (see Fig 2). Sample morphologies were generated with the same initial target point distribution altered to specifically match RInput of 0.6, 0.8 and 1.2 using PP_generator_tree.
Data in Fig 8 were obtained from three different datasets and processed in similar ways to obtain adequate values for RInput, RBP, and RTP. The same new TREES toolbox function r_mc_tree was used to calculate the R values separately on the synapse locations, branch points and termination points separately. The olfactory bulb dataset [29,30] was used similarly as previously [15]. The fly medulla dataset was obtained during a Hackathon at Janelia Farm Campus (https://github.com/janelia-flyem/SevenMedullaColumnConnectome) [31]. The fly larva dataset comes from an analysis of neurons of the peripheral nervous system at two larval developmental stages [32]. For the fly connectome datasets very tight hulls (α = 0.9) were used to avoid including lengthy axonal segments without synapses or branch points. Both inputs and outputs were included in the analysis since fly neurites contain both presynaptic and postsynaptic puncta on the same branches.
|
10.1371/journal.ppat.0030068 | Deletion of IL-4Rα on CD4 T Cells Renders BALB/c Mice Resistant to Leishmania major Infection | Effector responses induced by polarized CD4+ T helper 2 (Th2) cells drive nonhealing responses in BALB/c mice infected with Leishmania major. Th2 cytokines IL-4 and IL-13 are known susceptibility factors for L. major infection in BALB/c mice and induce their biological functions through a common receptor, the IL-4 receptor α chain (IL-4Rα). IL-4Rα–deficient BALB/c mice, however, remain susceptible to L. major infection, indicating that IL-4/IL-13 may induce protective responses. Therefore, the roles of polarized Th2 CD4+ T cells and IL-4/IL-13 responsiveness of non-CD4+ T cells in inducing nonhealer or healer responses have yet to be elucidated. CD4+ T cell–specific IL-4Rα (LckcreIL-4Rα−/lox) deficient BALB/c mice were generated and characterized to elucidate the importance of IL-4Rα signaling during cutaneous leishmaniasis in the absence of IL-4–responsive CD4+ T cells. Efficient deletion was confirmed by loss of IL-4Rα expression on CD4+ T cells and impaired IL-4–induced CD4+ T cell proliferation and Th2 differentiation. CD8+, γδ+, and NK–T cells expressed residual IL-4Rα, and representative non–T cell populations maintained IL-4/IL-13 responsiveness. In contrast to IL-4Rα−/lox BALB/c mice, which developed ulcerating lesions following infection with L. major, LckcreIL-4Rα−/lox mice were resistant and showed protection to rechallenge, similar to healer C57BL/6 mice. Resistance to L. major in LckcreIL-4Rα−/lox mice correlated with reduced numbers of IL-10–secreting cells and early IL-12p35 mRNA induction, leading to increased delayed type hypersensitivity responses, interferon-γ production, and elevated ratios of inducible nitric oxide synthase mRNA/parasite, similar to C57BL/6 mice. These data demonstrate that abrogation of IL-4 signaling in CD4+ T cells is required to transform nonhealer BALB/c mice to a healer phenotype. Furthermore, a beneficial role for IL-4Rα signaling in L. major infection is revealed in which IL-4/IL-13–responsive non-CD4+ T cells induce protective responses.
| Leishmaniasis is a disease induced by a protozoan parasite and transmitted by the sandfly. Several forms of infection are identified, and the different diseases have wide-ranging symptoms from localized cutaneous sores to visceral disease affecting many internal organs. Animal models of human cutaneous leishmaniasis have been established in which disease is induced by infecting mice subcutaneously with Leishmania major. Different strains of inbred mice have been found to be susceptible or resistant to L. major infection. “Healer” C57BL/6 mice control infection with transient lesion development. The protective response to infection in this strain is dominated by type 1 cytokines inducing parasite killing by nitric oxide. Conversely, “nonhealer” BALB/c mice are unable to control infection and develop nonhealing lesions associated with a dominant type 2 immune response driven by cytokines IL-4 and IL-13. However, mice deficient in IL-4/IL-13 signaling are not protected against development of cutaneous leishmaniasis. Here we describe a BALB/c mouse where the ability to polarize to a dominant type 2 response is removed by cell-specific deletion of the receptor for IL-4/IL-13 on CD4+ T cells. These mice are resistant to L. major infection similar to C57BL/6 mice, which highlights the role of T helper 2 cells in driving susceptibility and the protective role of IL-4/IL-13 signaling in non-CD4+ T cells in BALB/c mice.
| Experimental Leishmania major infection is widely used to explore the control of T helper 1 (Th1)/Th2 differentiation and elucidate mechanisms underlying susceptibility/resistance to intracellular microbial infection [1,2]. Typically, susceptible BALB/c mice infected subcutaneously with L. major develop severe pathology, manifested by progressive lesion development, necrosis, and death, while resistant C57BL/6 mice are able to control and heal dermal lesions [3]. Nonhealing disease in BALB/c mice is associated with a Th2 response characterized by secretion of mainly IL-4, IL-5, IL-9, and IL-13 [2,4–7], high anti-Leishmania antibody titres, arginase-1 production by macrophages [8,9] and visceral dissemination of parasites [10]. In contrast, resistance to L. major infection is mediated by development of a protective Th1 response, in which sustained IL-12 production, interferon-γ (IFN-γ) release and macrophage killing via effector nitric oxide (NO) production catalyzed by inducible NO synthase (iNOS) underlie protective responses [9,11–14].
CD4 T cell–derived cytokines drive L. major responses, and, as such, events that control T cell differentiation in response to L. major appear to be critical for disease outcome [15]. Disruption of Th1 differentiation by neutralization of IL-12 renders resistant C57BL/6 mice susceptible, whereas susceptible BALB/c mice treated with IL-12 become resistant to L. major infection [12]. IL-12 production must be sustained to control infection [13]. While both resistant and susceptible mice produce IL-4 early after infection [16,17], production of this cytokine is sustained in susceptible mice and transient in resistant mice [16–18]. Neutralization of IL-4 allowed control of L. major infection in BALB/c mice [19]. Subsequent studies in knockout mice proved that IL-4 was indeed important but not the sole mediator of susceptibility in BALB/c mice. L. major infection was controlled in BALB/c IL-4−/− mice, but parasite burdens remained greater than those of resistant animals [6,20]. These observations remain controversial, with some laboratory strains developing IL-4–independent susceptibility and indicating that further factors are involved [21]. IL-13 has been implicated as a susceptibility factor in L. major infection [4]. Susceptible IL-13 transgenic C57BL/6 mice develop impaired IL-12 and IFN-γ production during acute leishmaniasis, while IL-13−/− BALB/c mice remain comparatively resistant [4,22]. IL-13 can influence Th1 differentiation by modulating macrophage function and suppressing secretion of NO, IL-12, and/or IL-18 [22,23], partially attributed to IL-4/IL-13 activated alternative macrophages (aaMphs), recently demonstrated in mice deficient for this activation status [9,24].
IL-4 and IL-13 share a common signaling pathway through the IL-4 receptor α (IL-4Rα) chain. A functional IL-4R (type I) requires assembly of IL-4Rα with a γc chain, while interaction of IL-4Rα with an IL-13Rα1 subunit leads to formation of a functional IL-13 receptor (type II) [25]. IL-4Rα–deficient mice therefore lack responsiveness to IL-4 and IL-13. Careful analysis of footpad swelling and lesion development showed that initial control of L. major infection is equivalent in IL-4−/− and IL-4Rα−/− BALB/c mice. However, in contrast to IL-4−/− mice, IL-4Rα−/− mice develop progressive chronic disease. These data clearly indicate a protective role for IL-13 signaling in protection against chronic L. major infection, at least in the absence of IL-4 responsiveness [20].
Expression of IL-4Rα reflects the pleiotropic nature of IL-4 biology, as this receptor subunit is expressed upon a wide range of cells [26]. Given the central role of T cells in controlling L. major infection [15] and of IL-4 in driving Th2 responses [27], CD4+ T cell–specific IL-4Rα knockout mice were generated to elucidate the role of IL-4Rα–mediated signaling in CD4+ T cells independently of non-CD4+ T cell populations. Our results demonstrate a successful generation of transgene-bearing hemizygous LckcreIL-4Rα−/lox BALB/c mice that have effective deletion of IL-4Rα on CD4+ T cells, an incomplete deletion on CD8+ T cells and other T cell subpopulations, and normal expression on non–T cells. LckcreIL-4Rα−/lox mice infected with L. major developed a healing disease phenotype and clinical immunity similar to genetically resistant C57BL/6 mice. Consequently, our studies demonstrate that impairment of IL-4Rα–dependent Th2 polarized CD4+ T cells in the presence of IL-4/IL-13–responsive non-CD4+ T cells is required to transform nonhealer BALB/c mice to a healer phenotype.
Recently established IL-4Rαlox/lox BALB/c mice [24] were intercrossed with BALB/c mice expressing Cre-recombinase under control of the T cell–specific promoter Lck [28] and IL-4Rα−/− BALB/c mice [20] to generate LckcreIL-4Rα−/lox mice (Figure 1A). IL-4Rα hemizygosity (−/lox) increases probability of Cre-mediated deletion of the “floxed” allele [24]. LckcreIL-4Rα−/lox mice were identified by PCR genotyping (Figure 1B). Fluorescence-activated cell sorter (FACS) analysis of IL-4Rα surface expression confirmed efficient deletion on CD3+CD4+ T cells from LckcreIL-4Rα−/lox mice when compared with IL-4Rα−/− and IL-4Rα−/lox BALB/c (WT) controls (geometric mean channel florescence [geo. mean]: WT = 18.11, IL-4Rα−/− = 8.5, LckcreIL-4Rα−/lox = 9.48), but incomplete and variable deletion efficiency was observed on CD8+ T cells (Figure 1C and Figure S1) (geo. mean: WT = 18.69, IL-4Rα−/− = 9.06, LckcreIL-4Rα−/lox = 13.96) and γδ+ (geo. mean: WT = 7.6, IL-4Rα−/− = 3.15, LckcreIL-4Rα−/lox = 6.72) and NK–T cells (geo. mean: WT = 9.03, IL-4Rα−/− = 5.25, LckcreIL-4Rα−/lox = 7.28; Figure 1C). The cellular specificity of IL-4Rα deletion was confirmed because B cells (CD19+), macrophages, and dendritic cells (DCs; Figure 1C) of LckcreIL-4Rα−/lox mice maintained expression of IL-4Rα. Efficiency of deletion of IL-4Rα in CD4+ T cells was analyzed at the genomic level by quantitative real-time PCR. The number of exon 5 alleles (both present in all cells) relative to exon 8 alleles (deleted in −/−, one copy in −/lox mice) of IL-4Rα was determined in CD4+ T cells sorted to high purity. As expected, exon 8 was efficiently deleted in CD4+ T cells and B cells from IL-4Rα−/− mice (Figure 1D). Confirming FACS analysis, efficient deletion of lox-p–flanked IL-4Rα exon 8 was observed in CD4+ T cells from LckcreIL-4Rα−/lox mice. Analysis revealed 0.114 copies of exon 8 were retained relative to exon 5, equating to 95.48% ± 5.8% deletion efficiency of exon 8 within the CD4+ T cell population. In agreement, no CD4+ T cell exon 8 product was visible following 75 PCR cycles (Figure 1D). An equivalent ratio of exon 8 versus exon 5 was maintained in CD19+ B cells in LckcreIL-4Rα−/lox mice compared with WT controls. These data provide evidence of efficient deletion of IL-4Rα in CD4+ T cells from LckcreIL-4Rα−/lox BALB/c mice.
IL-4 promotes proliferation of naive CD4+ T cells in vitro [29]. In order to assess functional impairment of IL-4Rα on CD4+ T cells from LckcreIL-4Rα−/lox mice, naive CD4+ T cells were stimulated with IL-4, and proliferation was measured by [3H] thymidine incorporation (Figure 2A). CD4+ T cells isolated from naive LckcreIL-4Rα−/lox BALB/c mice were unable to proliferate in response to IL-4, as were those from IL-4Rα−/− mice. In contrast, WT CD4+ T cells showed dose-responsive proliferative responses to IL-4. Impairment of IL-4 signaling was IL-4Rα specific, as proliferative responses to IL-2, which shares a γc-chain with the type I IL-4R, were unaffected in all three strains (Figure 2A). Impairment of CD4+ T cell IL-4 responsiveness was further verified using the Th cell differentiation assay. Th1 versus Th2 differentiation of noncommitted CD4+ T cells may be achieved in vitro by treatment with either IL-12/anti–IL-4 or IL-4/anti–IFN-γ, respectively [29]. Naive CD4+ T cells stimulated with anti-CD3/CD28 and polarized with cytokine/neutralizing mAb treatment demonstrate that Th2 polarization, indicated by IL-4 production, was impaired in LckCreIL-4Rα−/lox and IL-4Rα−/− but not WT mice (Figure 2B). As expected, Th1 polarization was achieved in all three strains.
Functional macrophage IL-4Rα data from LckcreIL-4Rα−/lox mice were demonstrated in Figure 2C. NO production was suppressed by IL-4 and IL-13 in macrophages from LckcreIL-4Rα−/lox and WT mice (Figure 2C), but not IL-4Rα−/− macrophages, showing IL-4Rα specificity. As a positive control, IL-10 suppressed NO production in all three strains. Production of IgE antibodies is strictly dependent on IL-4 signaling [30]. IL-4Rα responsiveness of B cells in LckcreIL-4Rα−/lox mice was demonstrated in Figure 2D. Antigen-induced IgE antibody was present at slightly reduced levels in OVA-immunized LckcreIL-4Rα−/lox mice when compared with those of WT mice, while IgE production was completely abrogated in IL-4Rα−/− mice (Figure 2D). Together, these data provide evidence for effective impairment of IL-4Rα–mediated functions in LckcreIL-4Ra−/lox CD4+ T cells, but not in other lymphocyte subpopulations such as B cells and macrophages.
Controversy remains as to whether IL-4 [6,20,21] and/or IL-4Rα signaling [20,31] are key components of susceptibility to L. major infection. Polarized Th2 cells certainly play a significant role in contributing to susceptibility [32]. To investigate the consequence of CD4+ T cell–specific IL-4Rα unresponsiveness in leishmaniasis, mice were infected subcutaneously with 2 × 106 stationary phase metacyclic promastigotes of L. major LV39 (MRHO/SV/59/P; Figure 3A). As expected, WT mice developed rapidly growing nonhealing lesions (Figure 3A) within 12 wks of infection and were unable to control parasite burden with high parasite numbers in the footpads (Figure 3B) and LNs (Figure 3C). IL-4Rα−/− mice initially controlled infection with intermediate parasite load in the draining lymph nodes (LNs) and footpad. However, as previously described [20], global IL-4Rα deficiency does not confer resistance to L. major infection, as the mice progressed to develop necrotic lesions in the chronic phase (Figure 3A). In contrast, LckcreIL-4Rα−/lox mice were able to resolve infection with lesion growth comparable with resistant C57BL/6 mice (Figure 3A). LckcreIL-4Rα−/lox mice carried low parasite burdens in the footpad, with approximately 2,000-fold less parasites in the footpad compared with that of WT 6 wk after infection (Figure 3B), and maintained an intermediate parasite burden in the draining LNs when compared with C57BL/6 and WT mice (Figure 3C). Resistance to L. major infection in CD4+ T cell–specific IL-4Rα–deficient mice was profound, as parasite load in the footpad was equivalent to that observed in C57BL/6 mice at 36 wk after infection using PCR to detect kinetoplast DNA at the lesion site (Figure 3D). LckcreIL-4Rα−/lox mice were also shown to be resistant to reinfection. At 6 wk after L. major infection, mice were reinfected in the contralateral footpad. LckcreIL-4Rα−/lox mice were again comparable with genetically resistant C57BL/6 mice in lesion development, while L. major reinfection in WT mice progressed to necrosis in acute phase (Figure 3E). LckcreIL-4Rα−/lox mice were also resistant to the more virulent L. major (MHOM/IL/81/FEBNI) strain (Figure 3F), again with lesion kinetics comparable with that of C57BL/6 mice.
IL-10 is a highly immunosuppressive cytokine, profoundly reducing NO production by macrophages (Figure 2C) [33], and is a susceptibility factor in L. major infection [31]. Intracellular cytokine staining revealed increased numbers of antigen-specific CD4+ IL-10–secreting T cells in the draining LNs of WT mice compared with C57BL/6 and LckcreIL-4Rα−/lox mice (Figure 4A and 4B). In order to examine an in vivo correlate demonstrating IL-10 inhibition of protective parasite specific responses, IL-12/IFN-γ–driven delayed type hypersensitivity (DTH) responses were investigated in L. major–infected mice. C57BL/6 develop sustained footpad swelling when challenged with soluble L. major antigen (SLA; Figure 4C), and LckcreIL-4Rα−/lox mice showed intermediate sustained swelling, whereas minimal DTH responses were observed in WT mice (Figure 4C). As expected, addition of IL-10 to SLA diminished DTH responses in all mice (Figure 4D). Neutralization of IL-10 function by blockade of IL-10R lifted suppression of the DTH in the low-responder WT mice on a par with DTH responses observed in the resistant strains (Figure 4E). Confirming that increased DTH responses observed in LckcreIL-4Rα−/lox mice resulted from increased Th1 responses, significant levels of IL-12p70 (Figure 4F) and IFN-γ (Figure 4G) were detected in footpad lysates taken from resistant mice, while little or no IL-12p70 or IFN-γ were induced in susceptible WT mice (Figure 4F and 4G).
IL-12 is a key protective cytokine involved in inducing protective responses following L. major infection [34]. We therefore examined IL-12 expression in LckcreIL-4Rα−/lox mice. Although IL-12p35 mRNA production was equivalent at 1 wk after infection (unpublished data), levels of IL-12p35 mRNA were increased in draining LNs of LckcreIL-4Rα−/lox and C57BL/6 mice at 3 wk after infection when compared with those of WT mice (Figure 5A). Levels of IL-12p35 mRNA increased from 1 wk to 3 wk after infection in resistant mice while remaining low in susceptible mice (Figure 5B). IFN-γ–driven iNOS production by macrophages is a key control mechanism in L. major infection [35]. CD4 T cell antigen–specific IFN-γ cytokine production was therefore examined. CD4 T cells from LckcreIL-4Rα−/lox mice induced 2.5-, 1.6-, and 2-fold more IFN-γ when compared with those from IL-4Rα−/− and WT or IL-4Rα−/lox mice at 10, 6, and 12 wk after infection (Figure 5C), respectively. Furthermore, greater IFN-γ levels were detected in footpad homogenates from infected LckcreIL-4Rα−/lox compared with WT mice at 10 wk after infection (Figure 5D). Importantly, IL-4Rα–independent IL-4 production was observed in LckcreIL-4Rα−/lox mice with similar levels of IL-4 production being observed in WT and LckcreIL-4Rα−/lox mice in antigen-specific CD4+ T cell restimulation (Figure 5E) and footpad lysates (Figure 5F). Consistently increased IFN-γ production had an influence on downstream macrophage effector functions. This was shown at 6 wk after infection, when more copies of iNOS mRNA/parasite were observed in resistant strains of mice (Figure 5G). Together, these data demonstrate that resistance to acute leishmaniasis in LckcreIL-4Rα−/lox mice is associated with an early induction of increased protective type 1 immunity and reduced suppression of responses by IL-10–secreting CD4+ T cells.
IL-4 and IL-13 share a common signaling pathway through the IL-4Rα chain [26], and as such the combined role of both cytokines can be studied in vivo in IL-4Rα−/− mice. While IL-4 mediates multiple effects on T cells, murine T and B cells do not respond to IL-13 [7]. Generation of CD4+ T cell–specific IL-4Rα–deficient (LckcreIL-4Rα−/lox) mice therefore allows investigation into the role of IL-4 signaling specifically on CD4+ T cells while maintaining IL-4/IL-13–mediated functions on non-CD4+ T cells. CD4+ T cell–specific IL-4Rα–deficient BALB/c mice were generated using the Cre/LoxP recombination system in BALB/c embryonic stem cells. Previous studies have shown efficiency of cell-specific Cre-mediated gene disruption may vary between 38%–85% depending on recombinase efficiency and promoter activity [36]. Efficiency of CD4+ T cell–specific IL-4Rα disruption (95.48%) was increased by using hemizygous WT mice instead of IL-4Rαlox/lox as mating partners for transgenic LckCre mice, thereby reducing the LoxP substrate for Cre-recombinase by 50%. FACS analysis showed efficient disruption of IL-4Rα gene expression in CD4+ T cells and incomplete deletion in CD8+ and NK–T cells with variable deletion efficiency. γδ T cells and non–T cells retained unaltered receptor expression in LckcreIL-4Rα−/lox mice. The data suggest that while the Lck promoter is functional and mediates deletion of loxP-flanked DNA sequences in CD4+, CD8+, and NK–T cell subsets, deletion is more efficient in CD4+ T cells using this promoter construct. Functional analysis further demonstrated effective and specific impairment of the IL-4 responsiveness of CD4 T cells, while B cells and macrophages retained IL-4– and IL-13–mediated functions. Thus, LckcreIL-4Rα−/lox mice are CD4+ T cell–specific IL-4Rα knockout mice, whereas all other cell types remain responsive to IL-4/IL-13.
LckcreIL-4Rα−/lox mice infected with L. major developed similar kinetics of lesion development and resolution as those observed in C57BL/6 mice genetically resistant to two strains of L. major. In contrast, control IL-4Rα−/lox (WT) and IL-4Rα−/− BALB/c mice developed progressive lesion swelling leading to necrosis during the acute and chronic phases of disease as expected. LckcreIL-4Rα−/lox BALB/c and C57BL/6 mice also resisted secondary parasite challenge, unlike WT mice, which showed no signs of footpad pathology. A similar resistant phenotype to L. major infection was also noted in an independent line of mice in which IL-4Rα is efficiently deleted from CD4, CD8, NK–T, and γδ T cells (unpublished data), indicating that IL-4–responsive CD4+ T cells control susceptibility to L. major infection, and that the resistant phenotype is not associated with Cre activity in T cells or hypothetical mutations introduced by the transgene. Together, our study demonstrates that clinical immunity can be achieved in mice on a susceptible BALB/c background by abrogating IL-4Rα responsiveness on CD4+ T cells while retaining IL-4/IL-13–mediated function on non-CD4+ T cells.
IL-10 is a potent suppressor of macrophage activation [37], can abolish IFN-γ/LPS–induced killing of L. major by macrophages [38,39], and can suppress development of DTH responses [40]. In agreement, L. major–infected C57BL/6 and LckcreIL-4Rα−/lox mice developed DTH responses to SLA, inhibited by coadministration of IL-10. In contrast, DTH responses in WT mice were absent. Neutralization of IL-10 signaling allowed WT mice to mount a significant response to SLA. Together, DTH data demonstrated that IL-10 produced in response to SLA in susceptible mice was able to suppress protective cell-mediated immune responses.
IL-10 production is increased in BALB/c mice compared with resistant mice [41], can regulate parasite survival in resistant C57BL/6 mice [1,42], and is a susceptibility factor for L. major infection [31,39]. In agreement, the draining LNs of infected resistant LckcreIL-4Rα−/lox and C57BL/6 mice contained reduced numbers of CD4+ IL-10–secreting cells (4- and 9-fold less, respectively) compared with WT mice. Variable amounts of IL-10 staining were observed in the non-CD4+ T cell population; however, this was found to be nonspecific (Figure 4A). Increased IL-10 secretion was also observed in anti-CD3–stimulated CD4+ T cells derived from WT mice compared with T cells derived from LckcreIL-4Rα−/lox and C57BL/6 mice (not shown). IL-10 production by macrophages [43] and CD4+ T cells [31] has been linked to susceptibility to L. major infection. Using our assay system, IL-10–secreting cells were identified as CD4+ T cells. IL-10–producing CD4+ T cells have been implicated in controlling L. major parasite survival/infection in genetically resistant C57BL/6 mice. CD4+CD25+FoxP3+ IL-10–producing natural T regulatory cells (Tregs) have been elegantly shown to control parasite survival [44,45]. More recently, a novel disease controlling FoxP3− IL-10/IFN-γ–coproducing Th1 cell population has been identified [46]. The role for Tregs in control of L. major is unclear in BALB/c mice and potentially obscured by the predominant polarized Th2 response. The moderately specific method of Treg depletion using anti-CD25 antibody has produced contradictory results either enhancing [47] or reducing [48] susceptibility to L. major infection. Certainly, IL-4 has the ability to enhance the proliferation and function of CD4+CD25+ T cells in BALB/c mice [49,50]. However, the generation of CD4+FoxP3+ T cells was unaffected by IL-4Rα deficiency (unpublished data). Therefore, while not excluding a role for macrophage IL-10 production [43], our data suggest that IL-10 is predominantly produced by activated/effector T cells or Tregs, and further characterization of the CD4+IL-10+ T cells is ongoing.
The absence of IL-4Rα specifically on CD4+ T cells resulted in consistently higher levels of IFN-γ secretion by CD4+ T cells compared with WT mice. However, as previously shown, induction of increased IFN-γ responses alone does not guarantee control of L. major infection. Substantially increased L. major–specific CD4+ T cell IFN-γ production was observed in macrophage/neutrophil-specific IL-4Rα–deficient mice when compared with WT controls. However, infection also induced a potent polarized Th2 response, and lesion development was delayed but uncontrolled [9]. In contrast, in the absence of a polarized Th2 response, increased IFN-γ production correlated with protection against infection in LckcreIL-4Rα−/lox and C57BL/6 mice. Significant DTH responses upon injection of SLA into the footpad were observed as early as 3 wk after infection in LckcreIL-4Rα−/lox and C57BL/6 mice, but not in WT mice (unpublished data). Sustained tuberculin-like DTH responses are driven by IL-12–induced IFN-γ–producing Th1 cells [34,51], resulting in macrophage recruitment and activation, and are indicative of protective cell-mediated immune responses against intracellular pathogens. This was confirmed by increased IL-12 protein detected in tissue lysate of footpads of resistant mice compared with WT mice 24 h after DTH induction. Furthermore, increased levels of IFN-γ secretion were associated with increased expression of iNOS mRNA/parasite in infected footpads. Together, these results demonstrate that in the absence of IL-4Rα signaling on CD4 T cells, a polarized Th2 response, and IL-10 production, protective Th1 immune responses during cutaneous leishmaniasis result in effective macrophage activation and intracellular parasite elimination.
IL-4Rα−/− mice are susceptible to L. major infection in the acute [31] or the chronic [20] phase. Despite the absence of Th1 downregulatory signals through the IL-4Rα, IL-4Rα−/− mice do not produce increased amounts of IFN-γ following L. major infection when compared with WT controls [7]. Resistance to L. major in LckcreIL-4Rα−/lox mice has therefore revealed the protective role of IL-4/IL-13–responsive non-CD4+ T cells in control of infection in BALB/c mice. Crucial to resistance in LckcreIL-4Rα−/lox mice is CD4+ T cell IL-4Rα–independent IL-4 production. Not only induced following L. major infection [7,31] in IL-4Rα−/− mice, IL-4Rα–independent IL-4 production has been observed in response to Nippostrongylus brasiliensis [52] and Schistosoma mansoni [53] infections and following immunization with protein precipitated in alum [54]. As our study suggests, IL-4Rα–independent IL-4 production in LckcreIL-4Rα−/lox mice drives the induction of protective responses by non-CD4+ T cells. Both IL-4 and IL-13 are able to indirectly promote protective Th1 responses. Elegant experiments have demonstrated that IL-4 is able to instruct DCs to produce IL-12 and subsequent protection from L. major infection in BALB/c mice [55]. Furthermore, IL-4 is required for protective type 1 responses to Candida [56]. IL-13 can prime monocytes for IL-12 production [57] and drive protective cell-mediated immune responses during listeriosis [58]. Indeed, levels of IL-12p35 mRNA were increased in draining LNs of LckcreIL-4Rα−/lox and C57BL/6 mice by 3 wk after infection (Figure 5A), coincident with increased DTH responses (unpublished data). As macrophage IL-12 production is actively downregulated by L. major [18], it is likely that increased IL-12p35 mRNA levels in the LNs at 3 wk after infection were produced by DCs. In agreement, infected DCs appear in draining LNs in two waves; the first transient wave peaks at 24 h, and the second commences 15–21 d after L. major infection [59]. Therefore, IL-4Rα–independent IL-4 production and subsequent IL-12 production by DCs in the absence of Th2 polarization may explain the protection of LckcreIL-4Rα−/lox from L. major infection. Furthermore, the protective effect of IL-4 signaling in non-CD4+ T cells may also explain the requirement for IL-4 in effective treatments against visceral leishmaniasis [60,61].
In summary, in the absence of a polarized Th2 response where non-CD4+ T cells retain IL-4/IL-13 responsiveness, increased protective immune responses are induced by 3 wk in LckcreIL-4Rα−/lox mice. As IL-12 may also negate Treg cell action on activated T cells [62], this regulation is likely to enhance beneficial Th1 responses and immunity following L. major infection in LckcreIL-4Rα−/lox mice, possibly reflecting a similar scenario in the healer C57Bl/6. In contrast, IL-4Rα expression on CD4+ T cells allows Th2 polarization and induction of IL-10 production in the nonhealer BALB/c strain. As a consequence, Th1 responses and protective macrophage effector functions are downregulated, IL-10 is upregulated, and subsequently, BALB/c mice succumb to L. major infection in the acute phase. In conclusion, where CD4+ T cells are unable to respond to IL-4, IL-4/IL-13 signaling in non-CD4+ T cells is beneficial in BALB/c mice following infection with L. major.
Transgenic Lckcre mice [28] back-crossed to BALB/c for nine generations were intercrossed with IL-4Rα−/− and IL-4Rαlox/lox mice to generate LckcreIL-4Rα−/lox BALB/c mice. WT littermates were used as controls in all experiments. Mice were genotyped as described previously [24]. All mice were housed in specific pathogen–free barrier conditions at the University of Cape Town, South Africa, and used in accordance with University ethical committee guidelines. All experimental mice were age and sex matched and used between 8–12 wk of age.
DNA was prepared from CD3+CD4+ and CD19+ sorted LN cells from LckcreIL-4Rα−/lox, WT, or IL-4Rα−/− mice using a FACsvantage flow cytometer (BD, http://www.bd.com) to >99% purity. A standard curve was prepared from serial 10-fold DNA dilutions of cloned IL-4Rα exon 5 and exon 8 DNA. Primers: exon 5 forward 5′ AACCTGGGAAGTTGTG 3′, exon 5 reverse 5′ CACAGTTCCATCTGGTAT 3′; exon 8 forward 5′ GTACAGCGCACATTGTTTTT 3′, exon 8 reverse 5′ CTCGGCGCACTGACCCATCT 3′.
DNA was prepared from homogenized tissues samples. A DNA standard curve was prepared from serial 10-fold parasite DNA dilutions in PBS. L. major kinetoplast primers used: forward 5′ CGCCTCCGAGCCCAAAAATG 3′ and reverse 5′ GATTATGGGTGGGCGTTCTG 3′. Real-time PCR amplification and data analysis performed using the “Fit Points” and “Standard Curve” methods as described previously [63].
IL-4Rα was detected by anti-IL-4Rα–PE (M-1; BD), and leukocyte subpopulations were identified using anti-CD19 (1D3), anti–δ-TCR (GL3), anti-CD11c (HL3), anti-F4/80, anti–I-Ad (AMS-32.1), anti-CD11b (M1/70) (all from BD), anti-CD3 (145–2C11), anti-CD4 (GK1.5), and anti-CD8 (53.6.72) mAbs, which were purified from hybridoma supernatants by protein G sepharose (Amersham Biosciences, http://www.amersham.com) and labeled with FITC or biotin. Biotin-labeled antibodies were detected by streptavidin–allophycocyanin (BD). Dead cells were stained by 7-AAD and excluded from analysis (Sigma, http://www.sigmaaldrich.com). Acquisition was performed using FACSCalibur, and data were analyzed by Cellquest (BD).
CD4+ T cells, positively selected by anti-CD4 Dynabeads (Invitrogen, http://www.invitrogen.com) to a purity of >85% as described [7], were stimulated with serial dilutions of IL-4, IL-13, or IL-2 (BD) in complete IMDM containing 10% FCS, penicillin, and streptomycin, 1 mM sodium pyruvate, NEAA (Invitrogen), 10 mM HEPES, and 50 μM β2-ME (Sigma). After 48 h of incubation at 37 °C and 5% CO2, cells were pulsed with 1 μCi (0.037 MBq) [3H] thymidine (Amersham Biosciences) for a further 18 h. [3H] incorporation was measured in a liquid scintillation counter.
In vitro Th1/Th2 differentiation of purified CD4+ T cells was induced as described previously [7].
Suppression assay was performed as described [20]. Briefly, adherent macrophages derived from peritoneal exudate cells elicited with 3% Brewers thioglycollate (Difco Laboratories, http://www.bd.com/ds) were incubated for 16 h with medium or with IL-4, IL-13, or IL-10 at 1,000 U/ml (R&D Systems, http://www.rndsystems.com). Cells were subsequently stimulated with LPS (15 ng/ml; Sigma) and IFN-γ (100 U/ml; BD) and NO was measured by Griess reaction after 48 h.
Mice were immunized subcutaneously with 10 μg of OVA in CFA (Sigma) and boosted at 7 and 14 d with OVA intraperitoneally. IgE production was detected as described previously [20].
L. major LV39 (MRHO/SV/59/P) and MHOM/IL/81/FEBNI strains were maintained by continuous passage in BALB/c mice and cultured in vitro as described previously [20]. Mice were inoculated subcutaneously with 2 × 106 stationary phase metacyclic promastigotes into the left hind footpad in a volume of 50 μl HBSS (Invitrogen). Swelling was monitored every week up to a maximum of 40 wk using a Mitutoyo pocket thickness gauge (http://www.mitutoyo.com). For reinfection studies, 6 wk after primary infection, mice were injected subcutaneously with 2 × 106 stationary phase metacyclic promastigotes into the contralateral footpad. Footpad swelling was monitored for 18 wk.
Infected organ cell suspensions were cultured in Schneider's culture medium (Sigma). Parasite burden was estimated according to a previously described limiting dilution method [20].
Total RNA from footpad or LN was purified using mini-elute columns (Qiagen, http://www.qiagen.com) and cDNA was generated using the Inprom-II re-verse transcription system (Promega, http://www.promega.com). Primers pairs used to detect IL-12p35 message: forward 5′-GATGACATGGTGAAGACGGCC-3′, and reverse 5′-GGAGGTTTCTGGCGCAGAGT-3′. iNOS message forward 5′-AGCTCCTCCCAGGACCACAC-3′, and reverse 5′-ACGCTGAGTAC CTCATTGGC-3′. Data analysis was performed using the “Fit Points” and “Standard Curve” methods using beta-2-microglobulin as a housekeeping gene.
Mice were inoculated subcutaneously with 10 μg SLA into the right hind footpad alone or with 0.5 μg mouse rIL-10 or 1.5 μg anti–IL-10Rα (R&D Systems). Footpad swelling was measured every 24 h. Footpads were homogenized, and lysates were taken 24 h after induction of DTH.
CD4+ T cells were positively selected using anti-CD4 Macs beads (Miltenyi Biotec, http://www.miltenyibiotec.com) to a purity of >90% according to the manufacturer's instructions. Thy1.2-labeled splenocytes were T cell depleted by complement-mediated lysis (Cedarlane, http://www.cedarlanelabs.com) to produce antigen-presenting cells (APCs). APCs fixed with mitomycin C (50 μg/ml, 20 min at 37 °C) and washed extensively in complete IMDM. A total of 2 × 105 purified CD4+ T cells and 1 × 105 APCs were cultured with SLA at 50 μg/ml, supernatants were collected after 48 h, and cytokines were analyzed as previously described [20].
IFN-γ and IL-4 were detected in footpad tissues using the method previously described [24].
L. major–infected mice; popliteal LN cells at 2 × 105 cells/well were stimulated with SLA (5 μg/ml) for 24 h. Cultures were supplemented with monensin (2 μM) for the final 4 h of culture. Cells were stained with anti-CD4 FITC (mAb, GK1.5), fixed, permeabilized, and stained with anti–IL-10 APCs (BD).
Values are given as mean ± SD and significant differences were determined using Student's t test (Prism software, http://www.prism-software.com). |
10.1371/journal.pntd.0006381 | Development and validation of four one-step real-time RT-LAMP assays for specific detection of each dengue virus serotype | 4 one-step, real-time, reverse transcription loop-mediated isothermal amplification (RT-LAMP) assays were developed for the detection of dengue virus (DENV) serotypes by considering 2,056 full genome DENV sequences. DENV1 and DENV2 RT-LAMP assays were validated with 31 blood and 11 serum samples from Tanzania, Senegal, Sudan and Mauritania. DENV3 and DENV4 RT-LAMP assays were validated with 25 serum samples from Cambodia
4 final reaction primer mixes were obtained by using a combination of Principal Component Analysis of the full DENV genome sequences, and LAMP primer design based on sequence alignments using the LAVA software. These mixes contained 14 (DENV1), 12 (DENV2), 8 (DENV3) and 3 (DENV4) LAMP primer sets. The assays were evaluated with an External Quality Assessment panel from Quality Control for Molecular Diagnostics. The assays were serotype-specific and did not cross-detect with other flaviviruses. The limits of detection, with 95% probability, were 22 (DENV1), 542 (DENV2), 197 (DENV3) and 641 (DENV4) RNA molecules, and 100% reproducibility in the assays was obtained with up to 102 (DENV1) and 103 RNA molecules (DENV2, DENV3 and DENV4). Validation of the DENV2 assay with blood samples from Tanzania resulted in 23 samples detected by RT-LAMP, demonstrating that the assay is 100% specific and 95.8% sensitive (positive predictive value of 100% and a negative predictive value of 85.7%). All serum samples from Senegal, Sudan and Mauritania were detected and 3 untyped as DENV1. The sensitivity of RT-LAMP for DENV4 samples from Cambodia did not quite match qRT-PCR.
We have shown a novel approach to design LAMP primers that makes use of fast growing sequence databases. The DENV1 and DENV2 assays were validated with viral RNA extracted clinical samples, showing very good performance parameters.
| The co-existence of several dengue virus (DENV) serotypes within the same location and/or individuals as well as a single mosquito being able to carry multiple DENV serotypes highlight the necessity of specific diagnostic tools capable of detect and serotype DENV strains circulating worldwide. In addition, these methodologies must be highly sensitive in order to detect the genome at low levels (i.e., before the onset of clinical symptoms) and not cross-detect other flaviviruses. Isothermal amplification methods (such as reverse transcription loop-mediated isothermal amplification, RT-LAMP) are affordable for laboratories with limited resources and do not need expensive equipment. Because of the increasing number of publicly available full DENV genome sequences, traditional primer design tools are not able to handle such huge amount of information. Therefore, to be able to cover all the diversity documented, we developed 4 complicated oligonucleotide mixes for the individual detection of DENV1-4 serotypes by RT-LAMP. This approach combined Principal Component Analysis, phylogenetic analysis and LAVA algorithm. Our assays are specific and do not cross-react with other arboviruses and DNA pathogens included in this study, they are sensitive and have been validated with samples from Tanzania, Senegal, Sudan, Mauritania and Cambodia, showing very good performance parameters.
| Dengue is a worldwide public health concern annually affecting more than 100 million people in tropical and subtropical areas [1, 2]. It is caused by dengue virus (DENV), the most common vector-borne viral pathogen of humans, transmitted by mosquitoes of the Aedes genus (primarily A. aegypti and to a lesser extent A. albopictus), as previously reviewed [3]. DENV infection in humans results in a broad spectrum of disease manifestations, ranging from self-limiting, acute febrile illness (dengue fever) to more severe forms of the disease (dengue haemorrhagic fever and dengue shock syndrome), which may lead to death [4]. In 2013, the annual global incidence was estimated close to 390 million DENV infections, which was more than three times the dengue burden estimate of the World Health Organization [2].
DENV is an enveloped virus (genus Flavivirus, family Flaviviridae) with a genome that consists of a single-stranded, positive-sense RNA molecule of about 11 kb in length. The DENV genome encodes three structural proteins (C, capsid; prM, pre-membrane, and E, envelope) at the N terminus and seven non-structural (NS) proteins (NS1, NS2a, NS2b, NS3, NS4a, NS4b and NS5) [5, 6]. This virus is classified into four phylogenetically related and loosely antigenically distinct serotypes (DENV1, DENV2, DENV3 and DENV4), each of which contains phylogenetically different genotypes [7–9].
DENV outbreaks between 2006 and 2013, in India, Vietnam, Solomon Islands, Myanmar, China, Singapore, Malaysia and Portugal [10–14], highlight the necessity of rapid virus detection to identify DENV as the cause of an outbreak, in order to manage and control virus spread in infrastructure poor urban, peri-urban and rural settings. Notably, routine detection of DENV in children who are often asymptomatic carriers could improve outbreak control [15]. A first vaccine has recently been licensed for the prevention of dengue, which aims to reduce the number of hospitalizations per year, being approved for people aged between 9 to 45 years [16].
Traditional virus isolation is time-consuming, requires experienced staff, costly facilities and equipment and needs more than seven days to complete the assay [17, 18]. IgM- and IgG-capture enzyme-linked immunosorbent assay (ELISA) are most widely used but some degree of cross-reactivity against other flaviviruses is usually observed and this method is not useful when antibody titers are not sufficiently high (febrile viremic phase) [19]. Molecular amplification techniques to detect DENV RNA (RT-PCR, quantitative RT-PCR—qRT-PCR), which have emerged as a new standard, have a quick turnaround time and can distinguish DENV serotypes [20–26]. However, these techniques require sophisticated equipment and experienced staff, making them unpractical for laboratories with limited resources.
Loop-mediated isothermal amplification (LAMP) has the potential to substitute PCR-based methods because of its simplicity, rapidity, specificity, sensitivity and cost-effectiveness, as no special equipment is needed (just a heating block or water bath capable to maintain a constant temperature between 60°C to 65°C) [27–29]. Reactions can be visualised by monitoring either the turbidity in a photometer or the fluorescence in a fluorimeter, by visual inspection under UV lamp when using an intercalating dye or by colour change [8, 28–36].
Previously reported reverse transcription LAMP (RT-LAMP) assays for DENV target the 3’ untranslated region (UTR) [8, 30, 32, 34, 37], whilst other detect a fragment of the C-prM region [33], a conserved region of the NS1 [36], or regions of NS2A (DENV1), NS4A (DENV3), NS4A (DENV2) and the 3’ UTR (DENV4) [38]. In all cases information about the primer design is limited as only one sequence per serotype or reference sequences were considered or it is not clearly detailed how the sequence alignment was carried out or how many sequences were included in the design. An initial screen of all published DENV RT-LAMP detection amplicons quickly revealed that all of them fail to cover the documented sequence variation. To improve DENV RT-LAMP design we used the LAMP Assay Versatile Analysis (LAVA) algorithm [39] which solves the limitations of existing LAMP primer-designing programs by allowing designs based on large multiple sequence alignments. Our LAMP design is based on 2,056 whole-genome DENV sequences covering DENV strains from 2004 to 2014 and yielded 4 one-step, real-time RT-LAMP assays to specifically detect each DENV serotype.
Ethical approval for retrospective use of the anonymized samples in diagnostic development research was available: Tanzania samples (Ethikkommission Basel in Switzerland, Institutional Review Board of the Ifakara Health Institute and National Institute for Medical Research Review Board in Tanzania), IPD and IPC samples (Ministry of Health of Senegal and National Ethics Committee for Health Research of Cambodia, respectively).
Virus material: DENV isolates were provided and tested at the Institut Pasteur in Paris (Table 1). TriReagent extracts from flavivirus culture supernatants were provided by M. Weidmann. Inactivated strains ATCC VR-344 (DENV1), ATCC VR-345 (DENV2), ATCC VR-1256 (DENV3) and ATCC-1257 (DENV4) were provided by ENIVD / Robert Koch Institute. An inactivated Zika virus strain (ZIKV, H/PF/2013) was provided by Prof. Xavier de Lamballerie (Unité des Virus Emergents, Marseille, France).
An External Quality Assessment (EQA) 2015 panel was provided by QCMD (Quality Control for Molecular Diagnostics, Glasgow, UK) including ten unknown samples (15–01 to 15–10).
Patient samples: We used RNA extracts of 31 blood samples collected during a fever study in Tanzania, 2013 (Table 2) provided by the Swiss Tropical and Public Health Institute in Basel, Switzerland. These samples included 24 DENV qRT-PCR positive, 2 DENV positive (not characterized by qRT-PCR) and 5 negative samples. In addition, a negative sample from MAST Diagnostica GmbH (Reinfeld, Germany) was included. RNA extracts of 11 DENV qRT-PCR serum samples from Senegal, Sudan and Mauritania collected in November-December 2014 by the Institut Pasteur in Dakar (IPD), Senegal (Table 3) were tested by qRT-PCR and LAMP in Dakar. Additionally serum samples from Cambodia collected through the National Dengue Surveillance System [40] were tested. RNA was extracted and air-dried using pre-dried RNAstable 1.5 mL microfuge tubes (Biomatrica, USA) from 13 DENV3 and 12 DENV4 samples, collected by the Institut Pasteur du Cambodge (IPC) in 2004–2006 and between 2008 and 2014, respectively. Samples were shipped at ambient temperature. Moreover, samples were tested by qRT-PCR before shipment and after receipt and reconstitution in molecular grade water. Overall the qRT-PCR CT deviation was in a range of 0.8 CT. Five μL RNA of each sample were used per reaction.
RNA extractions were carried out using the RNeasy mini (DENV strains from Robert Koch Institute, QCMD samples) (QIAGEN, Crawley, West Sussex, UK) and the QIAamp Viral RNA mini (DENV samples from IPD and IPC and ZIKV strain from Unité des Virus Emergents) (QIAGEN, Courtaboeuf, France) kits. TriReagent extracts were processed according to the manufacturer’s extraction protocol (Sigma-Aldrich, Dorset, UK).
RNA extraction of the clinical samples from Tanzania was initially performed from 50 μL whole blood using a trial version of a nucleic acid isolation system equivalent to the protocol established for the MagSi-gDNA blood kit (MagnaMedics, Geleen, The Netherlands). RNA was eluted in 190 μL elution buffer, and 5 μL per sample were used for each RT-LAMP reaction. Additionally, an improved trial version of the MagnaMedics system for nucleic acid isolation, starting from 100 μL whole blood and eluting the RNA in 100 μL elution buffer, using 5 μL per sample for each RT-LAMP reaction, was used. RNA was extracted from the clinical samples from Senegal using the QIAamp Viral RNA mini kit.
A DENV RNA standard was transcribed from the DENV 3’ UTR region, ligated into pCRII and evaluated as previously described [41]. DEN FP and DEN P were as described with the probe now tagged 5’-FAM / BBQ-3’ but an adapted reverse primer DEN RP2 (5’-CTGHRGAGACAGCAGGATCTCTG-3’) as described [42]. DENV qRT-PCR was performed using the Light Cycler 480 Master Hydrolysis Probes (Roche, Mannheim, Germany) in a 20-μL reaction volume containing 1x LightCycler 480 RNA Master Hydrolysis Probes, 3.25 mM activator Mn(OAc)2, 500 nM primers DEN FP and DEN RP2, 200 nM probe DEN P, and 1 μL RNA template on the LightCycler 2.0 (Roche), as follows: reverse transcription for 3 min at 63°C, activation for 30 s at 95°C, followed by 45 cycles consisting of amplification for 5 s at 95°C and 15 s at 60°C and a final cooling step of 40 s at 40°C. Analysis of the reactions was conducted using LightCycler software version 4.1.1.21 (Roche).
The Institut Pasteur in Dakar performed a qRT-PCR [43], using the ABI7500 Fast Real-time PCR System (Applied Biosystems, Foster City, CA). An RT-PCR assay, which simultaneously detects the 4 DENV serotypes, followed by a nested PCR, that specifically detects each DENV serotype, were used [20].
A two-step approach was used. First, all available sequences of DENV1 to 4 were downloaded from the NCBI database. Searches were limited to the samples collected between 2004 and 2014. All sequences were then aligned (for each serotype) using GramAlign v3.0 [44], and diversity was assessed using the glPCA module of R/adegenet v1.4.1 [45]. Finally, based on the Principal Component Analysis (PCA) and phylogenetic tree (Neighbor-Joining tree using the R/ape 3.2 package), the sequences were manually split into different clusters in order to maximise the region of sequence identity. LAMP DNA signatures for each cluster (and all combinations to minimise the number of primer sets) were designed using a modified version [https://github.com/pseudogene/lava-dna] of LAVA [39] applying the loose parameters set for DENV1-3 and the standard parameter set for DENV4. Full scripts and methods are available on GitHub at https://github.com/pseudogene/lamp-denv.
All the designed sets of primers were first checked for primer dimerisation with the VisualOMP version 7.8.42.0 (DNA Software, Ann Arbor, MI). In addition, primer combinations for each of the DENV assays were tested for primer dimerisation by comparing amplification signals in positive and negative controls.
RT-LAMP reactions were run at 64°C using either an ESE-Quant TubeScanner (QIAGEN Lake Constance GmbH, Stockach, Germany) or Genie II (Optigene, Horsham, UK), in a final reaction volume of 25 μL. The Genie II device displays the annealing curve for specificity after the reaction has finished, by melting curve analysis from 98°C to 80°C (0.05°C/s).
Four RT-LAMP assays were developed, one for each DENV serotype (S1 File). Each reaction consisted of 1x RM Trehalose, 6 mM MgSO4, 5% polyethylene glycol (PEG), 1 μL fluorochrome dye (FD), 8 U Bst 2.0 DNA Polymerase (New England BioLabs, Hitchin, Herts, UK), 10 U Transcriptor Reverse Transcriptase (Roche) and 1 μL template (DENV RNA or H2O as negative control). For each primer set per RT-LAMP assay, the final concentrations was as follows: 50 nM F3, 50 nM B3, 400 nM FIP, 400 nM BIP, 200 nM FLOOP, 200 nM BLOOP. Before adding the Bst 2.0 DNA Polymerase, Transcriptor Reverse Transcriptase and template, mixes were incubated at 95°C for 5 min to melt any primer multi-mers and cooled immediately on ice for 5 min. Reaction times vary for each RT-LAMP protocol, running for 45 min (DENV1), 90 min (DENV2), 75 min (DENV3) and 50 min (DENV4).
RM Trehalose, MgSO4, PEG and FD were supplied by MAST Diagnostica GmbH.
Sensitivity analysis was performed in the ESE-Quant TubeScanner (QIAGEN). Ten-fold dilutions of viral DENV RNA samples (ATCC VR-344 (DENV1), ATCC VR-345 (DENV2), ATCC VR-1256 (DENV3) and ATCC VR-1257 (DENV4)), quantified by qRT-PCR, were used to analyse the sensitivity of the developed RT-LAMP assays (range from 104−105 to 10 molecules/μL) and 1 μL per dilution was added to the RT-LAMP reaction. The complete RNA standard was tested in eight separate runs. The values obtained were subjected to probit analysis (Statgraphics plus v5.1, Statistical Graphics Corp., Princeton, NJ) and the limit of detection at 95% probability for each RT-LAMP assay was obtained.
Cross-specificity tests for the four RT-LAMP assays were carried out at the Institut Pasteur (Paris) using the QuantStudio 12K Flex Real-Time PCR System, and results were analysed with the software QuantStudio 12K Flex v1.2.2. (Applied Biosystems, Carlsbad, CA). Each of the RT-LAMP assays was tested using 1 μL RNA extracted from the DENV strains described in Table 1. Cross detection of other flaviviruses, ZIKV, Yellow fever virus (YFV), West Nile virus (WNV) and Ntaya virus (NTAV), was analysed using the Genie II (Optigene) at the University of Stirling.
The RT-LAMP assays were also tested against several DNA pathogens (Salmonella Typhi, S. Paratyphi, Streptococcus pneumoniae and Plasmodium falciparum). DNA samples were provided by MAST Diagnostica GmbH.
The performance of the RT-LAMP assays (sensitivity and specificity) was additionally evaluated using the 2015 DENV EQA panel provided by QCMD. Results obtained from QCMD refer to 8 core and 2 educational samples. Core samples are those needed to assess the performance from the regulatory point of view and educational samples are additional samples related to questions such as limit of detection or specificity.
We used 31 blood samples from a fever study in Tanzania, 2013 (Table 2). Twenty-six samples had been confirmed as DENV2 positive by the Swiss Tropical and Public Health Institute (Basel, Switzerland) (2 of them were not tested by qRT-PCR). Aliquots of these blood samples were sent to MAST Diagnostica GmbH and stored at -20°C until RNA extraction was performed using the Magnamedics kit trial version. RNA samples were stored at -80°C. RT-LAMP reactions were run in the TubeScanner TS2 (QIAGEN), using 5 μL RNA of each sample per reaction.
The samples at IPD were analysed by both qRT-PCR [43], and the DENV1 and DENV2 RT-LAMP assays (in triplicates) in an ABI7500 Fast Real-time PCR system (Applied Biosystems), using 5 μL RNA of each sample per reaction.
Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were obtained for the DENV2 RT-LAMP developed when compared against the results obtained by qRT-PCR.
The RNA standard was tested 3 times and similar crossing point (CP) values were obtained for the different dilutions from 107 to 103 RNA molecules detected (S1 Fig), showing an efficiency (E = 10−1/slope—1) of 0.99 ± 0.04 (mean ± standard deviation, SD). Quantification of DENV1-4 RNA extracted from inactivated isolates ATCC VR-344 (DENV1), ATCC VR-345 (DENV2), ATCC VR-1256 (DENV3) and ATCC VR-1257 (DENV4) (Table 1) ranged from 6.9x104–9.4x104 (DENV1), 4x105–5.3x105 (DENV2), 1.5x105 - 3x105 (DENV3), and 1.8x105–2.7x105 (DENV4) RNA molecules/μL.
In total 1,145, 477, 376 and 58 genomic sequences were retrieved from the NCBI database for DENV1, DENV2, DENV3 and DENV4, respectively. Each serotype dataset was split into 4 to 21 clusters (Fig 1A and S2–S4 Figs), allowing for the LAVA algorithm to design LAMP primer sets, and was executed for each group separately as well as for all possible combinations of the groups.
Sets of primers that showed dimerisation when analysed with VisualOMP (DNA Software, Ann Arbor, MI) were discarded (Fig 2A). Remaining sets where sequentially combined and tested by RT-LAMP to discard cases of primer dimerisation, visualised by the non-specific amplification signal (intercalating dye) in the no template control (NTC) (Fig 2B). The final primer sets are described in Fig 1B and S1–S4 Tables and consist of 84 (14 amplicons, DENV1), 72 (12 amplicons, DENV2), 48 (8 amplicons, DENV3) and 18 (3 amplicons, DENV4) primers.
When combining the amplicon primer sets for each RT-LAMP assay, amplification was not observed when using published standard LAMP primer concentrations for each primer set: 0.2 μM F3, 0.2 μM B3, 1.6 μM FIP, 1.6 μM BIP, 0.8 μM FLOOP and 0.8 μM BLOOP. To determine the concentration window of the complicated primer mix, a 2-fold dilution series of the above primer mix was used. Amplification yielding the best possible detection without amplification in the NTC was achieved at a dilution of 1:4 (50 nM F3, 50 nM B3, 400 nM FIP, 400 nM BIP, 200 nM FLOOP and 200 nM BLOOP, Fig 2C).
Table 1 and Fig 3 show the results of the cross-specificity and cross-detection tests. All DENV cell culture RNA extracts were detected and no amplification was observed in the NTC. The RT-LAMP protocols for DENV2, DENV3 and DENV4 were specific for each respective serotype. The RT-LAMP protocol for DENV1 detected all DENV1 RNA strains, but also scored positive in RNA extracts KDH0010A and VIMFH4 containing RNA extracts from DENV3 and DENV4 isolates, respectively (Table 1). Additional testing of samples KDH0010A and VIMFH4 by nested RT-PCR (Fig 4A and 4B) indicated contamination of the cell cultures samples with DENV1 confirming the RT-LAMP results.
The RNA of other flaviviruses was not cross-detected (Fig 3 and Table 1). Specific amplification was also indicated by a specific single peak temperature in the melting curve analysis (Fig 3B, 3D, 3F and 3H), with mean values ± SD of 85.4 ± 1.1°C (DENV1), 83.1 ± 1.0°C (DENV2), 84.3 ± 0.9°C (DENV3) and 86.4 ± 0.3°C (DENV4). No amplification was observed when DNA from S. Typhi, S. Paratyphi, S. pneumoniae and P. falciparum was used as template in the different RT-LAMP assays (Table 1).
The 2015 DENV EQA panel analysis confirmed that the RT-LAMP assays developed passed 8 core and the 2 educational samples of that panel. Concerning the core samples, 5 positive samples were scored 3/3, and 1 positive sample was detected once (the other 2 samples were negative). Results obtained from the educational samples indicated that 1 sample was detected in the 3 repetitions whilst the other sample was detected in 1/3 repetitions.
DENV1-4 RNA samples, previously quantified by qRT-PCR, were used to analyse the sensitivity of the developed RT-LAMP assays. RT-LAMP protocols for DENV1, DENV2 and DENV4 detected as few as 10 molecules per reaction, although this amount was only obtained in 3, 5 and 2 of 8 repetitions, respectively, with the following mean times: 28.8 ± 6.3 min (DENV1), 78.2 ± 5.8 min (DENV2) and 44.6 ± 3.3 min (DENV4). RT-LAMP for DENV3 detected as few as 102 molecules, but only in 4 of 8 reactions, at 44.9 ± 18.6 min. The lowest amount of molecules detected in the 8 reactions, showing 100% reproducibility, were 102 (DENV1, mean time of 25.3 ± 2.6 min), and 103 (DENV2, DENV3 and DENV4, mean times of 69.2 ± 11.6 min, 37.2 ± 11.6 min and 26.8 ± 2.7 min, respectively) (Fig 5). Considering 8 independent reactions per protocol developed, the probit analysis revealed that the limit of detection at 95% probability for each RT-LAMP was 22 RNA molecules (DENV1), 542 RNA molecules with a confidence interval from 92 to 3.2x1013 RNA molecules (DENV2), 197 RNA molecules (DENV3) and 641 RNA molecules with a confidence interval from 172 to 1.2x105 RNA molecules (DENV4).
Tables 2 and 3 show the results of the blood and serum samples analyses when using both qRT-PCR and RT-LAMP.
Out of 26 DENV2-infected blood samples 24 scored positive in qRT-PCR with cycle threshold (CT) values ranging from 21.57–29.13 (Table 2, column 2). In a first test DENV2 RT-LAMP detected 17/24 (70.8% positive samples) with initial time to positive (TT) values between 37 and 89 min (Table 2, column 3).
RNA from 14 samples, including those with initial TT values over 60 min, negative in both RT-LAMP and qRT-PCR, and 6 DENV negative samples (Table 2), were extracted a second time using the optimized MagnaMedics extraction starting from 100 μL sample and yielding enhanced detection.
Five samples with initial TT values from 81–89 min, now tested positive with TT values from 55–77 min. Six samples initially negative by RT-LAMP became positive with TT values of 61.7–72.2 min. Three samples, 1 of which had scored positive in qRT-PCR, remained negative in RT-LAMP. Most RNA samples extracted with the optimized method scored positive in all 3 replicates. One sample was detected 2/3 times, and 2 were detected only once. All negative samples included in these analyses scored negative.
Calculation of the clinical sensitivity and specificity yielded 100% specificity (CI: 0.63–1.00), as no false positives were detected, and a sensitivity of 95.8% (CI: 0.79–1.00) with 23/24 positive samples, a PPV of 1.00 (CI: 0.85–1.00) and NPV of 0.86 (CI: 0.42–1.00).
Table 3 summarises the results obtained with samples collected by the IPD and IPC. All 11 RNA samples from IPD used in this study were analysed in parallel by qRT-PCR and with DENV1 and DENV2 RT-LAMP assays. All scored positive in qRT-PCR (CT 25.89–38.48), 4 samples scored positive in the DENV1 RT-LAMP, and 7 scored positive in the DENV2 RT- LAMP (TT values 20–45 min). Samples 267175, 267197 and 267174 were serotyped as DENV1 with the developed RT-LAMP.
Additionally, of 12 qRT-PCR positive DENV4 samples dried with RNAstable shipped by IPC, 10 tested positive by qRT-PCR after shipment, and 9 were detected by DENV4 LAMP. Of 13 DENV3 samples qRT-PCR positive before shipment, only 1 tested positive by qRT-PCR on arrival and only 3 by RT-LAMP.
Dengue is now prevalent in more than 100 countries of the tropics and subtropics and as DENV continues to spread, all four serotypes co-circulate widely [46–48]. The introduction of new DENV strains continues through travellers moving between dengue-endemic countries [11] and recently the capacity of individual mosquitoes to carry multiple DENV serotypes was described [49], while elsewhere acute simultaneous infection with several DENV serotypes was observed [10].
DENV detection methods include virus culture, which is time consuming [17, 18] as well as ELISA or immunofluorescence methods to detect IgM and IgG which suffer from cross-reactivity to other flaviviruses antibodies and which are only considered valid when antibody titers are sufficiently high [19]. The introduction of NS1 antigen detection has improved the situation and recent studies show a high sensitivity of NS1 detection [50], with some concluding that the combination with IgM detection can outperform PCR [51]. However, its use for routine screening in dengue epidemics is questioned in terms of clinical necessity [52].
For molecular RNA detection, nested PCR [20] and real time PCR-assays [21–26] with high specificity and sensitivity are being used but need expensive and sophisticated thermocyclers and experienced staff. In recent years, isothermal amplification assays have been described, such as RT-LAMP [8, 30, 32–38] and RT-RPA [53, 54]. These assays require less expensive equipment and can be delivered in dried pellet format, making handling easier and amenable to poor infrastructure settings.
Worldwide monitoring and the use of Next Generation Sequencing methods have increased the number of complete DENV genomes sequenced and deposited in GenBank to 2,988 (as of June 2016). It is virtually impossible to use this amount of sequence information to manually align and design amplicons for molecular detection methods. There have been several attempts to create algorithms to derive signature sequences for PCR techniques from sequence datasets or alignments [55, 56]. LAMP amplicons are inherently more challenging to design as they require a minimum of 4 and a maximum of 6 signature sequences. LAVA software was developed to facilitate the determination of signature sequences for LAMP primer design using a set of aligned sequences [39]. The original and modified version of LAVA take into consideration the limitations observed with other primer-design programs (LAMP DESIGNER [http://www.optigene.co.uk/lamp-designer/] and PRIMER EXPLORER [https://primerexplorer.jp/e/], such as preventing the use of extensive alignments or sequences longer than 2,000 nt.
We used this approach to design serotype-specific primers aiming to match all possible DENV strains circulating worldwide, by considering 2,056 available GenBank DENV sequences (2004–2014). This is the greatest difference compared to other previously published RT-LAMP assay designs in which primer design focused on the conserved 3’ UTR, NS1 or C-prM regions but detailed limited information about the DENV sequences used to develop the primers. As the LAMP primers were designed from different clusters of each DENV serotype obtained after PCA and phylogenetic analyses, the individual LAMP amplicons locate to several regions across the DENV genome conserved in these clusters (Fig 1). This allows an overall detection of DENV variability surpassing any other molecular amplification assay. The final amplicons were selected through a combination of in silico primer dimer formation assessment (Visual OMP) and in vitro assessment by checking amplicons selected in the first step for unspecific amplification in the NTC. A similar methodology has been used to design RT-LAMP primers to detect Chikungunya virus (manuscript submitted to PLoS Neglected Tropical Diseases) and we consider this approach would be suitable for the assay development of other infectious diseases. The final DENV1-4 specific RT-LAMP assays contained 84, 72, 48 and 18 oligonucleotides respectively. The challenge was to find a working concentration of these oligonucleotide mixes, which would allow for sensitive detection. A 2-fold dilution series approach for the individual final primer mix allowed to identify a working concentration window in the dynamic range of these assays. This however came at the cost of run time. In order to increase the reaction speed without losing sensitivity, several combinations of enzymes were tested. We tested the combination of AMV RT (Promega, Southampton, UK) and GspSSD DNA polymerase (Optigene) recommended by others who successfully developed rapid RT-LAMP assays with 10–15 minute run times [57] (Manuguerra personal communication). We also tested Bst 3.0 DNA polymerase (New England BioLabs), but found that none offered an advantage over the enzyme combination we used (Transcriptor Reverse Transcriptase and Bst 2.0). As a matter of fact, we saw an increased level of unspecific amplification with Bst 3.0 DNA polymerase (data non-shown).
Thus currently reaction times range from 45 (DENV1) to 90 minutes (DENV2). This was not correlated with the number of oligonucleotides in the mixture but may reflect the efficiency of the individual primer sets in the mixture detecting the respective standard strains we used for the validation, and the low oligonucleotide concentration. Alternative approaches to evaluate the sensitivity of each RT-LAMP would consist of having either a pool of RNA samples representative for each amplicon included or specific primer sets for each particular DENV strain that would be compared with the primer mixtures included in the developed assays.
We used an RNA standard evaluated by qRT-PCR to quantify viral RNA of DENV1-4. These quantified RNA were then used to test the analytical sensitivity of the 4 individual specific RT-LAMP assays for the detection of each serotype. The analytical sensitivities of the DENV1-4 RT-LAMP assays, as estimated per probit analysis, ranged from 22 to 641 RNA molecules detected, and 100% reproducibility after 8 independent runs was achieved for 102−103 RNA molecules detected.
Therefore, results were in the range observed for previously described RT-LAMP methods detecting all four serotypes in a single reaction [8, 33, 37] with sensitivities between 10 and 100 RNA molecules detected, and RT-LAMP assays distinguishing the serotypes in individual reactions [30, 38]. For the latter assays the analytical sensitivities determined were 10 to 100 plaque-forming units (PFU)/mL and 10 RNA molecules detected respectively. Our RT-LAMP assay for DENV1 showed a limit of detection as per probit analysis of 102 PFU/mL with a confidence interval from 20 to 7.8x103 PFU/mL (data non-shown).
The assays developed were serotype-specific, and no cross-detection of other flaviviruses was observed. Surprisingly, 2 viral preparations tested—KDH0010A (DENV3) and VIMFH4 (DENV4)—were also found positive for DENV1. Subsequent analysis by serotype-specific nested PCR [20] confirmed the presence of DENV1 RNA probably due to contamination during RNA extraction or virus culture, and indicating that the DENV RT-LAMP assays had picked up the contamination correctly.
EQA panels have been developed in order to evaluate the performance and reliability of current diagnostic methods in laboratories worldwide, by using different samples (both negative and positive samples, including different concentrations) that provide information about their specificity and sensitivity [58, 59]. The EQA panel used in this study, provided by QCMD, comprises strains for the 4 DENV serotypes, as well as negative samples. The analysis showed that our RT-LAMP assays passed all the samples included in the 2015 DENV EQA panel, consisting of 8 core and 2 educational samples.
For evaluation with clinical material, RNA was extracted from whole blood samples collected in Tanzania, confirmed as DENV2 positive by qRT-PCR. A bead-based extraction protocol was improved and, in addition, instead of using 50 μL whole blood and eluting in 200 μL RNA, the extraction commenced from 100 μL whole blood and RNA was eluted into 100 μL. Due to this improved extraction protocol, time to positivity reduced from 81–89 min to 55–77 min.
In some cases, there were disparate results between RT-LAMP and qRT-PCR. Sample 1232, negative by RT-LAMP, had a CT value of 28.78, and samples 1241 and 1473, with CT values of 24.27 and 29.13, showed current mean TT values of 70 and 73.9 min, respectively. These differences in results observed may not be related to the sensitivity levels of the individual assay and we suggest that the performance of isothermal amplification reactions could be compromised when not using fresh samples, as previously described [53].
All 11 serum samples collected by Institut Pasteur in Dakar (2014), tested positive by qRT-PCR and the DENV1 and DENV2 RT-LAMP assays. While 3 of the samples could not be characterised with the qRT-PCR protocol, they were successfully amplified by the DENV1 RT-LAMP, providing evidence that determination of serotype is possible when handling samples that have not been serotyped yet.
Based on the results obtained for the fever study in Tanzania, our DENV2 RT-LAMP scored a sensitivity of 95.8% (CI: 0.79–1.00) and specificity of 100% (CI: 0.63–1.00) in reference to the qRT-PCR used by the Swiss Tropical and Public Health Institute, indicating that all detected as positive by the LAMP assay were truly positive and no false positives were detected.
We used predried tubes of RNAstable for shipment of DENV4 and DENV3 RNA extracts from Institut Pasteur du Cambodge. The efficiency of this type of shipment at ambient temperature was disappointing. Surprisingly DENV3 sample RNA extracts suffered most from this type of shipment and this could not be improved in altogether three shipment trials. The results for DENV4 samples indicate specific detection which does not quite match the qRT-PCR sensitivity. DENV3 samples were detectable but sensitivity could not be assessed.
The determination of clinical sensitivity, specificity, PPV and NPV allows interpretation of diagnostic results for clinical decisions [60, 61]. The scores obtained for specificity, sensitivity, PPV and NPV were in the range observed for previously published assays [8, 30, 33, 36–38].
The scores obtained for PPV and NPV estimate the probability that the disease is present or absent depending of the result is positive or negative. Since the samples were collected in a fever study, the results obtained with the RT-LAMP (PPV = 100% and NPV = 85.7%) highlight a good performance of the method in determining true positive cases while excluding negative cases. PPV and NPV are very dependent of the number of positive and negative samples used, providing valuable information during naturally occurring infections in prospective trials. The values obtained in our study may not reflect this as only thirty samples were analysed and a larger number of both positive and negative samples would be needed to refine these results.
To conclude, we have shown a novel approach to designing LAMP primers that makes use of fast growing sequence databases. During the study time the number of complete DENV genome entries grew by 932 genomes deposited. To be able to cover all of the diversity documented, our approach devised 4 complicated mixes of oligonucleotides for the detection of the individual DENV1-4 serotypes. The DENV1 and DENV2 assays were validated with viral RNA extracted clinical samples and showed very good performance parameters. Finally the combination of PCA analysis and molecular detection assays design should also be considered for other molecular assay formats since the available sequence dataset of several pathogens has increased beyond what can be handled by traditional design based on simple alignments.
|
10.1371/journal.pgen.1005235 | Monoallelic Loss of the Imprinted Gene Grb10 Promotes Tumor Formation in Irradiated Nf1+/- Mice | Imprinted genes are expressed from only one parental allele and heterozygous loss involving the expressed allele is sufficient to produce complete loss of protein expression. Genetic alterations are common in tumorigenesis but the role of imprinted genes in this process is not well understood. In earlier work we mutagenized mice heterozygous for the Neurofibromatosis I tumor suppressor gene (NF1) to model radiotherapy-associated second malignant neoplasms that arise in irradiated NF1 patients. Expression analysis of tumor cell lines established from our mouse models identified Grb10 expression as widely absent. Grb10 is an imprinted gene and polymorphism analysis of cell lines and primary tumors demonstrates that the expressed allele is commonly lost in diverse Nf1 mutant tumors arising in our mouse models. We performed functional studies to test whether Grb10 restoration or loss alter fundamental features of the tumor growth. Restoring Grb10 in Nf1 mutant tumors decreases proliferation, decreases soft agar colony formation and downregulates Ras signaling. Conversely, Grb10 silencing in untransformed mouse embryo fibroblasts significantly increased cell proliferation and increased Ras-GTP levels. Expression of a constitutively activated MEK rescued tumor cells from Grb10-mediated reduction in colony formation. These studies reveal that Grb10 loss can occur during in vivo tumorigenesis, with a functional consequence in untransformed primary cells. In tumors, Grb10 loss independently promotes Ras pathway hyperactivation, which promotes hyperproliferation, an early feature of tumor development. In the context of a robust Nf1 mutant mouse model of cancer this work identifies a novel role for an imprinted gene in tumorigenesis.
| Cancer-causing mutations typically involve either allele inherited from parents, and the parental source of a mutant allele is not known to influence the cancer phenotype. Imprinted genes are a class of genes whose expression is determined by a specific parental allele, either maternally or paternally derived. Thus, in contrast to most genes, the pattern of inheritance (maternal or paternal-derived) strongly influences the expression of an imprinted gene. Furthermore, imprinted genes can be differentially expressed in different tissue types. This work identifies a novel link between cancer and Grb10, an imprinted gene involved in organismal metabolism and growth. In our mouse model of radiation-induced tumors, we found monoallelic Grb10 gene loss involving the parental allele responsible for protein expression. Tumors harboring genetic loss of the expressed Grb10 allele showed absent transcript and total protein levels, despite an intact remaining wildtype Grb10 allele identified by sequencing. When restored, Grb10 suppressed tumor growth by down-regulating Ras signaling. This work demonstrates a new role for an imprinted gene in tumor formation, and shows that Grb10 functions to negatively regulate Ras signaling and suppress hyperproliferation.
| Diverse types of somatic genetic alterations occur in cancers and play important roles in pathogenesis. A common cancer-promoting mechanism is the homozygous loss of a tumor suppressor gene, for example Tp53 [1]. Classically, loss of tumor suppressor genes requires bi-allelic loss or inactivation, conforming to Knudsen’s two-hit hypothesis.
Tumor-promoting somatic mutations involve either allele, and the parental source of a mutant allele is not known to influence the cancer phenotype. A small fraction of genes, known as imprinted genes, are characterized by monoallelic expression from a single parental allele [2]. Heterozygous loss of the expressed parental allele produces a functionally nullizygous state [3]. Thus, the imprinting mechanism modulates gene expression in a manner that defies Mendelian predictions. To date, imprinted genes are not known to have a role in promoting the development of malignancies.
The tumor suppressor NF1 gene, and its conserved murine homologue Nf1, encode the neurofibromin protein, which is ubiquitously expressed in mammalian cells and necessary for development [4]. Germline heterozygosity for NF1 causes Neurofibromatosis I (NF1), an autosomal-dominant inherited disease with an incidence of 1 in 3000 live-births [5]. The development of benign and malignant neoplasms, typically during childhood, is a well-recognized feature of Neurofibromatosis I [5]. Furthermore, tumor genome analyses of diverse cancers have identified NF1 mutations in sporadic but lethal cancers arising in adults, such as malignant brain tumors, ovarian cancers, and lung cancers [6–9].
The NF1 gene encodes the neurofibromin protein, which functions as a Ras GTPase activating protein (GAP) [10], and loss of neurofibromin promotes hyperactivation of Ras signaling [11]. Oncogenic, constitutively activated Ras is frequently found in human cancers [12] and has been shown to play a causal role in tumor formation in many genetic models [13]. Although neurofibromin is a tumor suppressor protein, NF1 loss alone is not sufficient to promote tumorigenesis. NF1-mediated tumorigenesis may thus require additional mechanisms to pathologically dysregulate Ras signaling, and as a consequence, additional therapeutically actionable steps may exist for inhibiting Ras signaling in the NF1 null context.
To identify novel mutations and mechanisms that promote tumorigenesis with Nf1 loss, we mutagenized mice heterozygous for Nf1 with fractionated ionizing radiation [14,15]. These mouse models recapitulate clinical second malignant neoplasm (SMN) induction observed in NF1 individuals, and provide a novel approach for identifying the molecules cooperating in this process. Ionizing radiation exposure induces mutations, some of which may cooperate with Nf1 heterozygosity to promote tumorigenesis. Mutagenizing Nf1+/- and wildtype mice with ionizing radiation generated diverse malignancies [14,15] from which we generated a unique panel of mouse tumor cell lines. Expression analysis of these lines revealed decreased Growth factor receptor bound protein 10 (Grb10) mRNA in Nf1 null tumor cell lines compared to controls.
Grb10 is an adaptor protein that interacts with multiple receptor tyrosine kinases (RTK) [16,17]. Grb10 possesses a plekstrin homology (PH) domain, a Ras-Association (RA) domain, and a C-terminal Src homology 2 domain (SH2) [18], and associates with the insulin receptor, insulin-like growth factor receptor and epidermal growth factor receptor with variable affinities [19,20]. Grb10 is described to interact with proteins functioning downstream of RTKs such as Raf1 and MEK1 although the biological significance of these interactions are unclear [19]. Grb10 is also linked to Ras signaling through mTORC1 (mammalian target of rapamycin complex 1), which can phosphorylate Grb10 and regulate its levels by influencing protein stability [21,22].
Although biochemical evidence position Grb10 at multiple nodes in RTK signaling, questions persist concerning Grb10’s basic functions. In cell-based studies utilizing cultured fibroblasts, Grb10 promotes cell proliferation and survival [23]. In vivo evidence, however, indicate that Grb10 is an important negative modulator of proliferation and growth in tissues. Overexpression of Grb10 in transgenic mice results in growth retardation and insulin resistance [24,25]. Conversely, in vivo loss of Grb10 in mouse models increases animal size due to hyperproliferation of peripheral tissues, although these animals have no apparent propensity to develop cancers [26]. Analysis of enlarged muscle in Grb10-deleted mice reveals increased myofiber number rather than size, a phenotype that is present at birth and maintained throughout adulthood [27]. Thus, in vivo data support a role for Grb10 as a negative regulator of proliferation and RTK signaling. Grb10 is not known to have a role in tumor suppression, although Grb10 expression is reduced in a wide range of human cancers [21].
Grb10 is an imprinted gene in human and mice [3]. In the mouse, Grb10 is expressed from the maternal allele in non-central nervous system (CNS) tissues [3]. Interestingly, radiation-induced tumors from our models all occur in non-CNS tissues, where Grb10 is expressed from the maternal allele. Analysis of maternal and paternal-specific genetic polymorphisms established that the maternally expressed Grb10 gene was lost in cis with wildtype Nf1 in most radiation-induced tumors, providing a genetic mechanism for functional Grb10 nullizygosity in tumors. Functionally, restoring Grb10 protein in Nf1 null tumors suppressed tumor growth in a MAPK-dependent mechanism. Conversely, Grb10 silencing promotes Ras signaling in and hyperproliferation of MEFs. This effect was Nf1-independent, although the most profound increase in Ras pathway activation occurred when both Grb10 and Nf1 were silenced. In human cancers, we found evidence for co-loss of neurofibromin and Grb10 expression in human glioblastoma, a tumor type in which NF1 is among the most significantly mutated genes [6]. Human tumor sequencing databases reveal that the Grb10 and NF1 genes can be co-mutated in diverse tumor histologies. In summary, this work identifies a role for an imprinted gene in promoting central features of tumorigenesis. In this context, we show that Grb10 is a negative regulator of Ras signaling, and contributes significantly to Ras dysregulation in the setting of Nf1-mediated tumorigenesis. These findings demonstrate a previously undescribed role for an imprinted gene in disease and suggest that mutations in imprinted genes should be considered with regard to parental origin.
In earlier work, we mutagenized Nf1+/- and control wildtype mice to model second malignant neoplasms, severe complications that individuals with the NF1 syndrome develop can develop after radiotherapy [15]. C57Bl/6/129Sv Nf1+/- mice exposed to focal, fractionated ionizing radiation developed diverse malignancies, including soft tissue sarcomas, mammary carcinomas and squamous cell carcinomas. We established multiple tumor cell lines from primary radiation-induced tumors, two arising from wildtype mice (cell lines 867 and 963) and ten arising from Nf1+/- mice [14,15]. Using single nucleotide polymorphism and microsatellite analysis as previously described [15], we found that tumors from mutagenized Nf1+/- mice commonly lose the wildtype Nf1 allele, rendering these tumors null for Nf1. Tumor formation in irradiated Nf1+/- mice is driven by complete loss of Nf1, which is also a hallmark of tumor formation in NF1 patients [28] and thus represents an early and necessary event. To identify mechanisms that are commonly altered in Nf1- mediated tumorigenesis and might function as second events, we interrogated our tumor cell lines using a targeted expression array to compare expression of known PI3K pathway regulators and effectors amongst our tumor lines (Fig 1A). This analysis identified Grb10 expression as most uniformly reduced in tumors as compared to control untransformed Nf1+/- MEFs. We then confirmed by immunoblotting that loss of Grb10 expression observed in Fig 1A was not neurofibromin-dependent (Fig 1B). Using quantitative PCR we independently verified Grb10 expression in wildtype and Nf1 mutant tumor cell lines relative to controls, and found that Grb10 expression varied among multiple postnatal organs included as controls (Fig 1C), with greatest expression in brain and muscle as previously described [29]. In addition to brain and muscle, mammary and skin tissues were included as controls to represent the tissue types in which tumors originated (mammary carcinoma, sarcoma and squamous cell carcinoma). Grb10 expression in 11 of 12 cell lines derived from our mouse models (the 867 cell line being the sole exception) was significantly reduced or undetectable (Fig 1C). Western blotting demonstrated that Grb10 protein was undetectable in all 12 tumor cell lines, including the 867 cell line (Fig 1D).
Grb10 belongs to a family of proteins that include Grb2, Grb7, and Grb14, whose members share protein domains but whose functions are not well-defined. To examine whether loss of Grb10 expression might result in compensatory overexpression of these related Grb family members, we performed quantitative PCR to determine whether Grb2, Grb7 or Grb14 transcript levels increased when Grb10 expression was reduced or lost. Similar to Grb10, Grb2, Grb7 and Grb14 expression varied among different control organs examined, however none were overexpressed in any tumor cell line compared to control tissues, arguing against compensatory overexpression of Grb family members in response to reduced Grb10 expression (S1 Fig).
The Grb10 gene, initially named Maternally expressed gene 1 (Meg1), is imprinted in mice and humans [26,30–32]. Grb10 expression from either the paternal or maternal allele segregates between the central nervous system (CNS) and non-CNS tissues [3]. In mice, paternally-derived Grb10 is expressed exclusively in the CNS, while maternally-derived Grb10 is expressed in remaining non-CNS tissues, such as muscles [3,33]. This pattern of tissue restriction is conserved in humans (CNS versus non-CNS), although the allelic contribution is reversed (i.e. maternal expression in the CNS). Because Grb10 expression is imprinted, monoallelic deletion, depending on the affected cell type, can functionally reproduce homozygous allelic loss [34]. Consistent with this, mouse models with targeted loss of either the paternal or maternal Grb10 allele demonstrate loss of expression in discrete tissue compartments and non-overlapping phenotypes [3].
Grb10 expression is reduced in a variety of human cancers [21], although the underlying mechanism for this reduction has not been defined. Earlier data implicate Grb10 as a candidate modifier gene cooperating with Nf1 loss to promote tumors. Indeed, astrocytoma formation in mice after co-loss of Nf1 and Trp53 is strongly influenced by the parental origin of the mutant chromosome 11 [32,35], on which Nf1,Trp53, and Grb10 genes reside. A mechanism for Grb10 contributing to tumorigenesis either in human or mouse tumors is currently undefined.
We have shown previously that the Nf1 gene and the Trp53 genes on chromosome 11 are co-lost in tumors arising in Nf1+/- mice [15]. To determine whether there was a similar genetic basis for Grb10 loss in our tumors, we exploited the F1 background of our model to analyze for loss of heterozygosity (LOH) in Grb10. Tumors from Nf1 mutant mice were analyzed using Illumina Medium density array-based SNP genotyping to assess for LOH along chromosome 11. Consistent with our earlier findings [15], LOH occurred in chromosomal regions spanning across Nf1 (Fig 2A, S1 Table), and this pattern of chromosomal loss was present in all three major tumor histologies arising from our mouse models (carcinoma, sarcoma and pheochromocytoma). Interestingly, LOH on chromosome 11 extended beyond the Nf1 locus to involve most of the chromosome, raising the possibility that a gene centromeric to the Nf1 gene drives this loss. The extent of LOH on chromosome 11 was similar between carcinomas, sarcomas, and pheochromocytomas arising in Nf1 mutant mice. Interestingly, the Grb10 gene localizes to a region involved with LOH (Fig 2A), suggesting that genetic loss of the expressed Grb10 allele may underlie the absence of Grb10 transcripts in Nf1 mutant tumor cell lines (Fig 1).
To orthogonally validate a genetic mechanism for the loss of Grb10 expression in our tumors we performed microsatellite-based loss of heterozygosity (LOH) analysis. The design of our mouse models employed a fixed breeding schema in which the wildtype Nf1 allele was always maternally-derived (Fig 2B). Our mouse models also utilize an F1 background so that parental alleles can be distinguished at specific loci (Fig 2C). Because the tumors isolated from our mouse models are non-CNS tumors, Grb10 expression would be derived from the maternal Grb10 allele. Analysis of primary tumor samples identified 19 tumors from Nf1+/- mice in which Nf1 and Grb10 status could be co-determined. 17 of these 19 (89%) demonstrated LOH of the maternal Grb10 allele, and 2/19 (11%) had LOH of the paternal allele. Among the 17 tumors demonstrating LOH of the maternal Grb10 allele, in 12/17 (71%) loss occurred in cis with the wildtype Nf1 gene. Loss of the paternally-derived Grb10 allele (in trans) with Nf1 and Trp53 alleles occurred in only 1/19 tumors, and 1/19 tumors showed LOH of paternal Grb10 without Nf1 co-loss. In summary, this analysis independently confirmed Grb10 LOH in the majority of tumors from Nf1+/- mice and localized loss to involve the maternal allele.
Among radiation-induced tumors from wildtype mice, Nf1 and Grb10 status was determined in 7 tumors. Four of 7 (57%) demonstrated maternal Grb10 loss, and the remaining 3 (43%) showed loss of the paternal allele. Of the 4 tumors with maternal Grb10 loss, 2 (50%) had loss of Nf1 (1 in cis and 1 in trans), and in the remaining 2 tumors LOH of Nf1 was not detected (heterozygosity was intact).
Taken together, these data localize the majority of Grb10 loss in both Nf1 mutant and wildtype tumors to the maternal allele, on which the wildtype Nf1 and Trp53 alleles are also lost in Nf1 mutant tumors (Fig 2C and 2D). The co-loss of all three of these genes in cis from chromosome 11 suggests a common genetic mechanism or event driving these losses.
We also performed Grb10 microsatellite analysis on our tumor cell lines (the corresponding primary tumors from which these lines are derived were analyzed above) shown in Fig 1. All lines except for 867 demonstrated LOH of the maternally-derived Grb10 allele (in cis with the wildtype Nf1 allele). The 867 tumor cell line, which was established from a tumor arising in an irradiated wildtype mouse and expresses Grb10 transcript levels similar to normal tissues (Fig 1), demonstrates loss of the paternal Grb10 allele (Fig 2C). 867 is a sarcoma cell line, and loss of the silenced paternal allele (with retention of the maternal allele) likely explains why this tumor has detectable Grb10 transcripts. However, immunoblotting for total Grb10 protein demonstrates absence of Grb10 protein in all tumor cell lines, including 867, suggesting that Grb10 loss in the 867 cell line results from alternative, possibly post-translational mechanisms. We sequenced the Grb10 exons in all our tumor cell lines, which revealed no mutations in the remaining paternal Grb10 allele (sequence data from representative cell lines 881, 963 and 989 are archived at http://www.ebi.ac.uk/ena/browse) and is consistent with preferential loss of the maternal Grb10 allele resulting in functional nullizygosity (Fig 2D). These data support the idea that genetically-mediated loss of Grb10 expression suggests a role for Grb10 in suppressing tumorigenesis.
As a negative regulator of growth factor signaling, Grb10 restoration is predicted to reduce Ras signaling through the downstream PI3K and MAPK pathways. To test whether Grb10 restoration altered tumor formation or Ras signaling, retrovirus was used to stably express either wildtype or mutant Grb10 protein (AA) in multiple Nf1 mutant tumor cell lines. The Grb10 AA mutant bears mutations of the mTORC1 phosphorylation sites Serine 501 and Serine 503 to alanines (AA) [21]. Although serine phosphorylation has been proposed to increase protein stability [21], the level of mutant Grb10 expression was similar to that of wildtype Grb10 in our cell lines (Fig 3A).
Restoring wildtype Grb10 significantly reduced soft agar colony formation by the Nf1 null cell lines 989 (shown in Fig 3B and 3C) compared to control and mutant Grb10. To explore the possible mechanism of Grb10-mediated inhibition of colony formation we assessed signaling downstream of Ras in these cells. Cells expressing wildtype Grb10 demonstrated reduced phosphorylated ERK after serum starvation as well as after insulin stimulation compared to control cells (Fig 3D). These data indicate that restoration of Grb10 can profoundly reduce colony formation, attenuate pro-oncogenic signaling, and decrease basal proliferation in Grb10 deficient tumor cell lines. These data thus identify the loss of Grb10 as a mechanism that supports a significant growth advantage that may be selected for during tumorigenesis.
Expression of the Grb10 AA mutant resulted in colony formation intermediate between wildtype Grb10 and control, suggesting that the mutated Serine 501/503 residues produce partial loss of function.
Mutant Grb10 expression had unexpected effects on signaling, characterized by suppression of MAPK pathway signaling after serum starvation as well as after stimulation with insulin, compared to control. Phosphorylated Akt levels were similar to control, but basal phosphorylated S6 levels were reduced and stimulation was delayed.
Phosphorylation of Akt after insulin exposure was similar in amplitude and duration among the three cell lines. Ras-GTP levels in cells expressing wildtype Grb10 paralleled the phosphorylation of ERK (Fig 3D). The difference in ERK1/2 phosphorylation kinetics after insulin exposure between the wildtype and mutant Grb10, coupled with the colony formation data in Fig 3B, suggests that the mutant Grb10’s less effective inhibition of MAPK signaling may underlie the more modest suppression of colony formation observed with this variant.
Hyperproliferation is a hallmark feature of transformation. Inappropriate cell proliferation occurs early in tumor development, and increasing cell proliferation rates typically correlate with tumor progression and aggressive clinical behavior [36]. Oncogenic activation of Ras signaling pathways in normal cells triggers compensatory mechanisms and negative feedback to suppress inappropriate proliferation, a phenomenon known as oncogene-induced senescence [37]. The feedback mechanisms underlying this protective response are not well-defined, but require p53 and Rb [37,38].
Grb10 expression is reduced in diverse human tumors [21], but the timing of Grb10 loss in tumorigenesis and whether Grb10 loss mediates early features of tumorigenesis are completely unknown. If Grb10 loss occurred as an important event early in tumorigenesis, a likely consequence could be increased cell proliferation, the basis for pre-malignant hyperplasia. Interestingly, the pattern of LOH characterized by Grb10 loss in cis with Nf1 and Trp53 in our tumors is suggestive of a chromosomal break that produced this extensive chromosomal loss early in tumor development. To replicate this genetic signature as an early event in untransformed cells we used lentivirus to stably express shRNA targeting Trp53 alone or with shRNA targeting Grb10 in Nf1 null MEFs (Fig 4A). Silencing Grb10 alone produced no significant change in basal MEF proliferation, as measured by cell counts (Fig 4B). Silencing of Trp53 alone, however, significantly increased cell proliferation rates (Fig 4B), and co-silencing of Trp53 and Grb10 further significantly increased cell proliferation by Nf1 null MEFs over that associated with Trp53 silencing alone (Fig 4B). Hyperproliferation associated with co-silencing was sustained over several days (Fig 4C), indicating that cells failed to invoke compensatory mechanisms to suppress proliferation. We then tested whether the MEF hyperproliferation mediated by Grb10 loss was influenced by Nf1 status, and silenced Trp53 with and without Grb10 in wildtype, Nf1+/- and Nf1-/- MEFs (Fig 4D). Adding Grb10 silencing to Trp53 silencing further increased cell proliferation by day 6 for all genotypes (Fig 4D), indicating that this effect does not require Nf1 loss. However, co-loss of Nf1 and Grb10 was associated with greater cell proliferation as compared to proliferation with silencing of each alone; thus over time, Grb10 loss produced the greatest relative increase in cell proliferation when combined with Nf1 loss, either heterozygous or homozygous (Fig 4D). To further determine whether the increased cell numbers associated with Grb10 silencing reflect proliferation as opposed to altered cell loss, we assessed the proliferative index by BrdU labeling (S2 Fig). This data indicates that Grb10 silencing in MEFs significantly increases BrdU incorporation, consistent with Grb10 silencing promoting proliferation.
We tested multiple unique shRNAs against Grb10 and expressed each of these in Nf1-/- MEFs, then assessed the efficacy of silencing by immunoblotting and the quantifying the effect on proliferation. Our panel of Grb10 shRNAs achieved variable degrees of silencing (Fig 4E), and cell hyperproliferation correlated with the degree of Grb10 silencing, with significantly increased cell proliferation requiring near-complete depletion of Grb10 protein levels (Fig 4F).
Grb10 restoration in Nf1 mutant tumor cells reduced colony formation, proliferation, and suppressed Ras signaling, suggesting Grb10’s tumor suppressive effects are mediated by modulating Ras pathway activation. As shown above, Grb10 silencing increased the proliferation of untransformed cells. To determine whether MEF hyperproliferation after Grb10 silencing, similar to tumor cells, involved Grb10-dependent effects on Ras signaling, we used phospho-specific immunoblotting to assess Ras effector activation in the presence or absence of Grb10 silencing. After stably silencing Grb10 and Trp53 in Nf1+/- or Nf1-/- MEFs, cells were serum starved for 18 hours, stimulated with insulin (75 nM), and lysates were collected at 0 and 5 minutes. Immunoblotting for total Grb10 protein confirmed the silencing achieved with shRNA, and interestingly revealed that basal Grb10 protein levels increase when Nf1 is fully lost (Fig 5A), suggesting that Grb10 can be dynamically modulated to negatively regulate Ras signaling. Immunoblotting with phospho-specific antibodies showed that silencing Grb10 increased phosphorylated Akt and ERK levels after serum starvation (Fig 5A, comparing lanes 1 to 3, and lanes 5 to 7) as well as after stimulation, with the greatest absolute levels of phosphorylated Akt and ERK attained in Nf1-/- MEFs after Grb10 silencing (Fig 5A). Paralleling the effects on Ras effectors, Grb10 silencing increased Ras-GTP levels in serum starved Nf1 mutant MEFs stimulated with insulin (Fig 5B). Assessing Akt and ERK phosphorylation at later timepoints revealed that Grb10 silencing was associated with hyperactivation of Akt and ERK at 45 minutes post stimulation with either insulin or EGF (Fig 5C and 5D). Ras signaling hyperactivation mediated by Grb10 loss was independent of Trp53 silencing, as MEFs expressing shRNA against Grb10 only demonstrated increased Akt and ERK phosphorylation after insulin (Fig 5C) and EGF (Fig 5D) stimulation.
To determine whether Grb10 silencing altered levels of the receptors through which stimulation is mediated, we assessed total protein levels of Insulin Receptor (IR), Insulin-like Growth Factor Receptor (IGFR) and Epidermal Growth Factor Receptor (EGFR) in cells after Grb10 silencing (Fig 5C and 5D). Grb10 silencing was associated with decreased total levels of IR, IGFR and EGFR (Fig 5C), indicating a multi-receptor downregulation in response to Grb10 silencing. Furthermore, Grb10 silencing in both wildtype and Nf1 null MEFs was associated with a reduction of total EGFR levels at 45 minutes post-stimulation (Fig 5D). These reductions in total receptor levels would be predicted to reduce proliferation and growth signaling in response to insulin or Epidermal Growth Factor. Although Grb10 silencing reduced total levels of insulin receptor, phosphorylated insulin receptor levels were increased in the setting of serum starvation as well as after insulin stimulation in Grb10-silenced MEFs (wildtype and Nf1-/-) (S3 Fig), suggesting that Grb10 protein suppresses receptor autophosphorylation and degradation. Despite this, Grb10 loss, in an Nf1-independent manner, renders cells resistant to downregulation of Ras signaling by serum starvation, and increases activation of Ras signaling in response to insulin and EGF. Together, these data indicate that Grb10 modulates Ras signaling in untransformed cells as well as tumor cells, and that loss of Grb10 in untransformed cells increases the magnitude and duration of Ras effector activation despite reduced insulin receptor levels.
Although Grb10 levels influence both PI3K and MAPK signaling in untransformed MEFs, in our tumor lines MAPK pathway activation was suppressed by Grb10 restoration. To determine whether Grb10-mediated suppression of colony formation by tumor cells requires inhibition of MAPK signaling, we expressed a constitutively activated Flag tagged-MEK (MEK DD) [39] alone or with wildtype Grb10 in 989 tumor cells. We confirmed expression by immunoblotting (Fig 6A) and assessed phosphorylated Akt and ERK levels (Fig 6A). Restoring Grb10 expression in 989 cells reduced phosphorylated Akt and ERK levels (Fig 6B). Expression of mutant MEK DD restored phosphorylated ERK levels but not phosphorylated Akt (Fig 6A). Comparing colony formation between the cell lines showed a significant reduction in colony formation by 989 tumor cells after Grb10 expression as compared to control, which was rescued with co-expression of MEK DD with Grb10 (Fig 6B and 6C).
Our data are consistent with Grb10 functioning to negatively regulate Ras activation, with Grb10 loss resulting in downstream activation of PI3K and MAPK pathways paralleling Ras-GTP levels. This function had been identified in an Nf1 mutant tumor cell line. To determine whether this effect in tumors was Nf1 dependent, we expressed Grb10 in the sarcoma line 963, which arose in a wildtype mouse and expresses neurofibromin protein. Grb10 restoration significantly decreased pERK levels and proliferation in these tumor cells, indicating that Grb10-mediated growth suppression does not require Nf1 loss (Fig 7A and 7B). We then tested whether Grb10 over-expression suppresses Ras signaling in cells transformed by oncogenically-mutated Ras, a common mechanism for Ras pathway hyperactivation in tumors. We expressed Flag-tagged wildtype Grb10 in human astrocytes transformed with retrovirus encoding V16HRas [40], a mutant Ras that is constitutively activated. Flag-tagged wildtype Grb10 was also expressed in the Nf1 mutant/Grb10 null mouse tumor cell line 881 (neurofibromin and Grb10 protein levels previously shown in Fig 1B and 1D) for comparison. Grb10 expression was confirmed by immunoblotting, which showed that Ras-transformed human astrocytes demonstrated marked overexpression of Grb10 compared to 881 tumor cells (Fig 7C). Grb10-overexpressing V16HRas-transformed human astrocytes demonstrated sustained phosphorylated ERK, in contrast to the Nf1 mutant tumor cell line 881, which similar to the 989 tumor cell line showed significantly reduced MAPK signaling and slightly reduced Akt phosphorylation (Fig 7C). We assessed cell proliferation to determine whether similar to the MEFs and Nf1 mutant tumor cell lines Grb10 expression suppressed cell proliferation of Ras-transformed cells. Grb10 restoration significantly decreased proliferation by the Nf1 null 881 tumor cells, and to a lesser extent, reduced cell proliferation in V16HRas-transformed human astrocytes (Fig 7D). This modest effect in V16HRas-transformed human astrocytes may reflect the known resistance of mutant Ras to modulation and may be mediated by Grb10’s effect on endogenous wildtype Ras present in these tumors cells. Together these results point to a mechanism of Grb10 action upstream of Ras activation.
We assessed human malignant peripheral nerve sheath tumors that arose in individuals with NF1 for total Grb10 protein levels. SNF02.2, SNF94.3 and SNF96.2 are established human MPNST cell lines harboring mutant NF1 [41]. Two out of the 3 lines examined (lines 02.2 and 94.3) showed decreased levels of Grb10 protein (Fig 7E). Restoring Grb10 expression in lines 02.2 reduced MAPK, but not PI3K activity, as measured by phosphorylated ERK and AKT, respectively (Fig 7F). Grb10 restoration also decreased proliferation in line 02.2, mirroring the previous effects observed in the sarcoma line 989 from our mouse model (Fig 7G).
This work identified Grb10 as genetically lost in most tumors developing in our Nf1 mutant-based mouse model. Our subsequent functional studies show that Grb10 loss promotes oncogenic signaling through hyperactivation of Ras signaling as well as the downstream MAPK pathway. The biochemical data are consistent with Grb10 protein functioning as a negative regulator of growth factor receptor signaling, and when lost with another negative regulator of Ras contributes to tumorigenesis. Although the initial identification of lost Grb10 expression was made in our Nf1 mutant-based mouse model of cancer, studies in human tumor cell lines as well as malignant cells generated from Nf1 wildtype backgrounds suggest that Grb10 loss may have tumor-promoting consequences in other oncogenic contexts.
This work sheds light on the cellular responses to Grb10 loss, alone and when combined with Nf1. Total Grb10 protein levels increased in MEFs with Nf1 loss alone. These findings indicate that Grb10 levels are modulated in response to alterations in Ras pathway control, suggesting that increased Grb10 expression is a compensatory response to Nf1 loss. However, this compensation is incomplete and the physiologic increases in Grb10 protein observed in Nf1 null MEFs fail to functionally replace neurofibromin, as gauged by Ras pathway activation. This is evidenced by the fact that Nf1 null MEFs demonstrate significantly increased Ras pathway activation compared to Nf1+/- MEFs (Fig 5A) despite increased Grb10 expression. Our experiments employing shRNA in MEFs show that silencing Grb10 increases Ras signaling in an Nf1-independent manner, and in an Nf1 null context Ras activation can be increased further by Grb10 silencing, and this is associated with a proliferative advantage exceeding that conferred by Nf1 loss alone (Fig 4). Total Grb10 protein levels are rapidly modulated in other cellular contexts with physiologic consequences, for example adipocytes exposed to pharmacologic mimetics of cold stress [42]. We also found that Grb10 silencing has upstream consequences at the level of membrane receptors. Grb10 silencing was associated with decreased protein levels of multiple growth-promoting receptors, namely EGFR, IR and IGFR. Together, these data implicate Grb10 as a player in the dynamic modulation of receptor tyrosine kinases and the regulation of Ras signaling.
Grb10 is a substrate for mTORC [21,22], and in brown adipocytes, it was recently described that Grb10 functions in a feedback loop with the insulin receptor and raptor to inhibit mTORC1 signaling [42]. In that work, transient overexpression of wildtype Grb10 in HEK293 cells was associated with decreased phosphorylation of S6Kinase after insulin exposure [42], an effect that was abrogated when mTORC1 phosphorylation sites on Grb10 were mutated to alanine. In our study, restoration of wildtype Grb10 expression in Nf1 mutant tumor cells lacking Grb10 expression failed to suppress phosphorylation of S6Kinase upon exposure to insulin (Fig 3D). This difference between HEK293 cells and our Nf1 mutant tumor cells may reflect the possibility that a Grb10-mediated negative feedback loop acting upon mTOR also requires fully intact Ras regulation, such as neurofibromin. Furthermore, Grb10 has tissue-specific functions in vivo, and an additional possibility is that Grb10’s function in a transformed cell may be specific to both the cell of origin as well as the mutational background.
Grb10 restoration in Ras-transformed human astrocytes decreased cell proliferation (Fig 7D), suggesting that even in cells expressing mutant Ras, Grb10 levels influence tumor growth. This finding does not invalidate the observation that Nf1 mutant cells benefit from Grb10 loss. Thus, Grb10 loss may not be limited to NF1 mutant tumors, but may operate in other genetic contexts to promote hyperproliferation.
As a negative regulator of Ras signaling, Grb10 also modulates PI3K and MAPK signaling pathways, although our experiments involving untransformed MEFs and tumor cells suggest that the precise downstream effects may differ depending on context. Grb10 knockdown in Nf1 WT and null MEFs increased both ERK and Akt phosphorylation, consistent with Ras activation and somewhat symmetric activation of downstream PI3K and MAPK signaling arms of the Ras pathway cascade. This activation pattern contrasts with the signaling consequences of Grb10 restoration observed in tumor cells. In the tumor cell lines tested here, Grb10-mediated growth suppression is largely mediated by dampening MAPK rather than PI3K signaling, possibly reflecting intrinsic dependencies of these tumors on activated MAPK signaling. The Grb10 protein is phosphorylated by mTORC1, which stabilizes it and supports its function as a negative regulator of PI3K, thus completing a negative feedback loop that has been described in MEFs [21]. Consistent with this model, we found that Grb10 silencing increased phosphorylated Akt levels, however phosphorylated ERK and Ras-GTP levels also increased, consistent with Ras activation. In Nf1 null tumors, restoration of Grb10 reduced levels of ERK and Akt, although the magnitude of this effect was unequal between these pathways.
Grb10 possesses multiple functional domains and phosphorylation sites [21], however the functional roles of these areas are not well understood. Our data suggest that alterations in Grb10 phosphorylation sites influence the dynamics of the signaling, which are evident in the signaling differences observed between wildtype Grb10 and mutant Grb10 AA. While 989 cells expressing wildtype Grb10 demonstrate the predicted response to insulin stimulation characterized by transient increase in pERK levels, peaking at 5 minutes post-stimulation, and followed by attenuation of the signal by 15 minutes, 989 cells expressing the mutant Grb10 AA have a decreased phosphorylated ERK with serum starvation and more sustained ERK phosphorylation in response to stimulation (Fig 3D). Wildtype Grb10 normalizes the signal upon stimulation with insulin, while Grb10 AA decreases ERK phosphorylation after serum starvation, without affecting the dynamics of the signal. Furthermore, expression of the mutant Grb10 AA was associated with elevated Ras-GTP levels at 5 minutes, which is not concordant with ERK phosphorylation. This difference may reflect specific functions of the Serine 501/503 residues that are mutated in the Grb10 AA protein. Functional analysis of Grb10’s domains will enable better characterization of Grb10’s effects on signaling by Ras and its effectors, which may be context-dependent.
Given Grb10’s functional connection to well-described tumor promoting signal transduction pathways, the literature suggests that Grb10 function could contribute to cancer development. Grb10 has not previously been shown to be intrinsically oncogenic either in vivo or in vitro, nor have tumor-promoting functions of mTOR, the best understood phosphorylator of Grb10, been shown to be Grb10-dependent. However, mTOR clearly links proliferative signaling to protein translation and plays an important role in many cancers [43–46]. mTORC1-driven inhibition of the PI3K pathway signaling occurs through Grb10 [22], and this negative feedback mechanism from mTOR to PI3K signaling suggests a potential role for Grb10 loss to promote cell proliferation. Although speculative, Grb10 likely functions in a complex network with mTOR and its other effectors, and its contribution to oncogenic signaling may depend upon both the signaling and cellular backgrounds.
Our data indicates that Grb10-associated tumorigenesis is conditional on Grb10 loss cooperating with other tumor-promoting genetic events, as Grb10 silencing alone failed to confer anchorage independent growth to MEFs. This is also consistent with the Grb10 mutant phenotype in vivo, as Grb10 knock-out mice do not spontaneously develop tumors [3]. Genetic loss was responsible for loss of Grb10 expression in the majority of our tumors. In humans, the genetic basis for GRB10 and NF1 co-loss differs from the genetics in murine tumors, because the GRB10 and NF1 genes reside on different chromosomes (chromosome 7 and 17, respectively). Therefore, loss of these genes in human cancers will be independent events rather than occurring by a single genetic event as suggested by our mouse model. In addition, non-genetic alternative mechanisms might be operational in human cancers promoted by GRB10 and NF1 co-loss. Indeed, our data suggest the presence of post-translational mechanisms mediating Grb10 loss, as among our tumor cell lines one tumor (the 867 tumor cell line) demonstrated Grb10 transcript levels comparable to normal control tissues but lacked detectable Grb10 protein. However, there are likely post-translational mechanisms mediating Grb10 loss, as among our tumor cell lines one tumor (the 867 tumor cell line) demonstrated Grb10 transcript levels comparable to normal control tissues but lacked detectable Grb10 protein. Given multiple potential mechanisms for Grb10 protein function, Grb10 protein levels may be more informative than genetic or transcript analysis alone in determining GRB10 status in human cancers.
This work illustrates that Nf1-mediated tumorigenesis can be promoted by loss of another negative regulator of Ras signaling, and raises the possibility that other negative regulators may contribute to Nf1-driven tumorigenesis in other contexts. Potential candidates for these alternative negative regulators might include some Grb protein family members. This family includes Grb7, Grb10 and Grb14, which are structurally related multi-domain adapter proteins with overlapping and distinct functions. All three family members share a polyproline stretch, PH, RA, SH2 and PBS domains, however their functions differ. While Grb7 was shown to play a role in adhesion/migration by associating with membrane regions of focal adhesion kinases and the EphB1 receptor [47], Grb10 and Grb14 have been implicated in the regulation of insulin receptor and insulin growth factor receptor signaling, and potentially other receptor tyrosine kinases [48]. Mouse models of Grb10 and Grb14 demonstrate the overlapping, distinct, and tissue-specific in vivo functions of Grb10 and Grb14 in insulin signaling regulation. Specifically, Grb14 has IR-mediated growth inhibitory effects in the liver and retina, whereas Grb10 plays a role in insulin signaling regulation in the muscle and adipose tissue [48–50]. As such, Grb10 mutant mice with loss of the expressed allele have a 30% increase in body size, whereas Grb14-/- mice are of normal size. Recent structural studies have also shown that Grb10 and Grb14 have different affinities for binding RTKs, with Grb10 having a higher affinity of binding to phosphor-inositol phosphates on the membrane through its PH domain, while Grb14 has greater affinity for binding Ras molecules [18,51]. In our studies, Grb10 was found to be uniformly lost in the mouse tumors but we found no evidence of a compensatory increase in the levels of other Grbs, including Grb14, the most closely related Grb family member. Perhaps this is not surprising, given that most of the tumors from the irradiated Nf1+/- mice are sarcomas derived from muscle tissue, where Grb10 has a more prominent role than Grb14 [27,52].
This work describes a previously unknown role for an imprinted gene as a tumor suppressor. Conceptually, restricting gene expression to originate from a single allele creates haploinsufficiency, which can expose an organism to disorders caused by loss of protein expression, cancer being an example. Presumably imprinting confers a compelling organismal advantage to offset this risk, and recent data support the idea that Grb10 imprinting is evolutionarily driven by nutrient utilization during development, and Grb10 influences proportions of lean and fat tissues during development [53]. An intriguing connection between Grb10 and tumorigenesis suggests that genetic mechanisms that initially developed to confer an advantage in energy storage and utilization during development might in later life limit a cell’s ability to suppress inappropriate proliferation.
Imprinting means that loss of either allele will impact different cell types variably, and that the cellular context determines the consequence of genetic loss. This feature of imprinted genes has implications for the types of tumors that can be expected to arise after Grb10 loss. The breeding schema of our mouse model produced F1 mice inheriting the mutant Nf1 copy paternally, and subsequent co-loss of Grb10 in cis with Nf1 from the maternal copy of chromosome 11 is equivalent to Grb10 nullizygosity in non-CNS tissues, where only the maternally-inherited Grb10 allele is expressed. This pattern of loss does not produce Grb10 nullizygosity in tissues expressing the paternal Grb10, for example tissues of the nervous system. In our breeding schema, the paternal Grb10 allele is located in cis with the mutant Nf1 allele, and we postulate that the Grb10 imprinting unique to tissues of the nervous system confers resistance to tumorigenesis in the nervous system. Indeed, irradiated F1 mice in our mouse models do not develop CNS tumors [14], despite the fact that optic pathway gliomas (OPGs) are a type of CNS tumor that arises in 15–20% of children with NF1 [54]. The genetic factors responsible for OPG development in NF1 are not fully understood, however our experimental model provides a robust context in which to assess possible tumor promotion by loss of the imprinted paternal Grb10 allele. In prior studies, Grb10 has been proposed as a candidate modifier gene in Nf1 mutant mice [32,35]. Analysis of tumors from our mouse models supports this role of Grb10 and sheds light on the mechanism and consequences of its loss. Additional experiments in Grb10 mutant mice will also be important for further elucidating the function of this protein in vivo. Apart from the development of malignant tumors in the Nf1 mutant mouse background, genetic modifiers may hold broader relevance for the NF1 mutant phenotype in humans. The severity of Neurofibromatosis I varies amongst family members sharing a germline mutation [55], and our data supports the concept that imprinted alleles function as disease modifiers and may be involved in mediating this variability. Candidate genetic modifiers have been identified in multiple sporadic human cancers [56,57]. Interestingly, polymorphism analysis of gliomas in NF1 patients found correlation between specific polymorphisms in the human adenylate cyclase 8 gene with glioma development in a sex-specific manner [58]. Sex-specific modifiers are driven by the sex of the affected individual and fundamentally differ from imprinted modifiers, which are defined by the parental allele involved. Finally, imprinted genes may influence an individual’s susceptibility to tumor development as well as potentially other diseases and represent a novel genetic mechanism that defines both normal physiology and disease.
Tumor samples obtained from previously described mouse models were analyzed. All animal procedures were approved by the UCSF IACUC (Approved protocol numbers: AN078941 and AN080665). These practices conform to regulations defined by the Animal Welfare Act and the US Department of Agriculture.
Mouse tumor cell lines were grown as previously described [40]. Cell lines used were established from tumors arising in irradiated Nf1+/- F1 mice [14,15,59]. Human cell lines were obtained from ATCC.
Retroviral vectors including HA-tagged retroviral WT-Grb10 and Grb10-AA (S501A–S503A) constructs were kindly provided by Dr. Yonghao Yu, (University of Texas Southwestern Medical Center). To generate retrovirus Ecopack packaging (5 μg) plasmids were co-transfected with the Grb10 plasmids (10–15 μg) into HEK 293 T cells using lipofectamine 2000 (Invitrogen). Lentivirus vectors including Flag-tagged WT-Grb10 and MEKDD constructs were cloned as described in the section Gateway Cloning.
To generate lentivirus, Grb10, MEKDD or GFP (9μg) plasmids were co-transfected with packaging plasmids Δ8.9 (9 μg) and VSVG (4.5 μg) into HEK293T cells using lipofectamine 2000 (Invitrogen).
Two days after transfection, retroviral or lentiviral supernatants were harvested and filtered. Recipient cells were infected in the presence of antibiotic-free, serum-containing medium supplemented with 8 μg/ml polybrene. Following infection, cell lines stably expressing the Grb10 constructs were selected using puromycin. Stable expression was confirmed by Western analysis to visualize the HA tag.
Cells from all experimental groups were plated in triplicates at the same density (ranging from 50,000 to 150,000 cells/well in 6-well plates depending on the experiment). Cells were trypsinized and re-suspended in 2 mLs of complete media. Cell counting and assessment of viability by trypan blue staining were performed using an automated cell counter, Vi-CELL XR (Beckman Coulter, Fullerton, CA).
Cell were plated 4 x 104 cells per 6 cm plate in 1x DMEM as previously described [40]. One day following plating, the plates were treated with drug to indicated concentration, or DMSO carrier as control, diluted in DMEM.
For signaling pathway Western blots, cells were collected at 70% confluence. Cells were washed twice in PBS and lysed in RIPA lysis buffer (1% Sodium Deoxycholate, 0.1% SDS, 25 mM Tris, 150 μM NaCl, 1% Triton-X, 0.2 mM EDTA, 10 mM NaF, 1 mM Sodium Vanadate, 10 nM Calyculin A, Protease Inhibitors). Lysate protein concentration was determined by Pierce BCA Protein Assay Kit (Thermo Scientific). 20–50 μg of cell lysate was run on SDS-PAGE 10–20% gradient gels (Novex). Most antibodies were purchased from Cell Signaling Technologies. Primary antibodies were diluted at 1:2000, and included Beta-actin (Cat No. 4967L), phospho-S473-Akt (Cat No. 4060), phospho-S6 (Cat No. 2211), phospho-p44/42 MAPK (Cat No. 4376), IRβ (Cat No.3025), Flag (Cat No. 8146), Phospho-Tyrosine (Cat No. 8954), HA (Cat No.2367). Anti-Grb10 and EGFR antibodies were purchased from Santa Cruz (Cat No. sc-1026, and sc-31157 respectively). Anti-IGFR antibody was purchased from Millipore (Cat No. 05–656). Secondary antibodies were diluted at 1:2000 and included anti-rabbit IgG-HRP (Cat No. 7074) and anti-mouse IgG-HRP (Cat No. 7076). NIH Image 1.49j or Li-Cor Image Studio Lite 4.0.21 softwares were used for densitometric quantification of phosphorylated proteins and receptor levels. For the quantification of receptor levels in the stimulation experiments, receptor band intensities were normalized to Actin as an internal reference. Subsequently all the values were normalized to the initial experimental condition: Luc/ time point “0”. For quantification of phosphorylated protein levels, Total AKT, pAKT, Total ERK, pERK intensity levels were first all normalized to Actin as an internal control, and the normalized values were used to calculate the ratio of pAKT/Total AKT and pERK/Total ERK. Quantifications displayed under the corresponding bands correspond to the values obtained for the represented experiment.
RNA was extracted with the RNAeasy kit (Qiagen). RNA purity was assessed by UV spectrophotometry using a Nanodrop p1000 (Thermo Scientific). cDNA was generated from 1 μg RNA with the Qiagen RT2 First Strand Kit. Genomic DNA was removed with the Qiagen Genomic DNA Elimination Mix. Reverse-transcription mix was added to purified RNA, mixed with the RT2 SYBR Green Mastermix and added to the 96-well PI3K (Cat No. PIMM058A) PCR array plates per Qiagen protocol. RT-PCR was performed over 40 cycles on a Stratgene MX3000P qPCR system. CT values were exported into a Microsoft Excel spreadsheet and uploaded into the Qiagen RT2 Profiler PCR Array Data Analysis Webportal software (v3.5) for analysis (http://pcrdataanalysis.sabiosciences.com/pcr/arrayanalysis.php). Data QC was verified and housekeeping genes were selected for data normalization. Fold change was calculated and used to generate the heat map.
Equal numbers (250,000 cells) of WT and Nf1 null MEFs expressing shRNA against Luciferase (Luc) or Grb10 were plated in 6-well plates. BrdU was added to the culture medium overnight, and the cells were trypsinized and harvested the next day. The cells were fixed and stained with APC-labeled anti-BrdU antibodies as recommended by the BD BrdU flow kit protocol (BDB552598). Percent BrdU incorporation was assessed by flow cytometric analysis (FACSDiva LSRII). For mitotic enrichment 250,000 cells were plated in 6-wells plates and treated with 25 μM MG132 for 2 hrs prior to labeling with BrdU as discussed previously.
Cells were lysed with modified Ripa buffer (50 mM Tris-Cl, pH 7.5, 150 mM NaCl, 1% Nonidet P40, 0.5% sodium deoxycholate, 0.1% SDS, 10 mM NaF, 1 mM Sodium Vanadate, 10 nM Calyculin A, Protease Inhibitors). Lysates were centrifuged at 15,000 g for 20 min at 4°C, and protein concentration was determined by Pierce BCA Protein Assay Kit (Thermo Scientific). 100μg of protein was diluted in a total of 200μl of modified Ripa buffer then precleared with protein A agarose beads (Life Technologies) at 4°C for 1 hr. The supernatants were then incubated with 1 μg of anti-insulin receptor antibody (Cell Signaling) at 4°C overnight. Twenty microliters of protein A agarose beads were then added/ml of lysate, and incubated at 4°C for 1 hr. The agarose beads (with the antibody-protein complex) were then collected by centrifugation (10 min, 14,000 g, 4°C). Supernatants were discarded and beads were washed 3 × with PBS. Finally, the beads were re-suspended in 20 μl of 2 × sample buffer, boiled for 5 min, electrophoresed on 10% Tris-glycine gels (Novex).
Prizm v.4 (GraphPad) was used to calculate paired student’s t tests. Experiments were performed at least three times and means with p <0.05 were considered statistically significant.
See S1 Text for additional experimental procedures.
|
10.1371/journal.pntd.0002721 | Costs Analysis of a Population Level Rabies Control Programme in Tamil Nadu, India | The study aimed to determine costs to the state government of implementing different interventions for controlling rabies among the entire human and animal populations of Tamil Nadu. This built upon an earlier assessment of Tamil Nadu's efforts to control rabies. Anti-rabies vaccines were made available at all health facilities. Costs were estimated for five different combinations of animal and human interventions using an activity-based costing approach from the provider perspective. Disease and population data were sourced from the state surveillance data, human census and livestock census. Program costs were extrapolated from official documents. All capital costs were depreciated to estimate annualized costs. All costs were inflated to 2012 Rupees. Sensitivity analysis was conducted across all major cost centres to assess their relative impact on program costs. It was found that the annual costs of providing Anti-rabies vaccine alone and in combination with Immunoglobulins was $0.7 million (Rs 36 million) and $2.2 million (Rs 119 million), respectively. For animal sector interventions, the annualised costs of rolling out surgical sterilisation-immunization, injectable immunization and oral immunizations were estimated to be $ 44 million (Rs 2,350 million), $23 million (Rs 1,230 million) and $ 11 million (Rs 590 million), respectively. Dog bite incidence, health systems coverage and cost of rabies biologicals were found to be important drivers of costs for human interventions. For the animal sector interventions, the size of dog catching team, dog population and vaccine costs were found to be driving the costs. Rabies control in Tamil Nadu seems a costly proposition the way it is currently structured. Policy makers in Tamil Nadu and other similar settings should consider the long-term financial sustainability before embarking upon a state or nation-wide rabies control programme.
| Rabies is a fatal viral disease. It is transmitted mostly through dog bites in greater parts of Asia and Africa. It is primarily a disease of the poorer population groups with children being the most vulnerable. Control of rabies among humans therefore requires interventions in the animal as well as the human sectors. Animal sector interventions include vaccination accompanied with or without sterilization of dogs. Human interventions are limited to individual vaccination following dog bites. We estimated the costs to the government of rolling out animal as well as human sector interventions across an entire state having a human population of 72 million. We also estimated the major drivers influencing program costs and the implications to the government of adopting such a strategy over a long time. We found that the animal sector interventions were many times more costly than the most expensive human interventions. We also found that in the absence of dog population control measures, it will require substantial financial commitment on the part of the government to be able to invest in dog vaccination strategies.
| While rabies has been identified as a priority zoonoses that needs to be addressed globally [1], it has a special relevance in South Asia. More than 55,000 rabies deaths have been estimated to occur among humans annually with little under half being contributed by India alone [2], [3]. Experts from animal as well as human health sectors agree on the controllable nature of the disease and on the importance of joint population level interventions for restricting disease transmission among animals and humans [4], [5].
Evidence from India and elsewhere demonstrates the efficacy of principle rabies intervention strategies. Indian researchers have studied the application of different post-exposure prophylaxis (PEP) regimens among humans [6]. Indian researchers have also used the experience of dog population control in specific urban settings to demonstrate the impacts of animal birth control strategies [7], [8]. Of late there is mounting evidence produced by international researchers related to the efficacy of anti-rabies immunization among animals in reducing rabies transmission [9]. Economic assessments have also been conducted in different parts of the world which study the economic impact of rabies [2], economics of rabies control [10] and cost effectiveness of different post-exposure prophylaxis regimens [11]. This body of work has been instrumental in development of national strategic plans for rabies control [12].
However, as previously documented, rabies researchers have not been able to satisfy the information needs of policymakers [13] and the economics of rabies control remains a “significant constraint” in rolling out rabies control programmes in low income countries [14], [15]. A possible explanation could be that to date, only a handful of studies have looked at combined costs of rabies across human and animal sectors [2], [10], [16]. Most of these analyses have been conducted from the societal perspective that is of limited use to program managers. Additionally, because of the design of cost effectiveness analyses, their findings are always relative in nature and are difficult to generalise in absolute terms.
Accordingly, we undertook a costing exercise building upon an earlier assessment [17] of rabies control initiative in the Southern Indian state of Tamil Nadu. Its objective was to determine the costs to the government of implementing different combinations of strategies for controlling rabies among human and animal populations in a state like Tamil Nadu.
Tamil Nadu is the southernmost state in India having a population of 72 million [18] and is considered one of the better performing states in public health [19]. According to the results of a study based upon verbal autopsy of deaths between 2001–03, it had 0.5 deaths or fewer per 100,000 human population due to furious rabies [20]. In response to calls for controlling dog bites and rabies, the state government formed a state level rabies coordination committee in 2008 to develop and manage a multisectoral response to dog bites and rabies in the state. This was the first time a large scale population level rabies control intervention was implemented in a large state in India [17].
As described in Table 1, the human interventions consisted of ensuring availability of anti-rabies vaccine at all government-run health facilities in the state as well as promoting awareness about rabies control across the state. Rabies antibody was not provided universally due to perceived high costs. The animal interventions involved outsourcing of ABC-AR operations to private veterinarians; dog catching operations were handled by local animal welfare organizations in selected urban areas of the state. ABC-AR was conducted throughout the year as specified in the guidelines of Animal Welfare Board of India [21]; vaccination-only strategies, whether parenteral or oral, were not considered. The interventions were supposed to be implemented in a continuous fashion throughout the year and not conducted in a campaign mode. The animal and human sector interventions were implemented by different departments and coordinated at the state and district levels through formal multi-stakeholder coordination mechanisms [17].
Using program data from the earlier assessment in Tamil Nadu, we estimated the annual costs of scaling up those interventions across the state, including rural areas. An activity based costing approach was used. The interventions for human and animal populations were calculated separately and the costs for different components within these interventions were disaggregated. System-wide and environmental interventions, such as waste management and vaccine supply chain management systems were not included in the costing framework (Table 1). Costs were calculated from the perspective of government, which was the provider for bulk of the services. Costs were estimated for five different combinations of interventions described in Table 1. All costs were inflated to 2012 Indian Rupees using national financial data [22] and converted into 2012 US dollars using historical exchange rates [23].
Based upon the existing interventions in Tamil Nadu [17], it was assumed that the entire population (rural as well as urban) would be covered by the expanded intervention. Costs were estimated for two combinations of interventions. Based upon the existing intervention model, the first set of interventions consisted of increased surveillance and awareness, in addition to provision of anti-rabies vaccine (ARV) to all patients reporting dog bites at public health facilities. The second combination of interventions involved an additional component of antibody administration to patients with severe dog bites in addition to the ARV.
Based upon the feedback received from local program managers [24], it was assumed that dog bite cases that report at peripherally located and low-throughput health centres would be provided with rabies vaccine through the easier intramuscular route, while those that report at high-throughput hospitals with better trained personnel would be provided vaccination through the intradermal route. The procurement costs of intradermal and intramuscular vaccine formulations (having different vial sizes) and antibodies were estimated from the state level procurement records [25] and market data, respectively. A standard 30% wastage rate was assumed for both the vaccine formulations in the absence of specific reference points. A lesser wastage rate of 15% was used for the antibody.
The annual number of outpatient visits for dog bites was calculated from the monthly dog bite visits reported by the state disease surveillance system over a twenty month period from January 2008 to August 2009. This was divided by the expected number of hospital visits for each dog bite case, to arrive at the annual number of dog bites in the state. While the national guidelines [26] recommend vaccination only for category 2 and category 3 dog bites, in practice, the vaccine was being administered to all reported dog bite cases, which was factored into our analysis. The proportion of dog bites categorised as ‘severe’ and requiring antibodies was assumed to be 63%, using estimates from other national studies [27]. Based upon the feedback received from program managers, some program administration costs were included to cover expenditure related to awareness generation, training and surveillance related activities.
The then-prevalent model of ABC-AR was selected as one of the intervention strategies. Parenteral vaccination using teams of dog-catchers and oral vaccination were selected as hypothetical intervention scenarios to determine the extent to which costs could be reduced by less resource-intensive exercises.
Using dog population density figures from the livestock census [28], the number of animal sheds (having capacity for 30–45 animals) required to cater to 100,000 human populations were calculated. The fixed and recurrent costs were then calculated for every 100,000 population. The costs for animal interventions were sourced from state program guidelines and adjusted for inflation. Subsequently, differential costing was conducted to include additional stay and veterinary fees for operating on female dogs. More vehicles were assumed to be required in rural areas because of the larger distances to be covered. Therefore, increased capital and fuel costs were considered for dog shelters in rural areas. All capital costs were depreciated over 5 years. Costing for dog catchers' and ambulance drivers' time was done on a monthly basis using state salary norms. Animal census costs were also included as an annual exercise and estimated accordingly.
A base case scenario was constructed for each of the five different combinations of human and animal interventions using the existing or most likely estimates of key input parameters. The values of input data for our analysis were sourced from our review of program documents, published research literature and from our personal observations in the state. A sensitivity analysis was conducted by varying the values of principle input factors. More than 224,000 scenarios of animal and human interventions were tested. The values of input parameters for the base case and alternative scenarios have been described in Supplementary Files. These were refined based upon the feedback received from experts at two different national consultations of Indian rabies experts organized in 2011 [29] and 2013 [30].
Rabies control is a long term proposition, requiring sustained levels of high coverage of interventions in the animal populations [10]. Accordingly, in addition to estimating the annual costs on the basis of a one-time assessment, we also assessed the long term implications of the animal sector interventions. Given the limited data on the impact of parenteral animal vaccination campaigns in mixed ecological settings such as India, we used data from an Indian study [7] describing the impact of dog population management interventions to assess the long term implications of the animal sector interventions.
We projected costs of four interventions—ABC-AR, injectable vaccination, oral vaccination, and a hypothetical intervention coupling injectable vaccination with injectable contraception for 20 years based on 2012 costs. For interventions involving contraception, a decrease in the dog population was estimated from a dog demographic model used earlier in India [7]. The model estimated the change in total stray dog population and the proportion of sterile dogs over a 20-year period given a sterilization rate of 62–87% in several mark-recapture study areas in Jodhpur city, from 2005 to 2007.
Since no other dog demographic models in the Indian context were available, we estimated the total number of stray dogs and its proportion that would be sterile for each year in Tamil Nadu assuming a similar setting and level of coverage. To project costs for future years, annualized capital costs (for 5-year depreciation) were assumed to be constant over 20 years, and recurrent costs were scaled to the projected dog population size in each year. Recurrent costs were calculated separately for the unsterile (requiring vaccination and sterilization) and sterile (requiring only vaccination) dog populations. Interventions which did not involve sterilization assumed a constant dog population. For the hypothetical injectable vaccination and contraception intervention, the additional cost of the injectable contraceptive was assumed to be negligible and the initial cost in 2012 was assumed to be the same as the cost of the injectable vaccine intervention alone in the base case scenario.
The study is based upon one-time costs data collected from state programme managers. Therefore the analysis only considers those human cases that were reported to the public health surveillance system. This is likely to be an underestimate. Moreover, there is limited data on the completion of treatment; and it is possible that a small portion of patients might not complete their treatment, leading to a further underestimate of dog bite incidence rate. Data on categorization of dog bites, dog bite burden among animals and dog ecology is limited. In the absence of more data, the upper and lower bounds of the input parameters were taken from a range of sources, including expert opinion, summarised in Supplementary Files 1 & 2. In the absence of longitudinal data, we used dog demographic projections from an Indian study [7] to estimate the long term resource requirements for different rabies control interventions. However, there is limited information related to the reliability of these findings in rural areas and other parts of India. While recent studies recommend canine vaccination in annual campaigns having coverage exceeding 60%[9], the current analysis estimates the cost of an year-long continuous routine vaccination strategy which is likely to provide a conservative estimate of likely costs. More long-term efficacy studies for different interventions are required to better comment upon their cost effectiveness.
The annual costs of providing post exposure prophylaxis with antibodies for severe dog bites for Tamil Nadu was calculated to be $ 2.2 million (Table 2). This was more than three times the costs of rolling out a vaccine-only program and translates into costs of $ 11 and $ 3, respectively for each dog bite patient vaccinated. Using base case scenarios, the annual costs of implementing ABC-AR, Injectable vaccination and oral vaccine programmes were calculated to be $ 44 million, $ 23 million and $ 11 million, resulting in each dog's vaccination costing $ 22, $ 11 and $ 5, respectively.
On varying the key input parameters, we found that the costs of the human interventions ranged from $3–$82 million, while the costs for animal intervention ranged from $9–$98 million annually. In order to compare the relative effects of different cost components on individual set of interventions, a tornado chart was prepared (Figure 1 & Figure 2) centred around the costs of the base case scenario for each combination of interventions. The value of each cost component was varied to its upper and lower bounds and the impact on the total program cost charted as red and blue bars, respectively.
In the case of human interventions (Vaccine only and Vaccine+Antibodies), health seeking patterns, cost and wastage rates of vaccine and antisera and the burden of dog bites were found to be the major cost drivers causing the greatest fluctuations in the cost of the program. In relation to the other drivers, the use of intradermal versus intramuscular vaccine regimes did not greatly influence program costs. Antibody procurement comprised around 70% of total costs of the human sector interventions, followed by vaccine procurement costs and training & health promotion costs.
In case of animal-sector interventions, the dog population and number of dog catchers required per team appear to be important drivers of ABC-AR programme costs. On the other hand, vaccine costs have greater role to play in influencing costs of vaccination-only programmes. Sex distribution of dogs does not affect total program costs in the long term, even for ABC-AR in which different surgical procedures are required for male and female dogs.
Assuming an average sterilization rate of 62–87%, the dog population size in Tamil Nadu would be expected to decrease by 70% over a 20 year period from an estimated 2,022,055 dogs in 2012 to 615,408 in 2032 (Supplementary File 4). The proportion of sterilized dogs would stabilize at 80% such that the number of dogs needing ABC would decrease by 94% from 2,022,055 in 2012 to 123,082 dogs in 2032.
Projected costs for ABC-AR, injectable vaccination, oral vaccination, and injectable vaccination-cum-contraception are shown in Figure 3. While costs are highest for ABC-AR in 2012, the cost drops quickly and is lower than that for injectable vaccination and similar to that for oral vaccination by 2032 (Table 3). Total costs over the 20-year period are highest for injectable vaccination and are comparable for oral vaccination and ABC-AR. The hypothetical joint injectable vaccination and contraception intervention would result in the lowest cost by 2032 and the lowest total cost.
In keeping with assertions about rabies control being the responsibility of the local governments [31], this study was conducted from the perspective of the state government of Tamil Nadu to inform its efforts to control rabies in the state. While by no means definitive, these results help in identifying major drivers of costs within a range of government sponsored intervention strategies.
The identification of major drivers of programme costs can inform programme management by identifying areas to improve program efficiency and direct research efforts towards development of precise estimates where required. Basic epidemiological parameters such as incidence of dog bites, their categorization and dog population density were found to be among the major drivers of the program costs. These knowledge gaps require more attention from researchers and can be easily filled through focussed research studies. From the program perspective, the procurement and wastage rates of vaccine and antibodies were found to greatly influence total program costs. Strengthening of local procurement and supply chain management systems and negotiating long-term procurement rates are some of the options that could help offset these costs.
Using base case scenarios, scaling up the existing animal interventions (ABC-AR) in Tamil Nadu would require 20 to 65 times the funds required for scaling up human post exposure prophylaxis alone. Moreover, a combination of human Post-exposure prophylaxis with ABC-AR would cost over 2.1% of the annual budgetary allocations for the departments of health, animal husbandry and municipal administration together in Tamil Nadu [32].
This is an important lesson for the proposed national rabies control programme in India which is currently structured around financing ABC-AR operations across selected cities [33]. Recent discussions have advocated parenteral vaccination of canines as a first step towards elimination of rabies [1], [9]. This would require a high level of coverage (>60%) costing 27% of the annual budget of the state department of animal husbandry, the likely implementing agency for such an intervention, and would need to be sustained continuously for multiple years or even decades.
Due to the challenges of achieving high vaccination coverage even among humans [34] and the high costs of existing animal interventions described above, the policymakers are unlikely to commit to a comprehensive rabies control programme yet. A more favourable case for rabies control among canines could be made by developing newer animal interventions that are not only efficacious but also affordable and effective, such as an inexpensive canine injectable contraceptive cum vaccine. In the absence of an intervention that promises long term sustainability, it is likely that ad hoc measures like post exposure vaccinations to economically productive animals continue.
Antibodies were not made universally available for human vaccines because of the costs involved. Our calculations show that the costs of a combined (antibody plus vaccine) rabies programme would be three times the costs of vaccine only intervention and is likely to cost an additional expenditure of $ 1.5 million (Rs. 82.5 million) annually.
Given the costs of different vaccine formulations in Tamil Nadu, choosing intradermal over intramuscular vaccine regimen is likely to result in annual savings of $ 13,000 (Rs 700,000) only. This relatively small amount should not be a deterrent to state public health programme managers in choosing a vaccine regimen that is more appropriate to the clinical setting and qualifications of their staff [24].
Rabies control efforts in Tamil Nadu seem a costly proposition as they are currently structured in the state. This would necessarily require high levels of technical, political and financial commitments before the government chooses to embark upon a long-term rabies control strategy. Given recent recognition of the need for a national rabies control programme in India by the National Centre for Disease Control [33] and the FAO/WHO/OIE tripartite statement on inclusion of rabies as an ‘entry point’ for demonstrating zoonoses control efforts at the global level [1], it is important that these discussions adopt a long term perspective and take local complexities into account before developing a national or a global rabies elimination strategy.
|
10.1371/journal.pbio.1000051 | Commitment to the Regulatory T Cell Lineage Requires CARMA1 in the Thymus but Not in the Periphery | Regulatory T (Treg) cells expressing forkhead box P3 (Foxp3) arise during thymic selection among thymocytes with modestly self-reactive T cell receptors. In vitro studies suggest Foxp3 can also be induced among peripheral CD4+ T cells in a cytokine dependent manner. Treg cells of thymic or peripheral origin may serve different functions in vivo, but both populations are phenotypically indistinguishable in wild-type mice. Here we show that mice with a Carma1 point mutation lack thymic CD4+Foxp3+ Treg cells and demonstrate a cell-intrinsic requirement for CARMA1 in thymic Foxp3 induction. However, peripheral Carma1-deficient Treg cells could be generated and expanded in vitro in response to the cytokines transforming growth factor beta (TGFβ) and interleukin-2 (IL-2). In vivo, a small peripheral Treg pool existed that was enriched at mucosal sites and could expand systemically after infection with mouse cytomegalovirus (MCMV). Our data provide genetic evidence for two distinct mechanisms controlling regulatory T cell lineage commitment. Furthermore, we show that peripheral Treg cells are a dynamic population that may expand to limit immunopathology or promote chronic infection.
| In mammals, CD4+ T cells are essential for controlling infections, but have the potential to attack host tissues as well, resulting in autoimmune disease. A subset of CD4+ T cells, regulatory T cells (Treg)—identified by the expression of the forkhead transcription factor Foxp3—serve to prevent immunopathology by dampening immune responses. These cells are unique among CD4+ T cell subsets, as only the Treg lineage can develop in both the thymus and periphery. Using a genetic approach, we identified a mutation in the gene Carma1, a key component of T and B cell signaling, which in mice distinguishes Treg cells derived from the periphery from thymic-derived regulatory T cells. The mutation caused an absence of thymic Treg cells. However, a small population of Treg cells was observed in the spleen, lymph nodes, and colon of Carma1-mutant mice that expanded after viral infection, suggesting that peripheral development of Treg cells could still occur. Indeed, Carma1-mutant CD4+ T cells could be converted into the Treg lineage in vitro. Our results demonstrate an organ-specific requirement for the CARMA1 signaling pathway that developing thymocytes need in order to become Treg cells, but that naïve CD4+ T cells can bypass in the periphery. This dichotomy suggests that Treg cells of thymic or peripheral origin may have different specificities or functions in vivo.
| Two major mechanisms enforce self-tolerance: negative selection in the thymus and dominant tolerance in the periphery. The importance of both mechanisms is underscored by the phenotypes of autoimmune regulator (Aire) knockout [1] and scurfy mice [2], which have defects in negative selection or dominant tolerance, respectively. Humans with orthologous mutations develop autoimmune polyendocrinopathy-candidiasis-ectodermal dystrophy (APECED) [3,4] or immune dysregulation, polyendocrinopathy, enteropathy, X-linked (IPEX) syndrome [5]. These mutations all result in systemic autoimmunity, though defects in dominant tolerance cause a more severe and fatal disease. The study of dominant tolerance accelerated after cloning of the scurfy locus, which identified forkhead box P3 (Foxp3) as an essential molecule [6]. Foxp3 is a transcription factor expressed predominantly in CD4+ T cells committed to the regulatory T cell (Treg) lineage [7]. Expression of Foxp3 programs T cells with suppressor function, allowing Treg cells to effect dominant tolerance [8].
The majority of Treg cells are derived from the thymus, although an unknown percentage of these cells may develop in the peripheral lymphoid organs. Thymic Treg lineage commitment occurs in CD4 single-positive (SP) thymocytes and requires intermediate affinity binding of the T cell receptor (TCR) [9], co-stimulation through CD80 and CD86 interactions with CD28 [10,11], and the cytokines TGFβ [12] and interleukin (IL)-2 or IL-15 signaling through the shared IL-2Rβ chain [13–16]. Peripheral commitment of naïve CD4+ T cells to the Treg lineage, modeled in vitro, requires exogenous TGFβ, in addition to TCR stimulation and concomitant IL-2 production, to induce Foxp3 expression and Treg function [17]. Foxp3 induction can be enhanced in vitro by inhibition of AKT-mediated signaling or transient TCR stimulation [18,19], and may be preferentially driven in vivo by retinoic acid made by macrophages and dendritic cells (DCs) residing in mucosal tissues [20]. The differences in signaling pathways used in the development of thymic versus peripherally induced Treg cells remain largely unexplored.
In this report, we describe the characterization and positional cloning of the king mutation. We identified the mutation by screening G3 mice, homozygous for germline mutations induced by N-ethyl-N-nitrosourea (ENU) [21], to detect defects in T cell effector function. In king homozygous mice, no thymic Treg cells were detected, but Foxp3 could be induced among peripheral CD4+ T cells in response to cytokines. Thus, the king mouse offered a model to explore differences in signaling pathways used for thymic versus peripheral Treg lineage commitment and to study peripheral Treg dynamics, which are normally obscured by the presence of thymic-derived Treg cells. Our studies provide genetic evidence of two pathways, operating in the thymus or periphery, that commit CD4+ T cells to the Treg lineage. We also show that viral infection can cause massive expansion of peripheral Treg cells, an event that may reduce immunopathology or contribute to persistent viral infections.
To identify genes with non-redundant roles in T cell development, priming or effector function, we designed a screen to detect defective cytotoxic CD8+ T cell (CTL) responses in mice immunized with ovalbumin (act-mOVA) [22]. Among 2,500 ENU-mutagenized G3 mice screened, we have thus far bred three non-responsive mutations to homozygosity. We termed one of these mutants king. While the primary screen used was an in vivo cytotoxicity assay [23], the mutation could be scored using an in vitro assay as well. To do so, we isolated T cells 7 d after immunization, at the peak of the CD8+ T cell response, and expanded antigen-specific CD8+ T cells in culture with SIINFEKL peptide. king CD8+ T cells did not undergo secondary expansion or produce interferon (IFN)γ after restimulation with peptide (Figure 1A). We hypothesized that a mutation affecting DC cross-priming of CD8+ T cells, T cell activation, or T cell proliferation could cause such a phenotype. To test DC function, we used FMS-like tyrosine kinase 3 (Flt3)-ligand to generate bone marrow-derived lymphoid DCs, a subset of DCs that efficiently cross-primes CD8+ T cells [22]. When lymphoid DCs were exposed to ovalbumin expressing apoptotic cells, king DCs primed ovalbumin-specific OT-I T cells as efficiently as wild-type DCs (Figure 1B). In addition, king DCs showed normal up-regulation of co-stimulatory molecules CD40, CD80, CD86, and major histocompatibility complex (MHC) class I and II after activation by Toll-like receptor ligands [24] or apoptotic cells [22] (unpublished data), suggesting that the mutation did not affect co-stimulation. These results indicated that the king mutation did not impair DC-mediated cross-priming of CD8+ T cells.
We next investigated T cell activation. king CD4+ and CD8+ T cells normally up-regulated CD69, but not CD25 (IL-2Rα) upon TCR activation (Figure 1C). As CD25 can be further up-regulated in response to IL-2, we measured IL-2 production by king CD4+ T cells activated by TCR ligation and found a lack thereof (Figure 1D). Since these data implied only a partial defect in T cell activation, we next assessed T cell proliferative capacity. king T cells failed to proliferate in response to TCR stimulation, although this could be partially rescued by exogenous IL-2 (Figure 1E). As IL-2 is required to maintain CD4+CD25+Foxp3+ Treg cells in the periphery [14], we assessed the development of these cells in king mice. CD4+Foxp3+ T cells were reduced by an order of magnitude in periphery (Figure 1F), but were absent in the thymus (Figure 1G), indicating that the king mutation blocked commitment of developing thymocytes to the Treg lineage. We also investigated the function of another population of self-reactive T cells that develop in the thymus, natural killer T (NKT) cells. 90 min after injection of the NKT cell-specific agonist alpha-galactosylceramide (αGalCer), elevated concentrations of IL-4 and IFNγ were measured in the serum of king mice, indicating that the mutation did not impair pan-T cell function (Figure 1H).
Unlike most other mutations that impair Treg development, king mice did not exhibit gross signs of autoimmunity. Even in a cohort of king mice monitored for over 9 mo, no detectable anti-chromatin auto-antibodies (Figure S1) were found in the serum, nor did the mice develop splenomegaly, lymphoproliferative disease, or signs of chronic inflammation—all aspects of autoimmunity normally controlled by Treg cells.
To find the causative mutation, we mapped the king phenotype by outcrossing the king stock (C57BL/6J background) to C3H/HeN mice, backcrossing to the king stock, and measuring the percentage of circulating CD4+ T cells expressing Foxp3 in the blood of F2 mice (Figure 2A). By analyzing 134 informative microsatellite markers dispersed throughout the genome on 39 meioses, we localized the king mutation to the distal region of Chromosome 5 with a peak logarithm of odds (Lod) score of 11.74 (Figure 2B). Further analysis of 268 meioses confined the mutation to a 1.03-Mb critical region, bounded by the markers D5Mit292 and D5Mit101. This region contained only six annotated genes (http://www.informatics.jax.org), and among these was Card11 (more commonly known as Carma1 [from CARD-MAGUK1]). We sequenced either genomic DNA or cDNA of all coding basepairs within the critical region and identified a single point mutation in Carma1 (Figure 2C), which resulted in an L525Q substitution. The mutation occurred in α-helix2, in the NORS (no regular secondary structure) domain, of the CARMA1 linker region (Figure 2D). In naïve T cells, the CARMA1 protein adopts a conformation in which the linker domain associates with the caspase recruitment domain (CARD). Upon T cell activation, protein kinase C (PKC)θ phosphorylation of residues in the linker domain reduces intramolecular affinity for the CARD domain. This liberates both the CARD and coiled-coil domains, allowing CARMA1 oligomerization and recruitment of B cell CLL/lymphoma 10 (BCL10) and mucosa associated lymphoid tissue lymphoma translocation gene 1 (MALT1) to the CARMA1 signaling module [25,26]. Following activation, degradation of BCL10 terminates CARMA1-dependent signaling [27]. CARMA1 has a similar function in B cells, downstream of PKCβ. No CARMA1 expression was detected by western analysis in the thymus, spleen, or lymph nodes of king homozygotes (Carma1k/k) (Figure 2E). Furthermore, CARMA1 was not detectable in CD4 or CD8 SP thymocyte lysates (Figure 2F). The L525Q mutation may have the effect of destabilizing the CARMA1-king protein or marking it for degradation in mature T and B cells.
Several other groups have generated targeted knockouts or hypomorphs of Carma1 [28–31]. Like these other mutant mice, Carma1k/k mice have reduced basal serum immunoglobulin levels (Figure S2A), fail to mount antigen-specific immunoglobulin (Ig)M and IgG responses after immunization with ovalbumin in Complete Freunds Adjuvant (Figure S2B), and exhibit impaired B cell proliferation (Figure S2C). Lymphocyte development was abnormal in Carma1k/k mice, as in other Carma1 mutants, including a deficiency in peritoneal B1 cells and skewed double-negative thymocyte populations (Figure S3A–S3D). Additionally, splenic natural killer (NK) cells, NKT cells, γδ T cells, memory CD4+ T cells, and mature B cells were reduced in both percentage and cell number (Figure S3E–S3L and unpublished data). With age, some Carma1k/k mice developed severe dermatitis, as reported in homozygotes for the unmodulated allele of Carma1 [30]. Collectively, these data suggest that the L525Q king mutation abolishes CARMA1 activity, and uncover an essential requirement for CARMA1 in thymic Treg development.
In addition to Foxp3, thymocytes committed to the Treg lineage also express CD25, glucocorticoid-induced tumor necrosis factor receptor (GITR), and cytotoxic T-lymphocyte antigen-4 (CTLA-4) [32]. To determine whether Carma1k/k thymocytes begin differentiation into the Treg lineage but fail to express Foxp3, a distal marker of Treg differentiation, we examined expression of these additional markers (Figure 3A). The lack of expression of these markers suggests CARMA1 acts early in Treg lineage commitment. An absence of thymic Treg cells in Carma1k/k mice could result from altered selection by thymic epithelial cells, a defect in the TCR and co-stimulatory signaling pathways, or a lack of signaling through the IL-2Rβ chain [13,15,16,33]. To understand the role of Carma1 in thymic Treg development, we generated reciprocal and mixed bone-marrow chimeric mice. Foxp3 expression was absent in Carma1k/k thymocytes that developed in a wild-type thymus (Figure 3B), but normal in wild-type thymocytes that developed in a Carma1k/k thymus (Figure 3C). Therefore, the Treg deficiency in Carma1k/k mice results not from an altered thymic environment, but rather from an intrinsic defect in hematopoietically derived precursors. When wild-type mice were reconstituted with Carma1k/k and wild-type bone marrow at a 4:1 ratio, 1:1 chimerism was achieved among lymphocytes. While wild-type thymocytes differentiated into the Treg lineage at normal frequencies, Carma1k/k thymocytes failed to develop into thymic Treg cells and expressed lower levels of CD25, GITR, and CTLA-4 among CD4 SP thymocytes (Figure 3D). As trans-acting IL-2Rγ chain cytokines produced by wild-type thymocytes did not rescue Foxp3 induction in Carma1k/k thymocytes, it is likely that impaired signaling downstream of the TCR or CD28 underlies the absence of the thymic Treg cells.
Pathways distinct from those involved in thymic development regulate commitment and homeostasis of peripheral Treg cells [14]. In wild-type mice reconstituted with Carma1k/k bone marrow, peripheral Treg cells were found at reduced frequencies, similar to those observed in Carma1k/k mice (Figure 3E). Conversely, in Carma1k/k mice reconstituted with wild-type bone marrow, peripheral CD4+Foxp3+ Treg cells occurred at frequencies similar to those observed in wild-type mice (Figure 3F). Therefore, the Carma1k/k environment can support Treg homeostasis and a cell-intrinsic defect in development causes the Treg deficiency observed in these mice. It has been proposed that IL-2 regulates Treg homeostasis and that Treg cells may function as an IL-2 “sink” [14,34]. However, in wild-type mice reconstituted with mixed wild-type and Carma1k/k bone marrow, the Carma1k/k Treg population did not expand in the periphery (Figure 3G). Interestingly, the wild-type Treg population expanded to comprise 20% of the wild-type CD4+ T cells, and 10% of the total CD4+ T cell pool. This suggests that a cell-extrinsic homeostatic mechanism regulates the size of the Treg compartment in the periphery of naïve wild-type mice.
While Carma1k/k mice lack thymic Treg cells, they do exhibit a peripheral Treg pool, though it is substantially smaller than that in wild-type mice. Peripheral expansion and conversion of CD4+Foxp3+ T cells can be modeled in vitro by culturing activated CD4+ T cells in the presence of the cytokine TGFβ [17]. When we activated Carma1k/k CD4+ T cells with plate-bound anti-CD3 and anti-CD28 in the presence of exogenous TGFβ, they did not proliferate or express Foxp3. However, the combination of exogenous IL-2 and TGFβ was sufficient to rescue both proliferation and Foxp3 induction (Figure 4A). Importantly, the percentage and number of undivided Carma1k/k CD4+Foxp3+ T cells increased, indicating that Foxp3 expression was induced from CD4+Foxp3− T cells. Additionally, the induced Carma1k/k Treg cells did not express stable CARMA1 protein (Figure S4).
To determine whether Carma1k/k CD4+ T cells were more or less prone than wild-type cells to express Foxp3, we preformed a dose-response analysis of Foxp3 induction by titrating anti-CD3, anti-CD28, and TGFβ concentrations. In the absence of CARMA1, the TCR-signaling threshold for Foxp3 induction was increased significantly (Figure 4B). Co-stimulatory signals were also required for efficient Foxp3 induction in wild-type and Carma1k/k CD4+ T cells (Figure 4C). CARMA1 deficiency did not alter the ability of CD4+ T cells to respond to limiting concentrations of TGFβ (Figure 4D). These data suggest that Foxp3 induction can occur without CARMA1, and reveal a partial role for CARMA1 in transmitting TCR-mediated signals for peripheral Foxp3 induction.
The absence of detectable thymic Treg cells and presence of a small peripheral Treg population in Carma1k/k mice suggests that induction of Treg cells can occur in the periphery of naïve mice without activation of the CARMA1 pathway. Also consistent with this conclusion, Carma1k/k mice had elevated numbers of CD4+Foxp3+ T cells in the lamina propria and mesenteric lymph nodes (Table 1). The lamina propria of the colon is a site where peripheral conversion of Treg cells may preferentially occur, and these lymphocytes drain to the mesenteric lymph nodes [20,35]. The CD4+Foxp3+ Treg cells in Carma1k/k mice showed normal expression of CD25, CTLA-4, and GITR, but interestingly expressed higher amounts of membrane-associated TGFβ in the mesenteric lymph nodes but not in the spleen (Figure 5A). In vitro induced Carma1k/k Treg cells showed a similar phenotype (Figure 5B) and released more soluble TGFβ than wild-type Treg cells in culture, with or without activation (Figure 5C). Together, these data suggest that peripheral induction of CD4+Foxp3+ Treg cells occurs in the absence of Carma1.
As IL-2 and TGFβ are sufficient to induce Foxp3 in Carma1k/k CD4+ T cells, we explored the possibility that these cytokines activated signaling pathways downstream of CARMA1, or induced Foxp3 expression via an alternative pathway. We first activated wild-type and Carma1k/k CD4+ T cells with the PKC activator phorbol myristate acetate (PMA) and ionomycin, without exogenous cytokines (Figure 6A). Upon activation, BCL10 phosphorylation, NF-κB inhibitor alpha (IκBα) degradation and phosphorylation of both Jun N-terminal kinase (JNK) isoforms occurred in wild-type T cells. In contrast, BCL10 was constitutively phosphorylated, amounts of IκBα were constantly elevated, and the JNKp54 isoform remained unphosphorylated in Carma1k/k T cells. Interestingly, decreased abundance and constitutive phosphorylation of BCL10 was observed in total Carma1k/k lymph node T cells, but not thymocytes (Figure 6B). Normally, BCL10 is recruited to CARMA1 after activation. After assembly of the CARMA1 signaling complex, phosphorylation marks BCL10 for ubiquitination and degradation [27]. The lack of CARMA1 protein (Figure 2D) and constitutive degradation of BCL10 in Carma1k/k T cells indicates that the CARMA1 signaling complex cannot be assembled; this is also reflected by the elevated amounts of TGFβ-activated kinase 1 (TAK1) in resting and activated Carma1k/k T cells (Figure 6A), likely due to reduced protein turnover.
Other groups have reported that Carma1-deficient T cells exhibit a lack of nuclear factor κB (NF-κB) nuclear translocation [28–31] and JNK2 phosphorylation [36] after TCR stimulation. While the CARMA1 signaling complex was inoperative in Carma1k/k mice, it was possible that exogenous IL-2 and TGFβ retained the ability to activate NF-κB or JNK. However, when we activated wild-type and Carma1k/k T cells with PMA and ionomycin in the presence of these cytokines, levels of IκBα degradation and JNK phosphorylation remained unchanged up to 60 min after activation (unpublished data). Similar results were obtained when T cells were activated by anti-CD3 and anti-CD28 antibodies.
We next investigated whether other cytokines could drive Foxp3 expression. To determine if trans-acting cytokines produced by wild-type CD4+ T cells could drive Foxp3 induction, we co-cultured activated wild-type and Carma1k/k CD4+ T cells in the presence of TGFβ. The presence of activated wild-type cells allowed proliferation and Foxp3 induction in Carma1k/k CD4+ T cells. Co-culture in the presence of IL-2 neutralizing antibody abolished proliferation of Carma1k/k CD4+ T cells and Foxp3 induction in both wild-type and Carma1k/k CD4+ T cells (Figure 7A). However, other IL-2Rγ chain cytokines can also substitute for exogenous IL-2 in Foxp3 induction [37]. To test whether this could occur in the absence of CARMA1, we activated Carma1k/k CD4+ T cells in the presence of TGFβ and IL-4 (an IL-2Rγ chain cytokine). Exogenous IL-4 induced proliferation and Foxp3 induction, although, as expected, it was less potent than exogenous IL-2 [38]. However, neutralizing IL-2 antibody abrogated this effect (Figure 7B), indicating that other IL-2Rγ chain cytokines can drive IL-2 production and T cell proliferation independently of TCR-mediated CARMA1 activation. It has also been reported that CpG DNA or associated proinflammatory molecules can act, in vitro, directly on T cells to restore TCR-mediated proliferation and induce CD4+ T cell polarization in cells from Pkcθ−/− mice [39]. However, neither proliferation nor Foxp3 induction occurred in Carma1k/k T cells cultured with TGFβ and CpG, TNFα, IFNα, or IFNγ (Figure 7B).
It remained unclear whether CARMA1 was required for Treg suppressor function. In addition to Foxp3, in vitro generated Treg cells express CD25, CTLA-4, and GITR at high levels, similar to Treg cells found in vivo. Carma1k/k Treg cells generated in vitro express normal levels of these markers (Figure 7C). Treg cells have the ability to suppress T cell proliferation by a cell-contact dependent mechanism in vitro [34]. Both wild-type and Carma1k/k induced Treg cells, generated in vitro with IL-2 and TGFβ, suppressed the proliferation of wild-type CD4+ T cells in a co-culture assay (Figure 7D). The difference in observed suppression at 1:4 and 1:8 dilutions likely reflects the proliferation defect in Carma1k/k Treg cells. These results suggest Carma1 is neither required for TGFβ-mediated induction of the Treg phenotype, nor for suppressor function.
Peripheral Treg cells comprise a small percentage of Carma1k/k CD4+ T cells in the steady-state. It has been suggested that peripheral Treg cells may expand during conditions of lymphopenia [9], at the site of tumors [40], or in response to pathogens [41]. The absence of thymic-derived Treg cells in Carma1k/k mice provides a model to study the dynamics of peripheral Treg cells during infection. To do this, we infected Carma1k/k mice with a pathogen that establishes persistent infection in mice—mouse cytomegalovirus (MCMV). Carma1k/k mice mounted T-dependent B cell responses that were reduced compared to the wild-type response (Figure 8A), but sufficient to allow survival without any signs of virus-induced immunopathology or detectable virus in the spleen 14 d after infection. Yet, at the peak of the effector CD4+ T cell response, 8 d after infection, no Treg expansion was observed in the spleen (Figure 8B). However, splenic CD4+Foxp3+ Treg cells expanded by an order of magnitude in Carma1k/k mice 14 d after MCMV infection (Figure 8C). MCMV establishes persistent infection in the salivary glands, and T cells drain from the salivary glands to the submandibular lymph nodes. Here, Treg expansion was also observed in Carma1k/k mice (Figure 8D). No Foxp3 expression was detected in the thymus 8 or 14 d after infection, suggesting that Treg expansion resulted from either de novo induction of Foxp3 or expansion of the pre-existing peripheral Treg pool. Additionally, MCMV infected Carma1k/k mice did not develop signs of autoimmune or lymphoproliferative disease when monitored for 80 d after infection.
Accumulating evidence suggests that diverse stimuli can drive proliferation or induction of CD4+Foxp3+ Treg cells in the periphery [20,40–43]. We have shown that cytokines or MCMV infection can drive Foxp3 induction and peripheral Treg proliferation in the absence of CARMA1, but found different requirements for thymic Treg development. Contrary to expectations that the Treg deficiency observed in mice lacking components of the CARMA1 pathway reflected an inability to produce IL-2 [33], cytokines produced by wild-type thymocytes or thymic epithelial cells did not rescue thymic Foxp3 induction in Carma1k/k thymocytes in mixed bone marrow chimeric mice. This may reflect the absence of CD4+CD25+ thymocytes capable of responding to IL-2 and inducing Foxp3 [44] among Carma1k/k thymocytes. Thus, unlike Foxp3 induction in peripheral CD4+ T cells, which can be driven by cytokines without activation of CARMA1, developing thymocytes require activation of a cell-intrinsic, CARMA1-dependent signaling pathway(s) that likely includes the transcription factor NF-κB. CARMA1-mediated NF-κB activation may act as a survival factor in thymic Treg development, preventing apoptosis of thymocytes with certain TCRs or pre-TCRs [28,45] that are destined to become CD4+Foxp3+ Treg cells.
Given the robust T cell expansion observed in Carma1k/k mice during MCMV infection, it is perhaps surprising that they do not develop spontaneous autoimmune disease as a consequence of thymic Treg deficiency. Likewise, deletion of other genes in the Carma1 pathway—Pkcθ [46], Bcl10 [46], Tak1 [47,48], Ikkβ [49]—impairs Treg development, but does not result in spontaneous autoimmunity (see Figure S5). Mice with a Ick-driven CD4-specific deletion of Tak1 develop colitis with age, although escaped CD4+ T cells retaining intact Tak1 might initiate disease in this model [47]. Additionally, hypomorphic mutations in regulators of TCR-mediated PKCθ activation, linker for T cell activation (Lat) [50], and zeta-chain-associated protein kinase 70 (Zap70) [51], also block Treg development. In Carma1k/k mice, T cells, including any with potentially auto-reactive TCRs, are normally quiescent, but can become activated during the “cytokine storm” of infection. The small Carma1k/k peripheral Treg pool may also contribute to dampening T cell responses directed towards self-antigens or commensal flora and may release more of the regulatory cytokine TGFβ than wild-type Treg cells. Additionally, the impaired B cell function in Carma1-deficient mice may prevent amplification of auto-reactive immune responses. We also demonstrated that NKT cell, but not conventional T cell cytokine production occurred normally without functional CARMA1, suggesting that in the absence of thymic-derived Treg cells, Carma1k/k NKT cells do not drive spontaneous autoimmune disease.
The development of the hematopoietic system and lymphocyte activation defects were similar in Carma1k/k mice and Carma1-knockout mice. Therefore, it was surprising to find reduced protein levels and constitutive phosphorylation of BCL10 in Carma1k/k CD4+ T cells. In CARMA1-deficient JPM50.6 cells, BCL10 expression is normal [52]. The normal expression of BCL10 in Carma1k/k thymocytes, but not lymph node cells suggests BCL10 degradation may be confined to mature T and B cells. As the king mutation occurred in a region of CARMA1 predicted to regulate accessibility of the CARD domain, we propose a model in which BCL10 is constitutively recruited to the CARMA1-king protein in mature T and B cells. This interaction is not sufficient to assemble the full CARMA1-BCL10-MALT1 signaling complex or to activate NF-κB, but does allow for BCL10 phosphorylation and the subsequent degradation of BCL10 [27] and the CARMA1-king protein. As a result, the CARMA1-BCL10-MALT1 complex cannot be assembled upon TCR stimulation.
The generation of peripheral Treg cells requires TCR signals in addition to TGFβ and IL-2 in vitro. Our data suggest that peripheral Foxp3 induction may not require CARMA1-mediated activation of NF-κB, as indicated by the lack of IκBα degradation or JNK phosphorylation after activation. When IκBα degradation was induced by culturing CD4+ T cells with TNFα [29], Carma1k/k cells still did not proliferate, nor did they up-regulate Foxp3 in the presence of TGFβ. Similar to our observations for Carma1k/k mice, exogenous IL-2 can also rescue the TCR-mediated proliferation defect in mice deficient for the NF-κB family members p50 and c-Rel [53]. T cells from these mice also cannot produce IL-2, and peripheral CD4+CD25+ cells are reduced 5-fold. It will be of interest to determine if thymic Treg cells develop in these NF-κB deficient mice and whether peripheral CD4+ T cells from these mice induce Foxp3 after exposure to IL-2 and TGFβ.
Recently, two groups reported TGFβ-independent Foxp3 induction in vitro when CD4+ T cells were activated in the presence of chemical or genetic inhibitors of the phosphatidylinositol 3-kinase (PI3K)-AKT-mammalian target of rapamycin (mTOR) signaling axis [18,19]. Weak or transient TCR stimulation, without co-stimulation, is also postulated to favor peripheral induction of Foxp3 in vivo [19,43]. However, Carma1k/k CD4+ T cells were less prone than wild-type CD4+ T cells to induce Foxp3 when given weak TCR stimulation. Therefore, our data suggest that unlike the AKT signaling axis, blockade of CARMA1-dependent signaling does not favor peripheral Foxp3 induction.
In vivo, we have shown that factors produced during MCMV infection can stimulate peripheral Treg expansion in Carma1k/k mice. Our in vitro data suggest that this expansion may require IL-2Rγ chain cytokines, such as IL-4, to drive IL-2 production. Expansion of Treg cells may contribute to the transient immunosuppression that can follow viral infection [54], and our data suggest these cells may arise peripherally. Such a response could protect the host against cross-reactive anti-viral T cells that have the potential to precipitate autoimmune disease. An alternative but not mutually exclusive hypothesis holds that MCMV may manipulate the host immune response to expand virus-specific Treg cells. Such a strategy has been well documented for the parasite Leishmania major [42], and may be utilized by other viruses, for example Friend leukaemia virus [41]. If these Treg cells localized to viral reservoirs, they could facilitate persistent viral infections.
The existence of two pathways for Foxp3 induction allows for a total Treg pool with two potential specificities [32]. Thymic Treg cells might primarily express TCRs with intermediate self-affinity [9], as Foxp3 induction would require TCR signals strong enough to activate CARMA1 without causing clonal deletion. Tonic signaling through the TCR along with consumption of cytokines could allow expansion of thymic Treg cells to fill the Treg niche. The generation of peripherally induced Treg cells might be driven by cytokines produced by the innate immune system at mucosal sites or during infection. In Carma1−/− mice, myeloid cells make normal amounts of IL-2, which may support the peripheral Treg population we observe in Carma1k/k mice [55]. These induced Treg cells could then prevent T cell responses to commensal flora and dampen potentially dangerous responses to pathogens [41] or innocuous antigens [43].
We obtained C57BL/6J and Jnk2−/− mice from The Jackson Laboratories. We generated king mice at TSRI using ENU mutagenesis [21]. They have been made available through the Mutant Mouse Regional Resource Centers (MMRRC:030114-UCD).
The MCMV smith strain was isolated from the salivary glands of 3-wk-old infected BALB/c mice. 1 × 105 PFU of MCMV was injected IP per mouse.
The following antibodies were used in this study for flow cytometry: CD4-APC (L3T4), CD25-FITC (PC61.5), CD45.1-FITC (A20), CD45.2-APC (104), CD69-APC (H1.2F3), Foxp3-PE (FJK-16s), GITR-FITC (DTA-1), IFNγ-APC (XMG1.2) (eBioscience); CD4-FITC (L3T4), CTLA-4-FITC (1B8) (Southern Biotech). Intracellular staining was performed for analysis of CTLA-4, Foxp3, and IFNγ. These antibodies were used for western blotting: BCL10 (Santa Cruz); CARD11, IκBα, pJNK, JNK, TAK1 (Cell Signaling). Purified CD3ε (145–2C11) and CD28 (37.51) antibodies (eBioscience) were used at indicated concentrations for T cell activation. Carboxyl fluorescent succinimidyl ester (CFSE) labeling was performed by incubating MACS (Miltenyi Biotech) purified CD4+ T cells, CD8+ OT-I T cells, or splenic B cells in 5 μM CFSE with 0.1% fetal calf serum in PBS for 10 min.
To assess the CD8+ T cell response, we immunized mice with 1 × 107 γ-irradiated (1,500 rad) act-mOVA splenocytes. 7 d later, 5 × 106 splenocytes were isolated and cultured with 10 nM SIINFEKL peptide in IMDM media supplemented with 10% FCS (Atlanta Bio). 5 d later, cells were restimulated with 100 nM SIINFEKL peptide in the presence of brefeldin A. Production of IFN-γ was assessed by intracellular cytokine staining.
To test cross-priming of CD8+ T cells, lymphoid DCs were generated by culturing 1 × 107 bone marrow cells with 200 ng/ml human Flt3-ligand (Peprotech) in supplemented IMDM media for 8 d. 1 × 105 DCs were then co-cultured with 2 × 105 γ-irradiated (1,500 rad) Kb−/−; act-mOVA splenocytes and 1 × 105 MACS purified CFSE labeled CD8+ OT-I T cells. CFSE dilution was assessed 3 d later by flow cytometry.
NKT cells were activated by injecting mice IV with 2 μg of αGalCer. Serum was collected 90 min later, and cytokine concentration was measured by ELISA (eBioscience).
MACS purified splenic CD4+ T cells were used to test T cell activation. To measure up-regulation of activation markers by flow cytometry, cells were activated for 24 h, and then stained for intra-cellular CD25 and CD69. IL-2 production was measured by culturing 2 × 105 cells/ml under activating conditions in supplemented IMDM media. Supernatant was harvested at 18 h and IL-2 was measured by ELISA (eBioscience). T cell proliferation assays entailed activating CFSE-labeled CD4+ T cells using 10 μg/ml CD3ε and 1 μg/ml CD28 plate-bound antibodies with or without 100 U/ml IL-2. CFSE dilution was measured after 4 d by flow cytometry.
Peripheral Treg cells were generated in 24-well plates by culturing 3 × 105 CFSE-labeled CD4+ T cells in plates coated with CD3ε and CD28 antibodies with or without IL-2 (R & D Systems) and/or TGFβ (R & D Systems) in supplemented IMDM media. CFSE dilution and intracellular expression of Foxp3 was measured after 4 d of culture. To examine the effect of other cytokines, 2 × 105 CFSE-labeled C57BL/6-CD45.1+ and king-CD45.2+ CD4+ T cells were co-cultured in plates coated with 2 μg/ml CD3 and CD28 antibodies in supplemented IMDM media with 5 ng/ml TGFβ and with or without 10 μg/ml neutralizing IL-2 antibody (JES6–1) (eBioscience). Additionally, CFSE dilution and Foxp3 induction was assessed in CFSE-labeled CD4+ T cells activated by 2 μg/ml CD3 and CD28 plate-bound antibodies in supplemented IMDM media with 5 ng/ml TGFβ and: 100 nM CpG oligonucleotides (SIGMA), 10 ng/ml IL-4, 10 ng/ml TNFα, 100 U/ml IFNα, or 100 U/ml IFNγ (all cytokines, R & D systems).
For western analysis of T cell signaling, MACS purified CD4+ T cells were activated with 50 ng/ml PMA and 500 ng/ml ionomycin (SIGMA) for the indicated time. Cells were lysed in a non-ionic buffer with protease and phosphatase inhibitors for 15 min on ice, then suspended at a 1:1 ratio in Lamelli sample buffer.
Recipient mice were γ-irradiated (2 × 500 rads) and injected with 1 × 108 donor bone marrow cells. 10 wk later, lymphoid tissues were harvested, homogenized, stained, and analyzed by flow cytometry.
The Treg suppressor assay was performed under conditions previously described [34,37]. Briefly, MACS purified CD4+ T cells were cultured in plates coated with 2 μg/ml CD3 and CD28 antibodies, in 100 U/ml IL-2 (R & D Systems) and 5 ng/ml TGFβ (R & D Systems) in supplemented IMDM media. After 4 d, CD4+ T cells were again MACS purified. Foxp3 induction was assessed by flow cytometry, and in all experiments at least 90% of CD4+ T cells expressed Foxp3. Induced Treg cells were harvested and co-cultured at indicated ratios with 5 × 104 MACS purified CFSE-labeled CD8+ T cells. Also included were 1 × 105 T cell-depleted, γ-irradiated (3,000 rad) splenocytes as bystander cells and 0.5 μg/ml soluble CD3ε antibody. CFSE dilution was assessed by flow cytometry after 3 d of co-culture.
To measure the concentration of MCMV specific Ig in the serum, 96-well plates were coated with virus, then blocked in 5% milk. Serum samples were diluted 1:200, then serially diluted threefold. Anti-Ig HRP conjugated antibodies were used for detection.
The Entrez GeneID numbers (http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene) for genes and gene products mentioned in the text are: Aire (11634), Akt (11651), Bcl10 (12042), Card11 (108723), Cd25 (16184), Cd28 (12487), Cd3e (12501), Cd40 (21939), Cd69 (12515), Cd80 (12519), Cd86 (12524), c-Rel (19696), Ctla-4 (12477), Flt-3 (14255), Foxp3 (20371), Gitr (21936), Ifnα (15960), Ifnγ (15978), Ikbα (18035), Ikkβ (16150), Il-15 (16168), Il-2 (16183), Il-2rβ (16185), Il-2Rγ (16186), Il-4 (16189), Jnk2 (26420), Lat (16797), Malt1 (240354), mTor (56717), p50 (18033), Pi3k (18708), Pkcθ (18761), Tak1 (26409), Tgfβ (21803), Tnfα (21926), and Zap70 (22637).
|
10.1371/journal.pcbi.1005643 | Deciphering the regulation of P2X4 receptor channel gating by ivermectin using Markov models | The P2X4 receptor (P2X4R) is a member of a family of purinergic channels activated by extracellular ATP through three orthosteric binding sites and allosterically regulated by ivermectin (IVM), a broad-spectrum antiparasitic agent. Treatment with IVM increases the efficacy of ATP to activate P2X4R, slows both receptor desensitization during sustained ATP application and receptor deactivation after ATP washout, and makes the receptor pore permeable to NMDG+, a large organic cation. Previously, we developed a Markov model based on the presence of one IVM binding site, which described some effects of IVM on rat P2X4R. Here we present two novel models, both with three IVM binding sites. The simpler one-layer model can reproduce many of the observed time series of evoked currents, but does not capture well the short time scales of activation, desensitization, and deactivation. A more complex two-layer model can reproduce the transient changes in desensitization observed upon IVM application, the significant increase in ATP-induced current amplitudes at low IVM concentrations, and the modest increase in the unitary conductance. In addition, the two-layer model suggests that this receptor can exist in a deeply inactivated state, not responsive to ATP, and that its desensitization rate can be altered by each of the three IVM binding sites. In summary, this study provides a detailed analysis of P2X4R kinetics and elucidates the orthosteric and allosteric mechanisms regulating its channel gating.
| Ligand-gated ion channels play a crucial role in controlling many physiological and pathophysiological processes. Deciphering the gating kinetics of these channels is thus fundamental to understanding how these processes work. ATP-gated purinergic P2X receptors (P2XRs) are prototypic examples of such channels. They are ubiquitously expressed and play roles in numerous cellular processes, including neurotransmission, inflammation, and chronic pain. Seven P2X subunits, named P2X1 through P2X7, and several splice forms of these subunits have been identified in mammal. The receptors are organized as homo- or heterotrimers, each possessing three ATP-binding sites that, when occupied, lead to receptor activation and channel opening. The P2XRs are non-selective cation channels and the gating properties differ between the various receptors. Previously, we have used biophysical and mathematical modeling approaches to decipher the kinetics of homomeric P2X2aR, P2X2bR, P2X4R, and P2X7R. Here we extended our work on P2X4R gating. We developed two mathematical models that could capture the various patterns of ionic currents recorded experimentally and explain the particularly complex kinetics of the receptor during orthosteric activation and allosteric modulation. This was achieved by designing a computationally efficient, inference-based fitting algorithm that allowed for parameter optimization and model comparisons.
| Purinergic P2X receptors (P2XRs) are a family of ligand-gated non-selective cation channels that are activated by extracellular adenosine 5'-triphosphate (ATP). In mammals, there are seven distinct subunits of this family of proteins, labeled P2X1-7. Each subunit contains intracellular N- and C-termini connected to the first and second transmembrane (TM) domains, respectively, followed by a large extracellular loop commonly referred to as the ectodomain. It is also well established that P2X subunits aggregate to form functional trimers [1–5]; receptors may be composed of either one type of subunit (homotrimer) or a mixture of more than one type of subunit (heterotrimer) [6]. When coordinated in a trimer, the interfaces between adjacent ectodomains form three binding pockets for ATP [7]. These ectodomains also form fenestrations which are lined by negatively charged amino acids that attract cations, and the cation selectivity of these channels is determined by the selectivity filter localized in the TM2 domain [8].
The binding of two or three ATP molecules to the extracellular binding sites induces conformational changes in the ectodomain and subsequently the TM domains, causing the channel opening. The gating of P2XR by ATP and other orthosteric agonists can be broken down into three distinguishable phases, activation, desensitization, and deactivation, defined by their ionic current kinetics in whole-cell recordings [9, 10]. Activation is a rapid phase of channel opening that corresponds to increasing inward current subsequent to agonist application. This is usually followed by the desensitization phase, a decay of current amplitude in the maintained presence of an agonist, with an onset that is slower than that of activation. After agonist removal from the medium, a relatively rapid decrease in current amplitude, referred to as the deactivation phase, is observed. Receptors differ in both their sensitivity to agonists and the kinetics of the phases (with desensitization being transient and controversial in P2X7Rs) [9, 11].
It has also been suggested that P2X2R and P2X7R are capable of exhibiting another phase in their gating, termed dilation, when the receptor pore was thought to progressively enlarge during sustained ATP application. Two observations were used as evidence for pore dilation: the ability of these receptors to become permeable to N-methyl-D-glucamine (NMDG+), a large organic cation (~7.3 Å in mean diameter), and the change in reversal potential (Erev) during a ramp protocol, when cells were bathed in a medium containing only NMDG+ with pipette containing only NaCl [12]. Recent investigations, however, have shown that changes in Erev during prolonged channel activation of P2X2R do not reflect pore dilation, but rather time-dependent alterations in the concentration of intracellular ions, specifically washout of intracellular Na+ and gain of NMDG+ through the initially opened large pore of P2XR [13, 14] and that permeation of NMDG+ can also occur without pore dilation [15].
In addition to orthosteric regulation, P2XRs also exhibit allosteric regulation [9]), as is evident from the action of ivermectin (IVM) on P2X4R channels. Extracellularly applied IVM increases current amplitude at low concentrations, and increases the sensitivity of receptors to ATP and partial agonists at higher concentrations. IVM also decreases the extent of desensitization in the continuous presence of agonist and prolongs deactivation of the receptor after the removal of agonists [16–18]. Furthermore, P2X4R is not substantially permeable to NMDG+ natively [16, 19], but displays a shift in Erev in the presence of IVM, suggesting that the channel pore is permeable to NMDG+ in IVM-treated cells [20].
The action of IVM on P2X4R gating is also time-dependent; i.e. the cells must be exposed to IVM for at least 30 s, in the absence of ATP, to alter the P2X4R gating (compared to ms for orthosteric activation) [20]. The onset of IVM’s potentiating effect on P2X4R current amplitude is faster than the effects of IVM on deactivation kinetics [17, 21]. Consequently it has been postulated that the two distinct effects of IVM are due to binding at two distinct sites [17]. Experiments with chimeric receptors containing domains from IVM-sensitive P2X4R and the IVM-insensitive P2X2R have provided evidence that TM domains play a critical role in this allosteric modulation by IVM [18]. The location of the IVM binding site has not yet been addressed in the context of the recent crystal structures of a zfP2X4R [2, 3]. However, IVM apparently inserts between pairs of neighboring subunits of the P2X4R channel in the membrane and interferes with the molecular rearrangement in the TM domains involved in channel gating, similarly to glutamate-gated chloride channels crystalized with IVMs [22]. Accordingly, there should be three potential binding sites for IVM in the P2X4R, as there are three clefts between subunits. Such a topography of IVM binding sites provides rationale as to why receptors (not previously stimulated orthosterically) must be exposed to IVM for a prolonged period for it to be effective. Subsequently, we have used the term priming to describe the time and concentration dependence of IVM to occupy its binding sites and the resulting development of its varied allosteric effects.
In recent years, mathematical modeling has begun to shed light on many aspects of P2XRs and to guide experimental designs to arrive at a more complete understanding of channel gating [19, 20, 23–25]. Biophysically detailed Markov models that describe individual orthosteric binding sites and their allosteric modulation, have been very successful in deciphering the kinetics of P2X homotrimers and succinctly explaining many phenomena [19, 20, 24–26]. They consider important biophysical details such as the conformational states of individual binding sites and other structural components of the receptor. One of these models for P2X4R is a simple Markov model that takes into account the sequential binding of ATP to its three subunits and assumes that IVM causes receptor sensitization upon the binding of three ATP molecules, that all ATP unbinding rates are decreasing functions of IVM concentration, and that IVM induces a change in ion selectivity caused by the assumed pore dilation [20]. In this model, the three allosteric effects of IVM on P2X4Rs are induced by a single IVM-dependent transition that allowed for generating the shift in Erev during the ramp protocol. However, the published model is unable to account for effects of pre-treatment with IVM before ATP application. The model also predicts a large (> 150%) increase in the unitary (single-channel) conductance of the receptor, in contrast to experimental evidence [17] indicating that there is at most 20% increase in unitary conductance.
To satisfy these constraints, we developed two substantially larger models that not only fit the data more closely in more experimental circumstances but offer better insights into how the kinetics of ATP and IVM sequential binding to P2X4R affect P2X4R activation, desensitization, and deactivation. They also illustrate how changes in ion selectivity of these receptors are manifested, as well as predict the previously unappreciated existence of receptor states (including the deeply inactivated and primed states) that are not directly observable in the experimental current recordings.
When HEK293 cells expressing rat P2X4R were bathed in Ca2+-containing medium, where Na+ was substituted by NMDG+ and a voltage ramp from −80 mV to +80 mV was applied to the cell, Erev was not found to change during sustained applications of 100 μM concentrations of ATP (S1 Fig, left). However, pretreatment with 3 μM IVM for 60 s caused a positive shift in Erev during sustained applications of 100 μM ATP (S1 Fig, right). This is consistent with our earlier work [20] and the finding of others that IVM potentiates ATP-induced responses and increases permeability for NMDG+, but cannot activate P2X4R channels on its own [16]. Because strategies that rely on changes in Erev to provide evidence for large pore formation during sustained stimulation with agonist were questioned [13, 15], we examined currents induced by ATP in Ca2+-free/NMDG+-containing medium. S2 Fig shows that in the absence of Ca2+, a 40-s application of ATP (100 μM) at −60 mV in bi-ionic NMDG+ out/Na+ in solution (where the reversal potential of ATP-induced current is about -70 mV [27]) evoked only outward Na+ current, whereas in the presence of IVM, outward Na+ current was followed by inward NMDG+ current. These experiments do not argue against findings with P2X2R presented recently [13] but provide evidence for the existence of two conductive pore states of P2X4R. These pore states, termed open1 and open2, differ in their selectivity for organic cations (a consequence of altered relative permeability PNMDG/PNa), and the priming of receptors by IVM is needed to switch from one conductive state to another.
We will demonstrate later that the shift in Erev does not require an increase in unitary conductance associated with the open2 state(s), but rather depends on the selectivity associated with Na+ and NMDG+, as suggested by the Goldman-Hodgkin-Katz equation,
Vrev=RTFlnPNa[Na+]out + PNMDG[NMDG+]outPNa[Na+]in + PNMDG[NMDG+]in,
where R is the gas constant, T is the absolute temperature, F is Faraday’s constant, PNa(PNMDG) is Na+ (NMDG+) permeability, and [Na+]out ([Na+]in) is Na+ concentration outside (inside) the cell, whereas [NMDG+]out ([NMDG+]in) is NMDG+ concentration outside (inside) the cell. Because the experimentally observed Erev shift is independent of the increase in unitary conductance, the term open2 state will be used to refer to both the (small) increase in unitary conductance and the (large) change in ion selectivity of the P2X4R pore.
The previous paragraph proposed that P2X4R opens with the open2 pore state(s) in the presence of IVM, which may have an increase in unitary conductance of as much as 20% [17]. At the same time, the ramp protocol shows a decrease in the slopes of the I-V curves (S1 Fig). To enforce such an outcome in any potential model of P2X4R with IVM-dependent allosteric transitions between open states, we require the rate of increase in the probability of open states due to allostery (open1 → open2) to be slower than the rate of decrease in the probability of open states (open1 → desensitized). We propose that the increased conductance of the open2 state(s) of the receptor pores is masked by desensitization in a time-dependent manner, similar to our previous finding with P2X2R [19]. We are thus led to assume that the probability of finding open receptors on the cell membrane, P(open1), is a strictly decreasing function of time (P˙(open1) < 0). On the other hand, we expect that the probability of finding a receptor whose pore is in the open2 state(s), P(open2) to be an increasing function of time (P˙(open2) > 0).
Without specifying a Markov model to describe P2X4R kinetics, we may consider a generic equation for current production in these receptors, capable of distinguishing open1 and open2 states based on their conductances and reversal potentials established after washout of intracellular Na+ and gain of NMDG+. According to the description above, we can write the equation for current as
I=g1P(open1)(V−E1)+g2P(open2)(V−E2)
(1)
where g1 is the maximum conductance of the open1 state(s), g2 (> g1) is the maximum conductance of open2 state(s), and E1 and E2 are the reversal potentials associated with the open1 and open2 states, respectively. The current equation can be rewritten in a standard form to isolate the total conductance and reversal potential of the cell, as follows
I=gtot(V−Etot),
(2)
where gtot and Etot are the total conductance and reversal potential of the cell, respectively. By equating Eqs (1) and (2), we obtain
gtot=g1P(open1)+g2P(open2).
The requirement for the slope of the I-V curves to decrease during the ramp protocol can be met if the total conductance of the receptor population decreases over time, i.e., g˙tot<0. Taking the time derivative of gtot and rearranging the terms, we obtain
g˙tot=g1(P˙(open1)+g2g1P˙(open2)),
which is strictly negative if we impose the condition
−P˙(open1)>g2g1P˙(open2).
(3)
It follows that
|P˙(open1)|>g2g1P˙(open2).
This result implies that the total conductance of the cell will decrease if the fraction of open receptors decreases more rapidly than the ratio of the open1-to-open2 maximum conductances times the rate of increase of the open2 state(s). Thus, in order to capture the decrease in the slopes of the I-V curves in any model development, we have to increase the rate of desensitization of the open states, reduce the rate of increase of open2 state(s) or decrease the ratio between the open2 and open1 conductances.
As a first approximation, we can attribute the decrease in the fraction of open states to two processes, desensitization and priming of receptors, related by the equation
−P˙(open1)=P˙(open2)+δ
(4)
where δ is the rate of change of open receptors due to desensitization. Furthermore, letting g2 satisfy
g2=g1(1+f)
where f is the fractional increase in unitary conductance, we can substitute this expression into Eq (3) to obtain
δ>fP˙(open2).
(5)
Inequality Eq (5) represents a new condition that can be used to produce the decrease in total conductance seen in the ramp protocol. For example, if we consider the experimentally observed value of ~0.2 for f in human P2X4R, then the rate of desensitization only needs to be one fifth the rate of the IVM-induced unitary conductance increase in order to mask its effect on the slopes of the I-V curves.
The desensitization rate of naïve receptors is well characterized by the current recordings produced during prolonged application of ATP, which can be used to constrain δ as a fixed parameter. The IVM-induced increase in unitary conductance has not been determined for rat P2X4R, nor its time-course. We determine these in the Markov model (discussed below) by fitting the total current, imposing Inequality Eq (5) to ensure that the total conductance of the cell decreases during the ramp protocol (due to desensitization).
We next consider the effects of IVM on desensitization and deactivation. During the pulse protocol (Fig 1A), where cells were repeatedly stimulated by 1 μM ATP for 2 s twice per min in the absence (black trace) and presence of 1 μM IVM (colored traces), we observed an initial increase in desensitization rate of the receptor (blue trace in Fig 1A), followed by a gradual decrease in desensitization rate at each subsequent ATP pulse (see the Methods section for quantification procedure). By the fifth pulse (green trace in Fig 1A), the desensitization rate reverted back to a value comparable to that seen in the absence of IVM (black trace in Fig 1A).
To assess if this phenomenon occurs consistently, we evaluated the statistical significance of the transient increase in desensitization rate. To quantify the amount of desensitization seen in the recordings, we used linear fitting to measure the rate of receptor desensitization normalized by the current amplitude of each pulse. As shown in Fig 1B, we did not see a significant change in the rate of desensitization at each ATP pulse in the absence of IVM (filled circles) (n = 7), suggesting that the desensitization proccess of receptors is far from equilibrium. However, in the presence of IVM, the first two ATP pulses following IVM application (indicated by the small arrows) exhibited a significant (p < 0.005 and p < 0.05; n = 7) increase in desensitization rates. The desensitization rate of current recordings in subsequent ATP pulses gradually drifted back to its original value before IVM was applied, further suggesting that the open state, exhibiting an increased desensitization rate, has reached an equilibrium with its corresponding desensitized state. At higher IVM concentrations, however, these transient effects were not observed, but an increase in non-desensitized current amplitude was found [20]. Thus, while the binding of IVM potentiates P2X4R, it also increases both the apparent rate of desensitization at low ATP concentrations and the rate of recovery from desensitization (i.e., it lowers the Gibbs free energy barrier for these transitions).
To assess the deactivation kinetics (i.e., decay of current amplitude following washout of agonist) of P2X4R, we used the same pulse protocol of 1 μM ATP for 2 s twice per minute (Fig 1C). In the absence of IVM (filled circles), receptors underwent fast deactivation with a time constant that remained roughly the same at each pulse, whereas in the presence of 1 μM IVM (open circles following the small arrow) the deactivation time constant progressively increased with incubation time, indicating a decrease in receptor deactivation rates. This effect became even more pronounced at higher IVM concentrations. At IVM concentrations greater than or equal to 10 μM, deactivation following washout of IVM was not always complete (see Fig 2A in [20]), suggesting that complex physiological processes might be initiated at these concentrations. These results are consistent with the idea that IVM increases the sensitivity of the receptor to ATP and decreases the rate of agonist unbinding following its washout from medium [16–18].
Our previous study showed concentration response curves for rat P2X4R stimulated by ATP in the presence and absence of 3 μM IVM (S3 Fig); ATP alone was found to produce a concentration response curve with an EC50 of 2.3±0.4 μM (blue line), and with 30-s pretreatment with IVM, the EC50 was 0.5±0.1 μM (green line) [20]. A similar conclusion was reached with human P2X4R [17]. A pretreatment period of 10 s was also considered and, whereas it did produce the same maximal current amplitude, an intermediate EC50 of 1.6±0.3 μM was measured (maroon line) [20]. This suggests that there are at least two distinct priming effects associated with IVM with separate time scales of action. First, IVM primes receptors by increasing the maximal whole-cell current response. Second, after prolonged exposures to IVM, receptors become further primed by an increased sensitivity to ATP (previously called sensitization). The model in [20] was only partially able to account for these behaviors and specifically was not able to account for their dependence on the duration of pre-treatment because there were no kinetics associated with IVM binding in the absence of ATP.
The concentration response curves of P2X4R (S3 Fig) reveal that not only do 10- and 30-s pretreatments with IVM increase sensitivity to ATP (maroon and green lines, respectively), but they also increase the maximum current amplitude evoked by ATP [17, 20]. The two hypotheses that can explain this behavior are: (i) the unitary conductance of individual channels increases; or (ii) the number of open receptors is rising (i.e., the maximal open probability increases). Although there is evidence that the former hypothesis holds [17], this does not preclude a change as well in the maximal open probability with IVM application [17, 20, 23]. In fact, it was reported that the maximal open probability in the absence of IVM is ~0.2 compared to ~0.8 in the presence of IVM [17, 23]. This phenomenon was previously explained by the Markov model in [20], which assumes that IVM modifies the connectivity between open and desensitized states, but that model required a large increase in unitary conductance. This assumption on the conductance is inconsistent with a study of human P2X4R [17], which showed that IVM produces a roughly 5-fold increase in maximal current amplitude while only inducing a 20% increase in unitary conductance. Those authors posited that, in the absence of IVM, desensitization plays a large role in reducing the current amplitude, whereas when IVM is applied, desensitization is greatly reduced, enhancing the observed current. While this is a plausible explanation, we are not aware of any receptors that function in this manner. Moreover, no quantitative analysis was made to assess to what extent such a mechanism produces the observed effect.
In order to test this hypothesis, we constructed a simple and generic mathematical scheme (hereafter referred to as a gating scheme) of a desensitizing ligand-activated receptor (Fig 2A). It consists of two rows: a naïve row comprised of two closed states (C1, C2) and two conducting states (Q1, Q2), and a desensitized row comprised of four nonconducting desensitized states (D1, D2, D3, D4). As was done in [20], we assume that channels open from states with two or three bound ATP molecules. This is in accordance with the finding that a single-bound receptor state does not lead to activation of P2X7 channels [28]. This is also consistent with previous models of P2XRs and the notion that a single kinetic model underlies the functioning of all receptor subtypes [19, 20, 24–26, 29]. Forward (backward) transitions between two states along each row represent a single ATP binding (unbinding) with rates k2, k4, k6 (k1, k3, k5), respectively, whereas upward (downward) transitions represent desensitization (recovery) with a rate kd (kr). Concentration response curves were generated for this gating scheme, each with a progressively increasing rate of desensitization kd (see Figs 2A and S5A). It was found that although reduced desensitization rates are capable of increasing the current amplitude at a given agonist (such as ATP for P2X4R) concentration, the mechanism proposed in [17] is unable to significantly increase the maximal current amplitude evoked by the agonist. Rather, it shifts the EC50 of the concentration response curves leftward as well as increases the Hill coefficient in such a way that the saturating phase of the concentration response curves are shifted by many orders of magnitude. A leftward shift in EC50 and modulation of the Hill coefficient by IVM have been observed experimentally [17, 20]. This was, however, consistently associated with an increase in the maximal current amplitude, which the desensitization mechanism cannot produce at saturating agonist concentrations (see Imax in the legend of S5A Fig). A mathematical model introduced by Silberberg et al. also used an IVM-dependent transition rather than modulation of desensitization by IVM in order to produce the increase in maximal current [23]. Therefore, the mechanism suggested in [17] seems unable, at least on its own, to explain the effects of IVM on the concentration-response relationship for the peak current of P2X4R.
After having ruled out decreased desensitization as a cause for the increased maximal current amplitude in the presence of a modulator, we tested an alternative hypothesis, that the closed states exist in equilibrium with a deeply inactivated state (C0) for which the agonist is not effective (Fig 2B). This mechanism has previously been used in Markov models of sodium channels [30]. Transitions linking the two states must be slow, but the equilibrium mixture of the closed-inactivated subsystem (C0 ↔ C1) establishes an upper bound on the maximal open probability in the absence of IVM, given by
POMax=H1H1+H2
(6)
even at the highest agonist concentration. In other words, before agonist application, only some fraction of receptors are in C1 and are susceptible to agonist-induced activation but in the presence of IVM, more can be recruited (from C0) into C1. Evidence of such recruitment was first seen during prolonged application of ATP in the absence and presence of IVM [20] and was obtained also by application of IVM to fully desensitized receptors (S4 Fig). To see how effective this mechanism is in producing the observed effects in [17], we tested the gating scheme of Fig 2B quantitatively, by progressively increasing the transition rate H1 (between C0 and C1), and plotting the concentration response curves (S5B Fig). Increasing H1 decreased inactivation, which increased the fraction of receptors in C1 and thus POMax.
Whereas reducing the occupancy of the deeply inactivated state is highly effective at increasing the maximal current amplitude, it does not significantly shift the concentration response curves or alter the Hill coefficients. Therefore, in order to match the experimental findings that IVM pretreatment of P2X4R not only increased maximal current but also shifted the EC50 leftward (S3 Fig) and increased the Hill coefficients, both reduced desensitization and rescue from a deeply inactivated state seem to be required. A model incorporating both features is described in the next section.
Based on the above considerations, we designed a one-layer Markov state model that describes the full kinetics of ATP and IVM binding to P2X4R and tested it against experimental data. For a detailed description of the model, see S6 Fig, Table A and Appendix A in S1 Text. Briefly, it is a revised version of the model of Zemkova et. al. [20] that now assumes 3 IVM binding sites, that the binding of IVM acts on P2X4R independently of ATP binding, and that IVM can bind to any ATP-bound state, not just the 3-ATP bound naïve state. Sequential binding of IVM causes three stages of receptor priming, depending on number of IVM molecules bound to receptor: primed-1, primed-2, and primed-3. Primed-1 receptors respond to ATP application with increased current amplitude, reflecting increased open probability. Primed-2 receptors exhibit modestly increased unitary conductance for Na+ and significantly increased unitary conductance for NMDG+, whereas primed-3 receptors show increased ATP binding affinity. The model also incorporates rescue from the deeply inactivated state by IVM, and therefore has a maximal open probability given by Eq (6) in the absence of IVM.
Although our analysis of this model (and several variations of it) revealed that it possesses many of the necessary ingredients to capture the gating properties of P2X4R and several aspects of its current recordings (S7 and S8 Figs), it includes the implausible assumption that receptors in the primed states must lose all bound IVM molecules in order to desensitize. This assumption led to two major issues in the performance of the model: (i) it did not capture accurately the short timescales of activation and desensitization robustly; and (ii) it produced discrepancies in current amplitudes when compared to experimental data during the pulse protocol. That motivated us to design a more accurate model of P2X4R kinetics.
Data-based Markov state models that describe the processes of ligand binding/unbinding to ligand-gated receptor are powerful tools to understand orthosteric and allosteric regulation of these channels. P2X4Rs are prototypical examples of such receptors with orthosteric and allosteric binding sites for ATP and IVM, respectively. They are associated with ion channels that are permeable to small cations, including Na+, Ca2+, Mg2+ and K+. The binding of ATP leads to receptor activation and channel opening, while IVM binding increases receptor unitary conductance and sensitivity to ATP. Here we analysed the kinetics of ATP and IVM binding/unbinding to P2X4R, and determined its gating properties using two detailed Markov models labelled the one-layer and two-layer models.
The one-layer model extended a previously developed, simple Markov model of P2X4R by taking into consideration a deeply inactivated state, nonresponsive to ATP but responsive to IVM (existing in equilibrium with the naïve ATP-unbound state), along with three additional gating schemes (per each ATP binding) representing the three IVM binding sites. The model also assumed that the IVM and ATP binding are independent of one another and that sequential binding of IVM can occur at any ATP-bound or unbound states. Our analysis revealed that the deeply inactivated state was essential for capturing the increase in the maximum response (Imax) in the ATP-dependent concentration-response curves (in the presence of IVM), with only a small increase in conductance between the open1 and open2 states. Although the model was able to capture many of the essential features of P2X4R recordings (S7 and S8 Figs), it assumed that IVM bound states can only desensitize by first becoming completely free of IVM. That made the model unable to robustly capture the short timescales of activation and desensitization, and it also produced current amplitudes incompatible with experimental data during the pulse protocol.
By allowing the IVM-bound states to desensitize, we were able to show that a simple gating scheme is able to capture the profiles of the first pulse (in the absence of IVM) or the last pulse (in the presence of IVM) of the pulse protocol very accurately when the scheme is fitted individually to each pulse, but not both simultaneously. However, when comparing the entire pulse protocol recording to the outcome of the scheme for each case, there was a gradual increase in discrepancy between them, suggesting that a mixture of gating schemes must coexist to be able to capture all aspects of P2X4R kinetics. That led us to propose the two-layer model, which assumes that IVM-bound states can desensitize.
The two-layer model was successful in capturing every aspect of P2X4R kinetics very accurately, including the short and long time scales of activation and desensitization, particularly the changes in the desensitization rate observed during the pulse protocol of Fig 1, as well as the current amplitudes. The observed shift in the EC50 along with the increase in the maximum current, during pre-stimulation with IVM, were also reproduced by the model (through the presence of the deeply inactivated state). Moreover, these gating schemes can be used to understand why ATP binding mutants with low amplitude of response tend to have significantly larger fold-increases in maximal current in the presence of IVM [21]. If we view such mutants as disproportionately populating the deeply inactivated state, where they cannot bind ATP, then their rescue by IVM from this state will produce a much larger fold-increase in maximal current. The existence of this deeply inactivated state was probed experimentally, by applying IVM to a cell whose receptors were almost completely desensitized from prolonged applications of 100 μM ATP. Upon IVM application, we observed an increase in the maximal current amplitude to about half of the initial maximal current (see S4 Fig), suggesting that IVM rescued receptors from a (deeply inactivated) pool corresponding to about one third of all receptors.
The two-layer model consisted of 4 gating schemes linked together through ATP/IVM binding/unbinding. Two of these gating schemes (the primed-2 and -3) contained conducting states exhibiting a 15% increase in unitary conductance compared to that of the open states in the naïve and primed-1 states. This increase is within the 20% limit seen experimentally [17], and is not required to produce IVM’s increase in maximal current amplitude within our model. Instead, the effect of IVM on maximal current amplitude is produced mainly by an increase in open probability. According to Eq (6), the two layer model predicts a maximal open probability of approximately 0.53 in the absence of IVM (with 47% of receptors in a deeply inactivated state), while it can easily reach values greater than 0.9 in the presence of IVM. It was suggested in [13] that the ionic conditions in the medium and the pipete are responsible for producing electrochemical effects which were long presumed to be evidence for pore dilation, particularly in P2X2R. While temporal changes in ionic gradients play a significant role in producing a shift in Erev associated with the I-V curves during the ramp protocol, our results suggest that the transition to open2 (which is permeable to NMDG+) is an intrinsic property of the pore in P2X4R, is independent of the increase in unitary conductance (S12B Fig) and is induced by priming with IVM. The two-layer model assumes that the increase in unitary conductance associated with this transition is masked by desensitization. This results in the shift in Erev being accompanied by a decrease in the slope of the I-V curves (due to desensitization). The observed decline in the slope of the I-V curves was a consequence of Inequalities Eqs (3) and (5). The previously developed model in [20] was capable of fulfilling these conditions and producing the decrease in the slope of the I-V curves, but it required a large increase (>150%) in unitary conductance to achieve it while simultaneously producing the increase in current amplitude induced by IVM. The inclusion of the primed-1 row with conductance g1 (and reversal potential E1) was an essential element for reproducing the shift in the I-V curves. Without this intermediate step, the model required a very positive reversal potential for the open2 state (E2), which does not reflect its loss of selectivity. Both the one-layer and two-layer models proposed here keep the increase in unitary conductance within 20% and produce the current growth with IVM pretreatment through IVM-induced transitions from the inactivated state C0 (by increasing POMax, as given by Eq (6), rather than increasing the maximal conductance). The two-layer model, however, is more plausible because it does not assume that desensitization necessitates the unbinding of IVM. Moreover, according to this model, IVM is able to transition receptors to the primed-3 states in the absence of ATP, allowing pretreatment with IVM to produce sensitization independently of ATP. It is important to note that effectively removing the 15% increase in unitary conductance associated with the primed states only slightly altered model simulations and did not abolish any experimental phenomena in symmetric ionic conditions (S11 and S12 Figs). Thus, the results of the two-layer model are independent of an increase in unitary conductance, but require a change in selectivity in order to capture the shift in the I-V curves of the ramp protocol.
Recently, the pore dilation hypothesis has become increasingly disputed. Molecular dynamics simulations indicate that NDMG+ is capable of permeating the open state of P2X4R pore, provided it is maintained in an open state long enough for the slow permeation event of NDMG+ to take place [15, 32]. One of the primary effects of IVM application on the single channel kinetics of P2X4R is to shift the distribution of open times from the sub-millisecond timecale to tens of millisecond [17]. We hypothesize that the drastic change in P2X4R’s permeability for NMDG+ upon IVM application results from the priming of receptors in such a way that their pores remain in the open2 state for long enough for the slow permeation of NMDG+. This increase in the permeability for NMDG+ (via an IVM-dependent transition to the open2 state) allows for its influx into the cell. Together with the efflux of Na+, these fluxes produce a more positive Erev as determined by the Goldman-Hodgkin-Katz equation.
While the two-layer model may seem to be a large departure from both the previously developed model in [20] and the one-layer model, it should be noted that in the absence of IVM, the remaining blocks of the models (or submodels) are identical, and that the P2X2R Markov model developed in [19] has a similar structure; it included a corresponding desensitized state for each of its closed and open states, although the desensitization pathway for primed (sensitized) states was calcium dependent. The increase in the number of states and number of kinetics parameters in the two-layer model was necessary to capture all the observed features of P2X4R, which previous models, including the one-layer model presented here, failed to do. A step by step validation of such an increase in complexity was provided through the use of coupled gating schemes, and the design of an extensive MCMC fitting algorithm that combined parallel tempering approaches with the t-walk method to estimate the kinetic parameters of the model efficiently.
The two major allosteric effects of IVM’s on P2X4Rs, observable from whole cell currents as an increase in maximal current amplitude or the deactivation time constant, exhibit distinct concentration dependencies [17, 20]. This suggests that they are likely caused by two independent processes. This existence of two distinct allosteric effects with differing concentration dependencies have also been reported for other P2XRs [11, 19], although for these receptors, ATP alone was sufficient to induce such effects. The models presented here reduce the concentration dependence of IVM’s allosteric effects on P2X4R to a single sequential binding process, and thus represent a major simplification of a more realistic model where all effects are assumed to arise from independent binding events. Despite this simplification, the model is quite capable of capturing all aspects of allosteric modification by IVM.
An important item for future work is cooperativity in the ATP and IVM binding, which has been investigated in the one-layer model but not yet in the two-layer model. By assuming correlations between the binding/unbinding parameters of ATP and IVM in the one-layer model, we were able to reduce the number of estimated parameters in that model significantly and found that there was negative cooperativity in the ATP binding in the naïve and primed-1 rows. Investigating if such cooperativity exists in the two-layer model is also warranted. This can be done by imposing correlations between the kinetic parameters of the two-layer model, which will again reduce the number of estimated parameters, and testing for cooperativity in ATP binding, IVM binding and between ATP and IVM binding. These variations of the model can then be compared to each other using Bayesian approaches to determine which is most likely.
In conclusion, here we present two novel models, one of which (the two-layer model) effectively mimics all experimental observations. In this model, receptors go through four stages of activation cycle during ATP and IVM binding: transitioning from functional to desensitized, from desensitized to internalized, from internalized to deeply inactivated and from deeply inactivated to functional. Functional and desensitized stages each exist as 16 distinct states, determined by the progressive saturation of three ATP and three IVM binding sites, whereas internalized and deeply inactivated receptors are single states. Binding of IVM influences ATP-induced gating properties of receptors, i.e. the rates of activation, desensitization and deactivation, open probability of channels, and the sensitivity of receptors to ATP. The channel pore state, open1 is predominantly permeable to small cations and open2 is permeable to large organic cations.
Experiments were performed on human embryonic kidney 293 cells (HEK293; American Type Culture Collection), which were grown in Dulbecco’s modified Eagle’s medium supplemented with 10% fetal bovine serum, 50 U/mL penicillin, and 50 μg/mL streptomycin in a humidified 5% CO2 atmosphere at 37°C. Cells were cultured in 75-cm2 plastic culture flasks (NUNC, Rochester, NY, USA) for 36–72 h until they reached 80–95% confluence. Before the day of transfection, ~150 000 cells were plated on 35 mm culture dishes (Sarstedt, Newton, NC, USA) and incubated at 37°C for at least 24 h. For each culture dish of HEK293 cells, transfection of wild-type P2X4R was conducted using 2 μg of DNA with 2 μl of jetPRIME reagent in 2 ml of Dulbecco modified Eagle’s medium, according to the manufacturer’s instructions (PolyPlus-transfection, Illkirch, France). After 24–48 h of incubation, the transfected cells were mechanically dispersed and re-cultured on 35 mm dishes of Corning 3294 CellBIND Surface for 1–4 hours before recording. Transfected cells were identified by the fluorescence signal of enhanced green fluorescent protein using an inverted research microscope with fluorescence illuminators (Model IX71; Olympus, Melville, NY).
Currents were recorded in a whole-cell configuration from cells clamped to −60 mV using an Axopatch 200B patch clamp amplifier (Axon Instruments, Union City, CA, USA). All currents were captured and stored using Digidata 1550A and pClamp10 software package. Patch electrodes were pulled from borosilicate glass tube with a 1.65 mm outer diameter (type GB150F-8P; Science Products GmbG, Hofheim, Germany) using a Flaming Brown horizontal puller (P-97; Sutter Instruments, Novato, CA). The tip of the pipette was heat-polished to a final tip resistance of 3–5 MOhm. During the experiments, the dishes with cell cultures were perfused with an extracellular solution containing: 142 mM NaCl, 3 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 10 mM HEPES and 10 mM D-glucose, adjusted to pH 7.3 with 10 M NaOH. The osmolarity of solution was 290 mOsm as determined by a vapor pressure osmometer (Model VAPRO 5520; Wescor, Logan, UT, USA). Experiments were done on single cells with an average capacitance of about 10 pF held at membrane potential of -60 mV. Patch electrodes used for whole-cell recording were filled with an intracellular solution containing: 145 mM NaCl, 10 mM EGTA and 10 mM HEPES; the pH was adjusted with 10 M NaOH to 7.2. The osmolality of the intracellular solution was 293 mOsM. Current-voltage relations were obtained by voltage ramps from −80 mV to +80 mV twice per second and used to estimate changes in reversal potential during 10–30 s of agonist application (Yan et al., 2008). Under a ramp protocol, cells were bathed in extracellular solution containing: 155 mM N-Methyl-d-glucamine (NMDG+), 3 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 10 mM HEPES and 10 mM D-glucose, adjusted to pH 7.3 with HCl, if not otherwise stated. IVM was dissolved in dimethyl sulfoxide, stored in stock solutions at 10 mM, and diluted to required concentration in extracellular solution prior to experiments. The control, ATP-containing and IVM-containing solutions were applied via a rapid perfusion system (RSC-200, BIOLOGIC, Claix, France) consisting of an array of 5 glass tubes each approximately 400 μm in diameter. The application tube was routinely positioned at about 500 μm distance and about 50 μm above the recorded cell. A complete change of the solution around the cell took between 5–20 ms, depending on the speed of the solution expelled.
The one-layer Markov state model developed here (S6 Fig), is a revised version of a previously developed Markov model describing P2X4R orthosteric activation by ATP and allosteric modulation by IVM [20], whereas the two-layer model (Fig 3) is more complex in nature, taking into consideration all processes involved in ATP and IVM binding. Both models assume the presence of three ATP and three IVM binding sites and were tested against current recordings to compare their performance in capturing the physiological properties of P2X4R. The following symbols were used to describe the various states of the model: C for closed, Q for conducting (open1 and open2), D for desensitized and Z for internalized states, each representing the fraction of receptors in a given state. The transition rates between the various states are in Tables A and C in S1 Text. Detailed balance was not explicitly incorporated into these models because it was assumed that P2X4R never reach absolute equilibrium during ATP and IVM stimulation. Stimulation with ATP during the pulse protocol (repetitive stimulation with 1 μM ATP for 2 s twice per minute in the absence and presence of various concentrations of IVM applied to the bath medium after two ATP pulses) and the prolonged protocol (stimulation with 100 μM ATP for extended periods, longer than 1 min, in the absence and presence of 3 μM IVM) were modeled as a square wave and a rectangular function for the duration of ATP application, respectively; stimulation with IVM was modeled as a rectangular function for the duration of IVM application. The ramp protocol was modeled as a sawtooth-like sequence of upstrokes with slope 320 mV/s (rising from −80 mV +80 mV over 500 ms). Detailed descriptions of the two Markov models along with their differential equations are provided in Appendices A and B in S1 Text.
Concentration-response data points were fit to a hill function
y=Imax1+(EC50/x)n,
where y is the amplitude of the current evoked at a particular ATP concentration x, Imax is the maximum current observed at 100 μM ATP, EC50 is the ATP concentration producing 50% of the maximum current, and n is the Hill coefficient. Deactivation kinetics of the current decay after agonist washout were fitted to a single exponential
y=A1exp(−t/τ1)+C
or to a sum of two exponentials
y=A1exp(−t/τ1)+A2exp(−t/τ2)+C,
where A1 and A2 are the amplitudes of decay for the first and second exponentials, τ1 and τ2 are their decay time constants, and C is the baseline current. In the case where the sum of exponentials fits the data better than a single exponential, we report the weighted time constant
τoff=A1τ1+A2τ2A1+A2.
In either case, we labeled the derived time constant of deactivation as τoff. Statistical significance (**p<0.01 and *p<0.05) was assessed using the Wilcoxon signed rank test. MATLAB (MathWorks, Natick, MA) was used to solve the differential equations of the models numerically, fit the models to the data and apply statistical tests.
In order to quantify the rate of desensitization from current traces of the pulse protocol (see Figs 1 and 4) which do not show complete desensitization, we note that, at low concentrations of ATP (1 μM), desensitization is a mono-exponential process and thus employ a mono-exponential model
I(t)=Ae−tτ,
where A is the magnitude of the current at the onset of desensitization and τ is the time constant of desensitization. By first normalizing the current and then evaluating its derivative, with respect to time, at the onset of desensitization (t = 0), we obtain
I˙(0)A=−1τ.
Thus the first derivative at time t = 0 of each desensitizing current, I˙(0), normalized by its maximum current (as has been plotted in Fig 1B), yields information about the time constant of desensitization (i.e., a small normalized initial desensitization rate corresponds to a large time constant of desensitization). Due to the simultaneous activation and desensitization of multiple receptors, I˙(0) was estimated using a linear fit for the small window of time (1–1.5 s) after cells achieved their maximal current and before agonist was removed. This window is relatively small compared to the 6 s time constant of desensitization for P2X4R, and therefore this approximation method provides a relevant estimate of I˙(0)/A, which serves as a proxy for desensitization time constant τ.
Parameter estimation was performed using MCMC techniques. Model simulations were generated using ode solvers in MATLAB and then fit to experimental recordings. Generally, MCMC produces Markov chains Λ = {x1, x2, ⋯, xM} of model parameters xm=(p1m,p2m,⋯,pNm), where N is the number of parameters (pi) and m = 1, 2, ⋯, M is the mth iterate of the Markov chain. The iterates represent samples from the posterior distribution π(x) determined using Bayes’ theorem as follows
π(x)∝L(x)P(x),
where L(x) = P(data | x), the likelihood function, is the probability of observing the data given the parameter values of x, and P(x) is the prior distribution of x, which reflects any prior knowledge about the parameter values independent of observed data. Proportionality, indicated by ∝, is sufficient—therefore there is no need to normalize the posterior.
In order to increase mixing of modes in parameter space, we used the parallel tempering algorithm which produces Markov chains in the product space
Xm = {xm(1), xm(2),⋯,xm(L)},
where each chain xm(l), l = 1, 2, ⋯, L, was sampled from a tempered distribution πβ(l) and β(l) is the inverse temperature of each chain. Parameter sets were stochastically swapped between chains according to the swap kernel of Miasojedow et al., and their strategy of adaptively updating the inverse temperature of each chain [33] was adopted. Because Metropolis-Hasting move-kernels can be difficult to tune for continuous-time Markov models of ion channels [34], we used the adaptive move kernel of the t-walk sampler [35] instead. Since the t-walk samples from the product distribution π(x)π(x′), the composite MCMC method samples from the product distribution
πβ(X)πβ(X′)=πβ(1)(x(1))πβ(1)(x′(1))∫πβ(1)(x)πβ(1)(x′)dxdx′×…×πβ(L)(x(L))πβ(L)(x′(L))∫πβ(L)(x)πβ(L)(x′)dxdx′.
Given that we have a set of discretely sampled whole-cell current recordings, we initially adopted the likelihood function from Gregory [36], defined by
L(x)=P(Iμ,σ|x)=exp{−∑i=1K(Iμ,i−Ix,i)22σi2},
where the index i refers to the ith discretely sampled data point, K is the number of data points in the experimental recording, Iμ,i and σi are the ith samples, respectively, of the mean current and current standard deviation estimated from the data set, and Ix,i is the ith sample of the current produced by the model given the set of parameters x. To circumvent (i) sampling inefficiency from the posterior distribution, which is exacerbated by the use of high data sampling rates, and (ii) very poor fitting of rapid transient behaviour, due to the value of the likelihood being dominated by slower portions of the signal with more data points, we opted to fit (using the least-squares method) both experimental and simulation data to appropriately chosen functions and to compare the fit parameters of the experimental and simulated data. For example, we have used exponential functions to measure deactivation kinetics (as described above). This results in the likelihood function
L(x)=P(τoff,μ,στ,Aoff,μ,σA|x)=exp{−(τoff,μ−τoff,x)22στ2−(Aoff,μ−Aoff,x)22σA2},
(7)
where τoff,μ and στ are the mean deactivation time constant and its variance, respectively, Aoff,μ and στ are the mean deactivation amplitude and its variance, respectively, and τoff,x and Aoff,x are the deactivation time constant and amplitude produced by the model corresponding to the parameter values x.Using this description-based approach (rather than the distance from data), we were able to simultaneously fit numerous aspects of P2X4R activation kinetics and their allosteric modulation. This was done by comparing experimental data and model predictions of (i) time dependence of the activation time and normalized rate of desensitzation in the absence and presence of 1 μm IVM (Fig 1A and 1B) (ii) maximal current, deactivation time constant, and desensitization at 1 μM ATP with increasing IVM concentrations (Fig 4E and 4F) (iii). Insensitivity to ATP removal at 10 μM IVM (Fig 4H) (iv) decay of current amplitude deactivation time constant following IVM washout (Fig 4A–4D) (v) activation, desensitization, and recovery after washout of 100 μM ATP in the absence and presence of IVM (Fig 5C and 5D) (vi) EC50 and Hill coefficient (n) of the ATP concentration-response curves for peak current (Fig 5A).
The degree of cooperativity in ATP binding was determined from the Markov chain xm of 5000 samples associated with each ATP binding and unbinding rate ki, i = 1, 2, ⋯, 24, that was generated from data fitting, followed by calculating the chains of ATP binding affinities
3k6n+2k6n+1,k6n+4k6n+5,k6n+63k6n+5, n=1,2,3
along each of the non-desensitized rows of the one-layer model (including the four lower rows) and the two-layer model (including the four rows in the upper layer). The posterior distributions associated with these affinities were used to compare the values of the most frequently sampled points along each row (n = 1, 2, 3). In the presence of a specific cooperativity between ATP bindings, correlations between the different binding affinities were detected and reported.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.